technicalmarketing.ai Free AI audit

Paper · November 2026

How technical marketing systems make AI profitable.

94,388 words · about 6.8 hours · 14 chapters · 294 sources

Abstract

AI use is spreading fast and documented returns are scarce. In the Census Bureau's survey of businesses, overall AI use ran between 17% and 20% from December 2025 to May 20261. In a controlled study of customer support work, access to an AI tool raised issues resolved per hour by 14% on average2. Neither figure says whether a firm earned more. This paper asks where the money is for a service business, and argues that AI pays when it is installed inside a measured customer pipeline and has no way to show a return otherwise.

The argument rests on the economics of general-purpose technologies, whose returns depend on complementary investment, and on the controlled studies of what AI does to work. Those studies concern large employers and earlier technology, so applying them to a small firm is an inference, and the paper labels it as one. We argue that three complements matter most to a service business: a flow of demand for the AI to act on, a process for it to sit inside, and measurement that can tell what it earned. A technical marketing system supplies all three.

The paper sets out a profit identity with a line for the demand a plan produces, four gains for the agent (capture, conversion, capacity and retention) and three costs, prices the cost side from published rates and official wages, shows how a small firm can attribute outcomes to an agent, treats legal risk as a cost, reads the public record of failures, and applies the model to dental, medical, legal and accounting practices. On published prices and stated assumptions a three-minute call costs between about $0.18 and $0.42 to run (Chapter 5), so the question that decides profit is how many customers an agent adds. No study of small firms that we found settles that question, and the last chapter says what would.

What this paper argues

  1. AI is a general-purpose input whose return depends on complements. In our judgment, the cheapest complement a service business is missing is a measured customer pipeline.
  2. A system earns through the demand its plan produces and through the agent's four mechanisms (capture, conversion, capacity and retention). Each can be zero: an agent that answers contacts the firm was already answering captures nothing. A fee counted as a cost has its product counted as a gain.
  3. On published prices the cost of running an agent is small beside the value of one customer, so the question that decides profit is how many customers it adds. Only measurement can answer that.
  4. Legal and reputational risk is a term in the profit equation. Statutes and cases price it, and scope limits, disclosure and hand-off to a person lower it.
  5. Demand and measurement come before automation. An agent bought before either is a cost with no way to show a return.

How to read this paper

~
Our estimate, built from the nearest cited figures. Not a measurement.
†
A vendor’s own figure, from its customers or its own survey. Not an independent sample.
NF
Not found: we looked for a published measure and found none.
( )
A figure followed by its working in brackets, or by “calc.”, is our arithmetic on cited figures.
Rules
We report what each rule says as read on the date shown. This is not legal advice, and counsel decides how a rule applies to you.
Sources
A small number points to the source in References, which says where we hold a source only through its abstract or a secondary account.

Chapter 1

The profitability problem

Is AI profitable? Wide use and thin evidence of return

In a Census Bureau survey, overall AI use by US businesses hovered between 17% and 20% from December 2025 to May 20261. In a report from MIT Project NANDA, 95% of organizations were getting zero return from generative AI3. The two figures come from different instruments and count different things, so they are not two readings of one quantity. They do mark the ends of the puzzle this paper takes up. By some counts, a large share of the organizations surveyed use AI. By other counts, estimates and a prediction, many AI projects fail, are abandoned or return nothing. An owner deciding whether to buy an agent needs a figure for return, and neither is one.

The puzzle is often stated more strongly than the evidence allows. Its strongest form says that adoption is nearly universal and few can show a return. The first half describes what one survey's respondents report and does not hold for the national survey described below. The second half rests on one report whose method we set out later in this chapter, on a forecast, and on an estimate that a research organization relays from others. The weaker form that the evidence does support is this: some counts report widespread use, and most of the available counts show little demonstrated return. The exception is a vendor survey. Thomson Reuters, a publisher that sells AI products, reports that more than half (53%) of the professionals in its 2025 survey said their organizations were already seeing ROI as a direct or indirect result of their AI adoption4. The figure is self-reported, and Chapter 2 weighs it.

People ask the question directly. In Google Ads keyword data for the United States, collected in November 2026, "is ai profitable" is searched 1,900 times a month, "ai profitability" and "ai roi" 590 times each, and "roi of ai" 70 times5. Broader phrases draw more: "ai for business" is searched 8,100 times a month and "ai for small business" 3,600 times5. The data have limits that apply throughout the paper. The phrases are ones an owner or a vendor would type. Volume shows interest, not purchase. Cost per click is what advertisers pay to appear, not what a customer is worth.

The same data show where most of the interest sits. The phrases "ai receptionist", "ai agent" and "ai voice agent" are searched 74,000, 49,500 and 33,100 times a month5. Those three total 156,600 searches a month (74,000 + 49,500 + 33,100). The five phrases that ask whether AI pays are the four above and "generative ai roi" at 10 searches a month5. They total 3,160 (1,900 + 590 + 590 + 70 + 10), so the first group is about 50 times the second (156,600 / 3,160). One person may type several phrases, so the totals overcount people, and "ai agent" can mean many things that the data cannot separate.

Cost per click is less uniform. A click on "roi of ai" costs $148.155. A click on "is ai profitable", the most searched of the five questions, costs $3.40, against $43.66 for "ai receptionist" and $49.07 for "ai voice agent"5. We think far more people are looking to buy AI than are asking whether it pays.

AI adoption: what the adoption figures count

The word "adoption" stands in the evidence for at least three different quantities, and each figure has to be read with its quantity in view. The first is a count of businesses. The Census Bureau's Business Trends and Outlook Survey (BTOS) surveys US businesses. Its first AI supplement found that 3.8% of businesses reported using AI to produce goods and services, and that the Information sector led with 13.8%6. Between September 2023 and February 2024 its bi-weekly estimates of firm AI use rose from 3.7% to 5.4%, and the authors expected about 6.6% by early fall 20247. For December 2025 to May 2026 the Bureau reports overall usage between 17% and 20%, with between 20% and 23% of businesses expecting to be using AI within six months1.

Read together, those readings invite a growth rate, but we do not treat the ratio between them as one. The low end of the 2026 range is more than three times the February 2024 reading (17 / 5.4 = 3.1). It is also far above the 6.6% that the earlier paper expected for the fall of 20247. Use may have grown faster than expected, or what the survey counts may have changed between releases; the cited claims do not say which. The first reading is specific about its question, which was use of AI to produce goods and services, while the 2026 claim says only "overall AI usage". We use the readings as separate descriptions of separate dates.

The second quantity is a count of organizations as their own respondents describe them. A McKinsey survey (secondary) reports that 78% of respondents say their organizations use AI in at least one business function8. The figure was 72% in early 2024 and 55% a year earlier8. The same survey reports that 71% say their organizations regularly use generative AI in at least one business function8. That is a low bar, since one team using an assistant to draft text would meet it. The claims do not say who the respondents are or how they were chosen, so we read the figure as what those respondents report about their own organizations.

The third quantity counts people. A study of how people use ChatGPT found that by July 2025 the consumer product had been adopted by around 10% of the world's adult population9. The same study found that non-work messages grew from 53% to more than 70% of all usage9, so a person-level count of one product says little about firms. Table 1 lays the instruments out side by side with what each one counts.

InstrumentWho is countedWhat counts as useFigureClaim
Census Bureau BTOS, first AI supplement (government survey)US businessesUsing AI to produce goods and services3.8%6
Census Bureau BTOS, December 2025 to May 2026 (government survey)US businessesOverall AI usage, as the release words it17% to 20%1
McKinsey survey (consultancy; secondary source)Respondents' organizationsUse of AI in at least one business function78%8
ChatGPT usage study (Chatterji and co-authors, 2025)Adults worldwideAdoption of one consumer productAbout 10%9

Table 1. Four adoption figures and what each one counts. The figures are not comparable with one another.

None of these instruments measures whether the use earned anything; they count use. Studies of usage add a different fact, which is what people do with the tools. In the ChatGPT study, "Practical Guidance", "Seeking Information" and "Writing" are the three most common topics and together account for nearly 80% of all conversations9. In the Anthropic† Economic Index (one assistant), the top 10 tasks made up 19% of all traffic in February, down from 24% in November10. Each describes the use of one vendor's assistant, and neither says whether a given use was part of a repeated business process.

AI ROI: what the failure and abandonment figures count

Three figures describe failure, abandonment or the absence of return, and they differ in kind more than the adoption figures do, so we name the kind each time. The first is a research report. Its authors at MIT Project NANDA titled it The GenAI Divide: State of AI in Business 2025, and it states that 95% of organizations are getting zero return3. Its stated method is a multi-method design: a systematic review of over 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders3. The cited claims do not say which of those three sources the headline figure is computed from, how return was defined, or over what period it was judged.

The second figure is an estimate that a report relays. A RAND Corporation report (secondary) on the causes of failure in AI projects states that by some estimates more than 80% of AI projects fail, twice the failure rate of information technology projects that do not involve AI11. The 80% is introduced with the words "By some estimates", and the cited claim does not say that the report counted anything itself. The RAND report also describes interviews with experienced AI practitioners from industry, and says that across all of the interviews five dominant root causes emerged11. The cited claims give the number of causes but not their content, and nothing in them defines what "fail" means.

Last comes a prediction. Gartner (secondary) predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value12. A prediction is not an observation. The year it named has ended, and this paper does not report a later check of it. Its unit is the generative AI project that reaches proof of concept, so projects that never get that far fall outside it, and so do projects that ship and then fail to pay. Table 2 sets the three figures side by side.

SourceKind of figureUnitWhat it saysClaim
MIT Project NANDA, 2025Report with a multi-method designOrganizations95% getting zero return3
RAND Corporation, 2024Estimate introduced as "By some estimates"; the report also rests on practitioner interviews (secondary source)AI projectsMore than 80% fail11
Gartner, 2024Prediction, from a secondary sourceGenerative AI projects after proof of conceptAt least 30% abandoned by the end of 202512

Table 2. Three failure and abandonment figures and what kind of figure each is.

The unit of each figure is a project or an organization, never a contact or a customer, so none of them can say what an agent earned on the calls it took. They also report outcomes but say little about the conditions that separated the projects that paid from those that did not, except through the causes that the sources name. Gartner's list of causes includes unclear business value. We think that is a problem of measurement.

AI for small business: why the question is sharper

For a ten-person service business, the question "is AI profitable" is sharper than it is for a company with a data team. The reasons are structural arguments from first principles, and the evidence in Chapter 2 does not test them. The first reason is that a small firm has no slack. The owner's attention and the office staff's time are the scarce inputs, and an AI project spends them. A large organization can assign a team to a pilot and absorb a failed one. A small firm that spends three months on a failed pilot has also spent three months in which the owner could have been serving customers. The second reason is that a small firm has no analyst. Someone must define what counts as success, collect the records and compare the period before with the period after, and in a small firm that someone is the owner.

Small numbers are a third reason. The average dental practice of one to seven locations books 30 new patients a month in Henry Schein One†'s benchmark13. Suppose an agent adds two of them. That is a 6.7% lift (2 / 30), and two bookings is a difference that ordinary month-to-month variation could produce. Without a longer record the firm cannot tell an effect from chance, and Chapter 7 works out how long a record is needed.

The fourth reason is the shape of the demand. We think much of a service business's demand arrives as a call or a form submission at a moment the customer chose, and a contact that goes unanswered is not queued. The AI product that fits such a business most directly is one that answers contacts, which is consistent with where the search volume in this chapter sits.

A fifth reason is that the cost of an error lands on the owner. A wrong answer from an agent is the firm's wrong answer, given in the firm's name, and a small firm often has no legal or compliance department between it and the customer. Chapter 9 treats that cost as a term in the profit identity. As the cited studies describe their samples, none of the controlled studies in Chapter 2 was run on a small service business. Their numbers show what is possible, not what a small firm should expect.

The thesis: three complements that turn AI into profit

Our thesis is that AI is a general-purpose input whose payoff depends on complements. For a small service business three complements matter most: a flow of demand for the AI to act on, a process for it to sit inside, and measurement that can tell what it earned. A technical marketing system supplies all three. The idea that the return to a general-purpose technology depends on complementary investment is not ours.

Brynjolfsson and Hitt wrote in 2000 that case studies and firm-level evidence suggest organizational investments have a large influence on the value of IT investments14. They also wrote that the benefits of IT are often intangible and disproportionately difficult to measure14. Bresnahan, Brynjolfsson and Hitt found econometric evidence in 2002 that IT use is complementary to new workplace organization and to worker skill15. Bloom, Sadun and Van Reenen, in 2012, combining pan-European firm-level IT data with a management practices survey, found that the US IT-related productivity advantage is primarily due to tougher "people management" practices16.

Those studies concern earlier information technology, not AI. Applying their lesson to AI is an inference, which Chapter 3 sets out and weighs against the economics of the productivity J-curve. The three complements need definitions. Demand is a steady flow of contacts (calls, messages, form submissions and booking requests) from people who want the service, arriving at a point where an agent can act on them. Process is a defined path for each contact: what the agent knows, what it may do, what it must pass on and to whom, how fast, and who receives the hand-off. Measurement is a record that links each contact to its source, its handling and its outcome, so that the firm can estimate what would have happened without the agent. We use the three words in these senses throughout.

The paper is organized around a change in what a website and a search ranking are for. Traditional web design treats a website like a billboard and Google like a phonebook: the site displays the business, the search engine lists it, and what happens after someone looks is left to chance. A technical marketing system turns Google into a funnel and the website into a working rep, which with AI answers questions, books, qualifies and hands off to the office team. The three complements follow from that picture. The funnel is demand, the rep's defined duties are process, and the record of what each contact became is measurement; a billboard supplies none of them. The picture also fixes an order of work, which is framing and plumbing first and the finish last, as in building a house, and Chapter 8 argues for that order.

Installed inside such a system, an agent has work to do, a way to do it and a ledger to be judged by. We argue that installed elsewhere in a small service business, an agent is at best a cost whose return the firm cannot see. That is a thesis: this paper reports no outcome data of our own on deployed agents, and none of the studies it cites tested the thesis directly. A randomized or field study of an AI agent answering inbound calls or scheduling in dental, medical, legal or accounting settings is NF; the nearest peer-reviewed measure is a prospective study at one hospital, without a control group, in which an AI voice assistant made outbound pre-procedure calls and completed 86.4% of the calls in the first of its two phases17. The evidence can show whether the thesis fits the pattern in the literature, and whether the rival explanation that the technology is simply immature leaves things unexplained. Chapters 2 and 3 examine both.

We also have a commercial interest in the thesis. Technical Marketing is the marketing practice of the Strata Modus team. We sell the kind of system this paper describes, including an AI voice agent priced at $299 a month with 250 call minutes included and $0.39 a minute after that18. A reader should weigh the argument with that in mind. The paper uses only cited prices and labeled assumptions, and the measurement designs in Chapter 7 are ones a small firm can run for itself, with or without us.

Terms: AI agent, technical marketing system and profit

AI agent

An AI agent, as the term is used here, is a software system built on a language model. It is given a bounded job, access to facts and tools, and rules about what it may do and when it must hand over to a person. It acts on a contact for the business, such as a phone call, a website chat or a form submission, across more than one step: it answers, asks, checks a calendar, books and records. A chatbot that matches questions to a fixed script sits at one end of the range, and a system that qualifies, quotes and routes sits nearer the other. The paper's main example is the voice agent that answers calls, sold under names such as AI receptionist and AI answering service, and the same definition covers a chat agent on a website.

The words are used loosely in the market: "ai agent" is searched 49,500 times a month in the US, "ai agents for business" 3,600 times and "ai voice agent" 33,100 times5. In our practice, callers and site visitors are always told that they are speaking with an AI assistant. We treat that as part of what makes an agent responsible, and Chapter 9 covers the law on disclosure.

Technical marketing system

A technical marketing system is the set of search presence, website, telephone and messaging, forms, scheduling, customer record and tracking that a service business uses to be found and to turn contacts into customers. It is built as one system and measured as one, with defined hand-offs to people. It does three jobs: it brings qualified demand to the point of contact, defines what happens to each contact, and records the outcome. Those jobs correspond to the three complements above. The phrase is little searched: "technical marketing" draws 210 searches a month in the US5. We use it in the defined sense only.

Profit as used here

Profit, in this paper, is the extra gross profit that can be attributed to the system and its agent over a stated period and against a stated alternative, less every cost of having them. Those costs are the cost of running the agent, the fixed cost of building and maintaining the system, and the expected cost of its errors. We write AI profit for this quantity. Chapter 4 sets out the identity term by term, with four mechanisms by which an agent can add gross profit: capture, conversion, capacity and retention. The system's ledger has a line above those four for the demand that the plan produces. Chapter 4 defines it as dN x c x V: the contacts a month the system adds, times the share of them that become customers, times the value of a customer. A fee counted as a cost has its product counted as a gain, and the plan's product is that demand.

First, profit is a difference against a counterfactual. The profit of an agent is what the firm has less what it would have had without the agent, and the second term is never observed. Second, profit is a dollar amount, not a ratio. Return on investment is the gain less the cost, divided by the cost, and it can be large on a small base. For illustration, a gain of $30 against a cost of $10 is a ratio of 200% and a profit of $20. A gain of $150 against a cost of $100 is a ratio of 50% and a profit of $50 ($150 - $100). The larger ratio is the smaller profit. An owner is paid in dollars, so we use ratios only where we also show the dollars.

Third, profit is not productivity. Controlled studies measure time saved, issues resolved or quality of output, and Chapter 2 reports them. Time saved becomes profit only if it is sold, spent on something that earns, or removes a cost. Suppose a firm saves 10 hours a week and fills none of them with work: its revenue and its payroll are unchanged. That is why capacity is only one of four mechanisms in the identity.

The plan of the paper and the evidence it uses

The paper has 14 chapters. Chapter 2 surveys the best evidence on what AI does to output, including the evidence that cuts against our thesis, and Chapter 3 explains the gap between adoption and return. Chapter 4 sets out the profit identity and Chapter 5 puts numbers on its cost side using published prices and official wages. Chapters 6 and 7 take demand and measurement in turn, and Chapter 8 describes the architecture that supplies all three complements with an agent inside it. Chapter 9 treats legal and reputational risk as a term in the identity, Chapter 10 reads the public record of deployments that failed or were reversed, and Chapter 11 applies the identity to dental, medical, legal and accounting practices. Chapter 12 describes how the system is run month to month, Chapter 13 gives a maturity model with a 90-day and a 12-month plan, and Chapter 14 answers the strongest objections and states what the paper cannot show.

Three kinds of material carry the argument, and the paper keeps them apart. The first is published research. Every statement of fact about the world in the paper is tied to a claim in an evidence file that records its source and how far we could confirm it. We give each figure with its year, its population and its hedges. Where a claim rests on a secondary source or on one that we could not read in full, we say so in the sentence that uses it and rest no conclusion on it alone.

The second kind is our own data. It includes a study that put 199 local questions to ChatGPT, Gemini and Perplexity across 16 US cities and 23 kinds of business on 2026-11-1019, and the keyword data used in this chapter. It also includes a check of the business associate agreement positions of 28 software tools that small practices commonly use20, which Chapter 12 uses, and our published prices18. The recommendation study asked each question once, matched names across assistants by spelling and read up to eight names per answer19. It used answers obtained through an API, which are a baseline, not what every signed-in user sees19. The third kind is argument and arithmetic. Where we reason from first principles we say so, and a worked example with invented numbers is introduced with the word "suppose".

A companion paper, technical marketing for professional practices, covers each profession's advertising and privacy rules in depth, reports our study of AI recommendations in full, and treats the measurement of marketing in general. Nothing in this paper is legal, medical, tax or financial advice. Where it describes a rule, it names the rule and says what it says. The paper makes no promise about rankings, AI answers, leads or revenue. Two of our plans carry a guarantee on a measure agreed after an audit18, and nothing in the paper's argument rests on it.

Chapter 2

What the evidence says about AI returns

AI ROI: what a controlled study can and cannot measure

Evidence about AI sits at three levels, and a finding at one level does not carry up to the next. Capability evidence asks whether a model can do a task. ChatGPT performed at or near the 60% accuracy passing threshold on all three exams of the United States Medical Licensing Examination21. A 2023 study by Ayers and co-authors (secondary) reports that in a blinded comparison licensed healthcare evaluators preferred chatbot responses to physician responses in 78.6% of 585 evaluations22.

Productivity evidence asks what happens to output when people use a model on a real task, and the controlled experiments in this chapter are of that kind. Return evidence asks whether the firm ends up with more money after paying for the tool, for the change in how work is done and for the tool's errors. The failure figures in Chapter 1 are attempts, of uneven quality, at the third level.

The controlled studies are the best evidence we have. They compare outcomes with and without access to the tool, so they isolate its effect on a task better than any survey of use can. What they measure is output per hour, time per task or quality. The cited studies report no cost of the tool, no cost of changing how the work is done and no cost of errors, and nothing on what became of the time saved. They supply a gain, not an ROI. People who search for "ai roi", about 590 times a month in the US5, are asking for a number that the studies in this chapter do not contain.

Four controlled studies: customer support, writing, consulting and software

Customer support: Brynjolfsson, Li and Raymond

In a 2023 study, Brynjolfsson, Li and Raymond examined customer support workers given a generative AI tool. Access to the tool raised productivity, measured as issues resolved per hour, by 14% on average2. The gain was 34% for novice and low-skilled workers, with "minimal impact on experienced and highly skilled workers"2. The authors also report that AI assistance improves customer sentiment, increases employee retention and "may lead to worker learning"2.

First, the outcome is a rate of work, not profit: more issues resolved per hour lowers the labor cost of each issue only if the firm needs fewer hours or handles more issues. A rise of 14% in issues per hour frees fewer than 14% of the hours. A fixed workload needs 0.877 of the hours it needed before (1 / 1.14), which is about 12% fewer hours2. Second, the average covers workers whose gains differ widely, so a team of experienced agents would see less than a team of novices. Third, the authors hedge the retention and learning results themselves, with "may" in the case of learning, and the claims give no dollar value for any of the three secondary results.

Professional writing: Noy and Zhang

In a 2023 paper, Noy and Zhang studied mid-level professional writing tasks. Access to ChatGPT cut the time taken on the post-treatment task by 10 minutes, or 37%, against a control group that took an average of 27 minutes23. The 10 minutes are 37% of 27 (10 / 27). Time taken fell by 0.8 standard deviations and output quality rose by 0.423. Inequality between workers decreased, because ChatGPT compressed the productivity distribution by benefiting low-ability workers more23.

The control group finished in 27 minutes on average, so the unit of work is short and self-contained. The cited claims say nothing about a working week in which writing is one task among many. Time saved on a 27-minute task is saved in any useful sense only if the worker has another task ready to start.

Consulting tasks: Dell'Acqua and co-authors

A 2023 study by Dell'Acqua and co-authors tested consultants using AI. For tasks inside the technological frontier, the consultants completed 12.2% more tasks on average and completed tasks 25.1% more quickly than a control group24. Their results were of more than 40% higher quality than the control group's24. Consultants below the average baseline performance gained 43% against 17% for those above it, each compared with their own scores24. The cited claim does not say how quality was scored, so the quality gain of more than 40% is a figure on the study's own scale24.

Software development: Peng and co-authors

Peng, Kalliamvakou, Cihon and Demirer, in a 2023 paper, studied developers who had access to GitHub Copilot, an AI pair programmer. The treatment group completed an HTTP server implementation task 55.8% faster than the control group25. The cited claim reports one task, and a task with a definite end point is easy to measure but a thin basis for a statement about a whole job.

StudyTaskOutcome measuredHeadline effectWho gained mostClaim
Brynjolfsson, Li and Raymond, 2023Customer supportIssues resolved per hour14% on averageNovice and low-skilled workers (34%); minimal impact on experienced and highly skilled workers2
Noy and Zhang, 2023Mid-level professional writingCompletion time and output qualityTime down 10 minutes (37%) from 27; quality up 0.4 standard deviationsLow-ability workers; inequality between workers decreased23
Dell'Acqua et al., 2023Consulting tasks inside the frontierTasks completed, speed and quality12.2% more tasks, 25.1% more quickly, quality more than 40% higherBelow-average performers (43% against 17%)24
Peng et al., 2023HTTP server implementationCompletion time55.8% fasterNot reported in the cited claim25

Table 1. Four controlled studies of generative AI at work.

Percentages that look alike are not alike

The headline numbers in Table 1 do not rank the studies, because they are not the same kind of number. Brynjolfsson, Li and Raymond report a rate and Noy and Zhang a duration. Dell'Acqua and co-authors report "more quickly" and Peng and co-authors "faster", and the cited claims do not say whether those percentages apply to time or to speed. The difference is large: a 37% cut in time is a rise in output per hour of about 59% (1 / 0.63 = 1.59).

That ambiguity changes the 55.8% a great deal25. If "faster" means 55.8% less time, output per hour is 2.26 times the control group's (1 / 0.442). If it means 55.8% more speed, time falls by 35.8% (1 - 1 / 1.558). Comparing the 14% in the support study2 with the 55.8% in the software study25 would set a rate against a figure whose meaning the claim leaves open. We do not make that comparison.

Who gained most, and what that implies for a small firm

Three of the four studies report who gained most (Table 1), and they agree on the direction: gains were largest for the workers at the lower end2,23,24. The fourth study's claim is silent on the point.

The agreement is looser than it looks. The studies define their groups differently, and in all three, people used the tool as an aid to their own work. That is how the studies were designed, and it limits what the results say about the subject of this paper. An agent that answers a call does the work itself.

That direction cuts against a simple version of our case. Where the gain goes to the less experienced, an owner-operator who already does the task well would be expected to gain least from a tool that helps with that same work. The results do not speak to a different use, in which an agent takes contacts that no one was going to take, such as the call that arrives when the office is closed. That use is the capture mechanism of Chapter 4, and none of the controlled studies tests it.

Where AI made work worse: the jagged frontier

The study by Dell'Acqua and co-authors that produced the consulting row of Table 1 also included a task selected to lie outside what the AI does well. For that task, consultants using AI were 19 percentage points less likely to produce correct solutions than consultants without AI24. The paper's title calls the technological frontier jagged. As we read the idea, the boundary between tasks that AI handles well and tasks it handles badly is irregular. Two tasks that look alike to a user can fall on opposite sides of it.

First, the authors selected the task, so the 19 points show that such tasks exist and what happened on one of them, not how common they are in real work. Second, the claim gives the size of the drop but not its mechanism; we infer that the harm came from wrong answers that looked right. Third, the results do not point one way. The tool that produced results of more than 40% higher quality on tasks inside the frontier lowered the chance of a correct answer on a task outside it24. The two outcomes are measured differently, so we do not subtract one from the other.

Accuracy studies of tools tested without a person in the loop show another side of the same problem. Stanford researchers found legal hallucinations occurring 58% of the time with ChatGPT 4 and 88% of the time with Llama 2 when the models were asked specific, verifiable questions about random federal court cases26. A second study found that commercial AI legal research tools each hallucinated between 17% and 33% of the time27. An audit of commercial generative search engines found that, on average, 51.5% of generated sentences were fully supported by their citations and that 74.5% of citations supported their associated sentence28.

These are error rates of tools tested alone, on tasks where a statement can be checked against a source, not productivity effects of tools used by people. We include them because they show the other half of the frontier in a domain where a wrong answer has a price. Such an error rate can be found only by checking answers against sources, and a firm that does not check will not see its own. Chapter 9 treats the expected cost of error as a term in the profit identity.

AI profitability at the level of firms and the economy: small measured effects

Acemoglu, in a 2024 paper, estimates the macroeconomic effects of generative AI over 10 years as "nontrivial but modest", no more than a 0.66% increase in total factor productivity29. The estimate is a calculation, not an observation. As the same sentence puts it, it uses existing estimates on exposure to AI and productivity improvements at the task level29. Table 1 reports results of the second kind. Spread evenly over 10 years, the ceiling of 0.66% is about 0.07% a year (0.66 / 10).

Task gains of the sizes in Table 1 can still yield an economy-wide figure below 1%. For illustration, suppose a role spends 20% of its hours on tasks that AI affects and AI cuts the time on those tasks by 37%, the figure from Noy and Zhang23. The saving is 7.4% of the role's hours (20% x 37%). Suppose instead that the exposed share is 5% of hours: the saving is 1.85%. Neither exposure share is a finding. The product of the exposed share and the saving on exposed work is small even when the saving is large, and it counts only if the freed hours are put to use.

We think a second reason an aggregate can be small, one that Acemoglu does not state, is the share of firms that use the tools at all. The Census series reported 3.8% of businesses using AI to produce goods and services at its first reading6 and overall use of 17% to 20% for December 2025 to May 20261. Chapter 1 explains why we do not treat the two as one series. An effect on adopters shows up in an aggregate only in proportion to how many adopt.

Two results across organizations speak to return directly, and they point in opposite directions. The first is the MIT Project NANDA finding that 95% of organizations are getting zero return3. The second comes from a survey by Thomson Reuters, a publisher that sells AI products. More than half (53%) of the professionals surveyed said their organizations were already seeing ROI as a direct or indirect result of their AI adoption, and only 22% said their business had a visible, defined AI strategy4. Unlike the experiments, neither isolates a tool's effect on a task. Each reports return at the level of the organization, the first as a research finding and the second as what respondents say, and they count different populations, so we do not average them.

A figure of zero return cannot tell a return that did not occur from one that the organization could not see. Brynjolfsson, Rock and Syverson give a reason to take the second possibility seriously: they argue that unmeasured complementary intangible investments lead to initial productivity slowdowns followed by later surges30. The NANDA figure does not separate the two kinds of zero3. Chapter 3 returns to the problem, and Chapter 7 to how a small firm can measure.

Humlum and Vestergaard report, in an April 2025 working paper titled Large Language Models, Small Labor Market Effects, that AI chatbots had no significant impact on earnings or recorded hours in any occupation31. The confidence intervals ruled out effects larger than 1%31. The same paper reports modest productivity gains, with average time savings of 3%, combined with weak wage pass-through31. The revised version, retitled Still Waters, Rapid Currents, estimates precise null effects on earnings and recorded hours and rules out effects larger than 2% two years after the launch of ChatGPT32. The two versions state different bounds, and we do not merge them. These are null results on earnings and hours, not on any firm's profit, and a ceiling calculated from task results says nothing by itself about whether any single firm earned a return.

Failure, abandonment and zero return: what the figures can support

RAND's contribution in our evidence is the pattern in its interviews, not the 80%11. Interviews with experienced practitioners are a good way to learn why projects fail and a poor way to learn how many do. The RAND report says that five dominant root causes emerged across all of its interviews11.

If the estimate of more than 80% is twice the failure rate of information technology projects that do not involve AI, the comparison implies a failure rate of more than 40% for those projects (half of the 80%)11. That would itself be high, and it would suggest that failure is common in technology projects generally. The claim does not say that both rates use one definition of failure or how either was estimated, so the inference is conditional and we rest nothing on it.

Gartner's prediction (secondary) gives four stated reasons for abandonment in one sentence: poor data quality, inadequate risk controls, escalating costs and unclear business value12. We think unclear business value is the reason that measurement addresses directly, because a firm that cannot say what a project earned cannot show its value. Poor data quality and inadequate risk controls concern the process the AI sits inside, and escalating costs is the subject of Chapter 5. Gartner gives no weights. None of the four reasons is that the model could not do the task.

The NANDA report describes outcomes as starkly divided across buyers3. The other 5% (calc.) are the organizations outside that group, and they are the ones a small firm most needs to learn from. The cited claims say nothing about how they differ from the rest, and the thesis of this paper is a hypothesis about exactly that difference. The nearest thing to an answer is the Thomson Reuters survey. It reports that organizations with visible AI strategies were twice as likely as those with informal or ad hoc adoption to experience revenue growth as a direct or indirect result of AI adoption4. That is an association between what respondents say about strategy and what they say about growth, and a survey cannot show which came first. We think it is consistent with the thesis and not a test of it.

None of the three figures gives a firm the probability that its own project will fail, because their units, definitions and populations differ and in places are not given. The reasons Gartner is reported to name do give a firm a list of questions to answer before it buys: whether the data are good enough, whether the risks are controlled, whether the costs are bounded and whether the value can be seen.

What the studies have in common: our reading

We read the evidence as showing that gains appear where the task is frequent, the output can be checked and the work sits inside a measured process. The pattern is drawn across the studies above, and none of them tested it.

A frequent task is one that the worker does many times, so that a small gain per repetition accumulates and shows up quickly in the data. The support study measures a rate, issues resolved per hour, which presupposes work that recurs2. The cited claims for the other three studies do not say how often the task is done in practice. A checkable output is one that can be judged against a standard cheaply and soon: an issue is resolved or not, a solution is correct or not. A measured process is one in which outcomes are recorded for each worker or task, so that a difference can be seen at all.

Some of the evidence is consistent with the reading. Every study that found a gain reports a measured outcome, which is what a measured process provides. The one controlled result in which AI-assisted work was worse is about correctness24, and the legal error rates were measured on specific, verifiable questions26. Gartner's reasons for abandonment include unclear business value12, which is what a missing measurement would produce. And the older studies of information technology suggested that organizational investments had a large influence on the value of IT14. They also found econometric evidence that IT use was complementary to new workplace organization15.

Against that, the first problem is selection. The studies were built to measure a gain, so each has a measured outcome, and they cannot show whether gains would have appeared without measurement. A literature made of measured tasks will always look as if gains need measurement, because unmeasured gains, if they exist, are by definition missing from it. The J-curve argument holds that unmeasured complementary intangible investments can be followed by later increases in productivity30.

The second is that no study varied frequency, checkability or measurement on its own, so the pattern is an association across four studies that differ in many other ways. The third is that the failure figures do not report whether failed projects lacked frequent tasks, checkable outputs or measured processes, so they cannot confirm that the pattern predicts failure. The fourth is that skill also mattered, with gains smallest for the most experienced workers in the support study2, and the reading says nothing about that.

For a small service business, incoming calls are frequent in aggregate, though a given firm's volume sets how frequent. A booking can be checked, because it is in the calendar or it is not, and a recorded call can be read. The measurement exists only if a system links each call to its source and its outcome, which is the third complement defined in Chapter 1. The first two conditions are properties of the task, and the third is a property of the system around it, which a firm can build or fail to build.

We treat the reading as a working hypothesis that organizes the rest of the paper, and it makes a prediction that could fail. An agent in a small firm with no flow of contacts, no defined process and no measurement would be expected to show no demonstrable return. An agent in a firm with all three would be expected to show gains of the kind the studies report. A field experiment could test that, and Chapter 14 includes one in the research agenda.

Whether a small service business has enough contacts for the task to count as frequent depends on the business. In health care, Call Box† reports an average of 299 calls per office location in February 202033. Patient Prism† (call scoring) counted 11,552,668 calls at 8,280 dental locations in 202534, which is about 116 a month per location (11,552,668 / 8,280 / 12). A measurement of call volumes at law and accounting firms is NF; the nearest evidence is the health care and dental counts above.

What the evidence does not yet show for AI for small business

The evidence in this chapter leaves six gaps for a small service business. As the cited studies describe their samples, none was run on one. All four controlled studies test AI as an aid to a person, and none tests an agent that handles contacts alone, which is the subject of this paper. They report no cost of the tool, of changing the work or of errors, so none can supply a profit. None counts demand either, the line that sits above the agent's four gains in the system ledger of Chapter 4 (dN x c x V): they study workers who have tasks and ask how well the tasks are done, not where the tasks come from. The Census figures count use, not return, and this paper reports them as national totals and for one sector, without a breakdown by firm size. And the usage studies cited in Chapter 1 each describe the use of one vendor's assistant, without showing whether any of the use sat inside a business process with a recorded outcome.

Surveys inside the professions count use too. The AMA's 2026 physician survey found that 81% of physicians reported awareness or use of AI in their practice, and its list of uses has no phone or front-desk use35. In a February 2026 MGMA poll with 177 applicable responses, scheduling (31%) and calls (27%) ranked first and second as the areas where leaders were turning to AI and automation to improve access36. That states a priority, not a deployment. For accounting, Intuit's sponsored 2026 survey of 725 accounting and bookkeeping professionals reports that 88% used AI for at least one client service in the last 12 months, which is a low bar37.

One company's announcement (secondary) describes an AI assistant handling customer conversations directly. Klarna reported that in its first month its assistant had 2.3 million conversations, two-thirds of its customer service chats, and was doing the equivalent work of 700 full-time agents38. It reported that customers now resolve their errands in less than 2 minutes against 11 minutes before38. These figures are the company's own statements, not an independent measurement. Taken at face value, they describe a very large flow of contacts inside a process that the company could measure. That is consistent with our reading but does not test it for a firm with far fewer contacts. Chapter 10 reads the company's statements further.

Two kinds of study would settle more than anything above. Field experiments with small service firms, in which some firms or some hours receive an agent and others do not, would give the effect on contacts and bookings. Independent reporting of the share of contacts an agent completes without a person, and of the share that become bookings, would let a firm compare itself with others.

A randomized or field study of an AI agent answering inbound calls or scheduling in dental, medical, legal or accounting settings is NF, and so is a peer-reviewed containment rate, which is the share of calls finished without a person. The nearest peer-reviewed measure is a study at one hospital, without a control group, of an AI voice assistant making outbound pre-procedure calls. It completed 86.4% of its calls in the first phase, and 36.6% of the calls in that phase were fully automated and needed no human intervention17. Relatient† (patient scheduling) said its Voice AI product answered 62% of inbound calls, of which 72% were either autonomously handled or escalated39. That count includes calls passed to staff and is not containment39.

Audits of AI answers in regulated advice exist, and the cost term in Chapter 9 needs more of them. The audit of legal research tools above used 202 queries, which its authors call small, on early-2024 products, so it is not a current error rate40. A randomized Nature Medicine study of 1,298 UK participants shows why an audit of a model alone is not enough. Tested alone, the models correctly identified the underlying condition in 94.9% of 10 medical scenarios41. Participants using the same models identified relevant conditions in fewer than 34.5% of cases, no better than the control group41. The error rate of a person using an answer, which is what a firm's customers do, can differ from the error rate of the answer. Chapter 14 sets out this agenda.

Chapter 3

Why returns lag: complements, the J-curve and measurement

The AI productivity paradox: Solow's remark and the gap between use and return

In 1987 Robert Solow wrote in the New York Times Book Review: "You can see the computer age everywhere but in the productivity statistics."42 Our evidence file marks the remark as plausible and not confirmed, so we quote it only as the file prints it, and the file describes it as the articulation of the productivity paradox. We use the paradox in a plain sense: a technology is visible in daily work, and the figures that should show its payoff do not. Chapters 1 and 2 left the same puzzle for AI. Several instruments report use, controlled studies document task-level gains, and a return at the level of the firm or the economy is hard to find.

Resolving a paradox of this kind means finding which measured half is wrong or incomplete. Either the technology does not do what the task-level studies suggest, or a return exists and the statistics do not capture it, or the return comes later than the statistics look. The economics of general-purpose technology gives reasons to take the second and third possibilities seriously. A general-purpose technology, as we use the term, is one that many kinds of work can use. Brynjolfsson, Rock and Syverson put AI among such technologies, and their argument about them is the subject of this chapter43.

This chapter reports what the literature on complementary investment, the productivity J-curve and data-driven management found. Several of the studies are held only as abstracts, and for those we say what the abstract says and no more.

Complementary investment: organization, process and skills in AI implementation

Two inputs are complements, in the economist's sense, when more of one raises the return to the other. The evidence for complements to information technology (IT) comes from firm-level studies of technology adopted before AI. In 2000, Brynjolfsson and Hitt reported that case studies and firm-level econometric evidence suggest that organizational investments have a large influence on the value of IT investments14. They also reported that the benefits of IT investment are often intangible and disproportionately difficult to measure14. These two findings carry this chapter's argument: the payoff depends on something other than the technology, and part of it does not show up where a firm would look.

Bresnahan, Brynjolfsson and Hitt, in a 2002 paper whose title calls its evidence firm-level, found that IT use is complementary to a new workplace organization that includes broader job responsibilities for line workers, more decentralized decision-making and more self-managing teams15. They also found that both IT and that organization are complements with worker skill, measured in a variety of ways15. Complementarity changes the unit of analysis: when inputs are complements, the thing that earns a return is the package. A study that sets the technology's cost against the firm's output will credit the technology with a result that belongs to the package, or find nothing because the package is incomplete.

Time is the second thing complements change. Brynjolfsson and Hitt used data on 527 large U.S. firms over 1987 to 199444. They found that the productivity and output contributions associated with computerization are up to 5 times greater over long periods, measured as differences over 5 to 7 years, than over one-year differences44. The authors link the gap to large, time-consuming investments in complementary inputs such as organizational capital44. If the ratio of 5 held, a one-year comparison would show as little as a fifth of the long-run contribution (1 / 5 = 0.2). The abstract reports large firms and computerization, and we do not carry its ratio to any other setting.

A fourth study ties the pattern to management. Combining pan-European firm-level IT data with their own management practices survey, Bloom, Sadun and Van Reenen found that the US IT-related productivity advantage "is primarily due to its tougher 'people management' practices"16. The study's title, Americans Do IT Better, indicates that it concerns US multinationals. The abstract does not give the number of firms. It adds that a complement can be a way of managing people and need not be something a firm buys.

The literature we cite speaks to organization, skills and management, and it does not map directly onto the three complements in our thesis. The process an AI agent sits inside has its nearest counterpart in what the studies call workplace organization and management practice. Measurement appears in them as data-driven decision making, covered below. The first of our three, a flow of demand for the technology to act on, is not in the cited studies at all. It is this paper's own addition, argued in Chapter 6, and the complements literature does not report it.

Scale is a further caution. Self-managing teams and decentralized decisions are features of organizations with many people, and a firm of ten may have no teams to make self-managing. We think the analogue there is smaller and plainer: a written path for each contact, a named person who receives each hand-off, and one calendar that holds the truth. The studies do not make that translation.

The productivity J-curve and AI ROI: why measured returns may dip before they rise

The productivity J-curve is a model of what the previous section implies for timing. Brynjolfsson, Rock and Syverson write that their model "generates a Productivity J-Curve that can explain the productivity slowdowns often accompanying the advent of GPTs, as well as the increase in productivity later"30. They argue that general purpose technologies such as AI "enable and require significant complementary investments, including co-invention of new processes, products, business models and human capital"43. The same authors describe these investments as often intangible and poorly measured in the national accounts43. We infer, from the logic of the model and no cited claim beyond it, that the unmeasured investment makes the early years look poor.

Suppose a firm spends the first months of a new technology writing rules for it, changing who does what and training people. The hours come out of work that would otherwise have produced output, and no asset appears in the accounts, because rules and habits are not on the balance sheet. Output measured in the usual way stalls while the building goes on, and when the rules begin to pay and output rises, nothing in the accounts records that the rise was paid for earlier. The early period looks worse than it was, and the later period looks better than the technology alone could explain. The authors make this argument for economies; we apply the same logic to a firm's own books below, labeled as an application.

Those authors also size the gap between an adjusted and an official measure of productivity in the aggregate. In the version of the paper we read, their adjusted measure of total factor productivity is 11.3% higher than official measures at the end of 2004 and 15.9% higher at the end of 201743. The published version may state different numbers. We think the adjustment concerns the unmeasured investment just described, though the cited claim does not say so. The comparison with "official measures" is an economy-wide calculation from a model, a statement about national statistics. It shows that the gap between official and adjusted productivity can be large, and it says nothing about the size of any such gap inside a firm.

Applied to one firm, the logic yields a timeline an owner could recognize. Table 1 sets out a supposed deployment in three phases. A door, in this paper, is any way a contact arrives, such as a phone number, a form or a chat window; the pipeline record is defined in a later section. The phases illustrate order and say nothing about how long each lasts in a small firm. The horizon in the cited large-firm data is the 5 to 7 year difference over which it showed the larger contribution44.

PhaseWhat the owner's books showWhat has actually happenedWhat a pipeline record adds
SetupA new monthly fee. The owner's hours are on no pageRules written, calendar connected, doors tagged, first calls readA baseline: contacts by door, answered or not, bookings by source
LearningThe fee, with bookings from the agent merged into total revenueSome contacts held, some errors found and fixed, rules revisedBookings labeled by source, contacts held and handed off, errors logged
Steady useRevenue flat or up, with no line that says whyThe new process is in use and the staff have adapted to itA comparison of the period against the baseline, term by term

Table 1. A supposed three-phase deployment and what a firm can see at each phase. The phases illustrate order and carry no duration.

An owner who must decide at the end of the learning phase whether to keep paying has the least information of the three phases, because the bookings the agent produced sit in the same revenue line as everything else. Chapter 2 noted that a zero-return figure, such as the 95% of organizations getting zero return in the MIT Project NANDA report, cannot distinguish a return that did not occur from one the organization could not see3. The J-curve gives a reason to expect the second case in the early period, and the pipeline record lets a firm tell the two apart.

Data-driven decision-making and management practice as complements

Two lines of work in our evidence ask what kind of firm gets more from its inputs: one concerns the use of data in decisions, the other the quality of management. We use data-driven decision making (DDD) to mean basing decisions on recorded data and analysis and not on recollection. Of the complements in this chapter, it is the closest to measurement.

Brynjolfsson, Hitt and Kim used survey data on the business practices and information technology investments of 179 large publicly traded firms45. They found that firms adopting DDD have output and productivity 5% to 6% higher than would be expected given their other investments and IT usage45. Because the 5% to 6% is measured above what the firm's other investments and IT usage would predict, we think DDD is a complement separate from the technology itself, though the finding is an association in survey data and not a causal estimate45.

Brynjolfsson and McElheran used data collected by the Census Bureau for a large representative sample of US manufacturing plants46. They found that DDD rates nearly tripled, from 11% to 30%, between 2005 and 201046. The same paper reports that the key correlates of adoption are size, the presence of potential complements such as information technology and educated workers, and firm learning46. The abstract does not give the direction of the size effect, and the sample is plants that make things, not practices that serve clients. We infer that the second finding is consistent with data use depending on complements, so a firm without them may find it hard to start.

Bloom and Van Reenen collected management-practice data from 732 medium-sized manufacturing firms in the US, France, Germany and the UK. They found management practices strongly associated with firm-level productivity, profitability, Tobin's Q, sales growth and survival rates47. They also found significant differences in management quality within each country, with a long tail of extremely badly managed firms47. In the paper's estimates, a move from the lower to the upper quartile of management scores between firms, a difference of 0.971 points, is associated with an increase in total factor productivity47. Under ordinary least squares the increase is between 3.2% and 7.5%47.

Table 2 lists the studies. Apart from the J-curve model, which names AI but is not a study of any firm's deployment, every study concerns technology or practices that predate AI. Where the cited claims give the sample, it is large or medium-sized firms or manufacturing plants, and for the studies whose abstracts give none we cannot say. None of them describes a service business of a few dozen people or fewer, which is why the next section is needed.

StudySample and scopeWhat it foundClaim
Brynjolfsson and Hitt, 2000Case studies and firm-level econometric evidence; sample size not in the abstract we read; ITEvidence suggests organizational investments have a large influence on the value of IT investments; IT benefits are often intangible and hard to measure14
Bresnahan, Brynjolfsson and Hitt, 2002Firm-level evidence; number of firms not in the abstract we read; ITIT use is complementary to a new workplace organization, and both are complements with worker skill15
Brynjolfsson and Hitt, 2003527 large US firms, 1987 to 1994; computerizationContributions up to 5 times greater over 5 to 7 year differences than over one-year differences44
Bloom, Sadun and Van Reenen, 2012Pan-European firm-level IT data with a management practices survey; US multinationals by the titleUS IT-related productivity advantage primarily due to tougher "people management" practices16
Brynjolfsson, Rock and Syverson, 2021A model; AI named among general purpose technologiesA productivity J-curve; complementary investments are required30,43
Brynjolfsson, Hitt and Kim, 2011Survey data on 179 large publicly traded firmsFirms adopting DDD show output and productivity 5% to 6% above expectation45
Brynjolfsson and McElheran, 2016Census Bureau data, a large representative sample of US manufacturing plantsDDD nearly tripled, 11% to 30%, from 2005 to 2010; adoption correlates with complements46
Bloom and Van Reenen, 2007732 medium-sized manufacturing firms in four countriesManagement practices strongly associated with productivity, profitability and survival47

Table 2. The studies of complements, with the sample and scope of each. None concerns AI deployed in a small service business.

From large firms to AI for small business: the step this paper takes

The paper makes three moves that none of the studies in Table 2 makes: from earlier IT to AI, from large and medium-sized firms and manufacturing plants to small service businesses, and to a named set of complements that matter in such a firm.

  1. In earlier IT, the studies in Table 2 associate the return with complements, and one reports a larger contribution over longer horizons.
  2. On the J-curve authors' account, AI is a general-purpose technology of the same kind, so its return in a firm should depend on complements too43.
  3. In a small service business, the complements that matter are a flow of demand, a process for each contact and measurement. This is our argument, and Chapters 6, 8 and 7 develop demand, process and measurement in turn.

Each move is an inference, and each can fail. Table 3 states what would have to be true for the step to hold. This paper tests none of the five conditions, and the last column names what would.

ConditionWhy the argument needs itHow it could failWhat would show it
Task-level gains turn into profit only with complementsThe mechanism in Table 2 is about returns to the firm, and Chapter 2's controlled studies are about output on a taskA firm might profit from an agent with no process or record, because the tool alone moves the numbersFirms using the same agent, with and without a defined process and a record, compared on profit
A small firm has complements of the same kind, at smaller scaleThe studies' organization and skills must have an analogue in a firm of tenWhat mattered in large firms may be specific to scale, leaving nothing in a small firm that plays the roleField trials in small firms that vary the process and the record
The complement causes the return and does not merely mark good managementThe studies report associations, and well-run firms may do everything wellA record installed in a poorly run firm adds nothing, because the cause lay in the managementA trial that assigns the record to firms at random
The lag is short enough for a small firm to bearA firm with no slack cannot wait for yearsIf a small firm needs a horizon as long as the 5 to 7 year differences over which the large-firm data showed the larger contribution44, it runs out of time before the return arrivesThe duration of the dip in small-firm trials
A record is cheap enough to build and accurate enough to useWe think the record is the cheapest complementStaff may not enter it, sources may be mislabeled, and a poor record misleadsThe share of contacts with a complete record after six months

Table 3. Conditions for carrying the large-firm findings to AI in a small service business.

The third condition deserves the most attention, because the literature we cite is least able to meet it. As far as the abstracts we read show, each study in Table 2 reports an association or builds a model, and none says it assigned a complement to firms at random. A reader who thinks that well-managed firms both measure and prosper will read the results as saying that management quality explains returns, and nothing in the cited claims refutes that reading. If it is right, a spreadsheet installed by a poorly managed firm will not raise its returns, and we cannot rule that out. The argument in the next section does not need the record to cause the return. It needs only that without a record the return cannot be seen, which is a weaker claim.

Marketing attribution first: the cheapest missing complement, in our judgment

This section makes the paper's central argument. We think that for a service business the cheapest missing complement is a measured customer pipeline, meaning a record that links each contact to its source, its handling and its outcome. Chapter 7 gives the data model. A firm that cannot say where its last ten customers came from cannot say what an AI agent added.

The claim can be tested at a kitchen table. Take the last ten customers who booked. For each, the owner either can read from a record how the customer found the firm and what happened between first contact and booking, or recalls it, or guesses, or cannot tell. The test is ours, not a validated instrument, and ten customers is too few for any statistical statement. It checks whether a record exists, which is the question that matters first. If the honest answers are mostly guesses, the agent's customers are among the ten and cannot be told from the others.

Profit from an agent is a difference against what would have happened without it, as Chapter 1 defined profit. To estimate a difference, a firm needs to know how many contacts arrived, from where, and what each became. An agent leaves a trail of its own in transcripts and logs, but the trail ends at the call. Whether the caller booked, came, paid and returned is held in the calendar and the customer record, and unless a field links the two, the agent's work and the firm's revenue sit in different places.

We think the pipeline is the cheapest missing complement, for four reasons. First, the other complements consume the owner's time, which Chapter 1 described as one of the scarce inputs: reorganizing a front desk, retraining staff and integrating systems all take the owner's attention, while a source tag on each door and an outcome column on each contact are small by comparison. Second, a descriptive record needs a field and a habit, not a redesign. Third, the record pays whether or not an agent is bought, because it tells the firm which doors bring customers, whereas the reorganization of a front desk pays chiefly in combination with the tool. Fourth, the record is the precondition for judging every other complement, since a firm that changes its process and has no record cannot tell whether the change helped. This paper does not price the complements for a small firm, so the comparison is of kinds of effort.

Measurement is also the complement for which the cited studies supply the nearest evidence. The DDD result is an association between recorded data in decisions and output above what IT predicts45. A pipeline record is the smallest case of data-driven decision making a service business can run. We make no claim that 5% to 6% carries over to such a firm45. In our keyword data, "marketing attribution" is searched about 720 times a month in the US and "incrementality testing" about 4805. The phrases are ones a marketer or vendor would type, and the volume says nothing about how many small service firms hold a record.

Seeing which doors brought customers is a descriptive task. Estimating what an agent caused is a causal one, and it is hard even for large firms. Lewis and Rao reported twenty-five large field experiments with major US retailers and brokerages48. They found the median confidence interval on the return on investment of advertising to be over 100 percentage points wide48. They add that informative advertising experiments can easily require more than 10 million person-weeks, which makes them costly and potentially infeasible for many firms48.

Gordon and co-authors, in 15 experiments at Facebook, found that observational methods often fail to reproduce what the randomized experiments found, even after conditioning on extensive demographic and behavioral variables49. Both findings concern digital advertising, not agents on a phone line. We think they warn that a record by itself will not settle a causal question, and Chapter 7 describes the designs a small firm can run.

The other two complements are not cheaper. A process for each contact takes decisions from the owner about what the agent may say and to whom it hands off. A flow of demand takes continuing marketing work and money, and an agent attached to no flow of contacts has nothing to act on. We do not say the pipeline matters most. All three are needed, and the pipeline is the one a firm can begin without spending anything on the agent.

The objection that the technology is simply immature

A strong objection to this chapter's reading is that returns lag because the technology is not yet good enough and will come as models improve. Part of it is certainly right. The four controlled studies in Table 1 of Chapter 2 date from 2023, and the accuracy studies from 2023 and 2024. Many of the figures in Chapter 2 therefore describe tools that are two or more years old as we write, in November 2026. If the tasks that matter to a service business lie outside the frontier of today's models, no complement can repair that, and the sensible course on those tasks is to wait.

Error figures show how far the frontier still was in the cited studies. Consultants using AI were 19 percentage points less likely to produce correct solutions on a task selected to lie outside the frontier24. Legal hallucinations occurred 58% of the time with ChatGPT 4 and 88% of the time with Llama 2 when the models were asked specific, verifiable questions about random federal court cases26. The commercial AI legal research tools tested each hallucinated between 17% and 33% of the time27. In an audit of commercial generative search engines, on average only 51.5% of generated sentences were fully supported by their citations28. The last three are error rates of tools tested alone, and the first is a result for people using a tool. A better model could lower all of them, and the objection would then be right about those tasks.

Two field experiments bear on contact handling directly and cut toward the objection. At a large telecommunications company, introducing a voice-based AI system to replace the phone menu temporarily increased machine-service duration and customers' demand for human service, and persistently reduced customer complaints50. On Alibaba's Taobao platform, a randomized experiment in text chat found that an agentic AI system reduced average chat duration and had limited effects on retrial rates, but substantially lowered ratings for AI-eligible chats51. The second is a 2026 arXiv preprint. Both concern an AI system taking customer contacts, which is closer to this paper's subject than the assisted-worker studies, and each shows a cost alongside any gain. We also cannot say how much of the zero-return figure in the MIT Project NANDA report, 95% of organizations, is immaturity, since the cited claims for that report give no causes3.

Even so, the evidence counts against immaturity as the whole explanation. First, the tools already produced measured gains on tasks. Access to a generative AI assistant raised customer support productivity by 14% on average2. In mid-level professional writing it cut completion time by 37%23, and on tasks inside the frontier consultants completed tasks 25.1% more quickly24. Those are gains from AI used as an aid to a person, so they do not show that an unassisted agent works. They do show that the tools moved output on tasks, while firm-level returns were still hard to find in the later reports that Chapter 2 describes. An account that says only that the tools were not good enough has to explain why tools that moved output on tasks have not been shown to move returns.

Second, Gartner gave a list of causes. It predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 202512. It named poor data quality, inadequate risk controls, escalating costs and unclear business value12. This is an analyst firm's prediction (secondary), and this paper reports no check on whether it came true. None of the four causes is that the model could not do the task.

Third, the older record: the contribution of computerization in the large-firm data grew with the horizon, and the authors link the growth to complementary inputs44. We cannot say how much of the difference was the technology improving, so the record does not decide between the two readings. It shows that those who measured the contribution for an earlier technology attached an explanation other than immaturity.

Fourth, the profit identity that Chapter 4 sets out shows an asymmetry: better models act on some of its terms and not on others. They can raise the share of missed contacts that the agent answers and holds (a) and the share of answered contacts that become bookings (c). They can lower the run cost per contact (k) and the expected cost of error (E). They cannot supply the inbound contacts (N) or the contacts a month that the plan adds to them (dN). The system's ledger carries dN as a line above the agent's four gains, written dN x c x V in Chapter 4. They cannot change what a new customer is worth (V) or record the outcome for the owner. Capture is a product of its factors, so raising one factor leaves the product at zero while another factor is zero. Table 4 sets the two explanations side by side.

ObservationIf the technology is immatureIf complements are missingWhat the cited evidence shows
Gains on tasks in controlled studiesSmall or absentPresent on the task, not always in profitGains on tasks, as an aid to a person: 14% more issues resolved per hour2; time cut by 37%23
Errors on tasks outside the frontierFall as models improveExist, and are absorbed by review and hand-offA 19 percentage point drop in correct solutions24; legal hallucination 17% to 33% of the time in research tools27 and 58% to 88% of the time in two general models26
Return differs between firms using the same toolSmall differencesLarge differences tied to process and recordNo controlled comparison is cited here; the claims for the report behind the 95% figure3 do not say how the organizations outside that 95% differ. A survey by Thomson Reuters† (a publisher that sells AI products) reports that organizations with visible AI strategies were twice as likely as ad hoc adopters to see revenue growth, a self-reported association4
Reasons given for abandonmentCapabilityData, controls, cost and valueGartner's four causes, none of them capability (a prediction; secondary)12
Return and the length of timeRises as models improveRises as complements are builtContributions up to 5 times greater over 5 to 7 year differences, linked by the authors to complements44

Table 4. What each explanation predicts, and what the cited evidence says.

A second source, a press release from Clio† (practice-management software), leans toward the second column and tests neither. It says that firms with wide AI adoption were nearly 3x more likely to report revenue growth than firms that had not adopted AI52. It also says that 77% of firms that increased revenue with AI attributed it to improved operations such as document generation, workflow automation and client communication52. Both compare groups of firms by what they say about themselves, so they cannot show whether adoption made firms grow or growing firms adopted more. The attribution is the firms' own.

The evidence cannot separate the two explanations, and they are not exclusive: both can be partly true, and slow now and faster later is what each predicts. Our argument does not depend on winning the contest. Suppose the objection is right. A firm that has built a pipeline record will see when the tools cross the line from costing money to earning it, and a firm that has not will see nothing. Suppose the complement reading is right. The firm with the record holds one of the complements, and the firm without it is missing one. The firm with no record loses under both, so we infer the record is worth building before the agent whichever story is true.

Three results would lead us to revise the reading. The first is a trial in which small service firms receive the same agent, some with a defined process and a record and some without, and the groups earn the same. The second is evidence that returns track the version of the model and not the firm's own process and record. The third is evidence that the lag in small firms is so short that the complements are not the binding constraint. Chapter 14 includes the first in its research agenda.

Chapter 4

A theory of AI profit for a service business

The AI profit identity: a model of AI profitability

Monthly AI profit, in this paper, is the agent's four gains (capture, conversion, capacity and retention) less three costs: the run cost of the agent, the fixed cost of the system around it and the expected cost of its errors. A ledger for the whole system adds a fifth gain, demand: the new contacts that the plan produces and the agent then answers. The block below states the identity in the symbols that every later chapter uses, and Table 1 defines each symbol once.

Monthly AI profit = demand (dN x c x V)
                  + capture (N x m x a x c x V)
                  + conversion (N x (1 - m) x dc x V)
                  + capacity (H x w)
                  + retention (B x dr x L)
                  - run cost (k x n) - F - E
SymbolMeaningUnitDeveloped in
dNcontacts a month the system adds through search, AI answers, listings and reviewscontactsChapters 4 and 6
Ninbound contacts a month (calls, forms, chats)contactsChapters 4 and 6
mshare of those contacts that go unanswered or unanswered in time without the agentshare, 0 to 1Chapters 4 and 7
ashare of those otherwise-missed contacts that the agent answers and holdsshare, 0 to 1Chapters 4 and 7
cshare of answered contacts that become booked customersshare, 0 to 1Chapters 4 and 7
dcchange in c on contacts that were already being answered (conversion lift)change in a shareChapter 4
Vvalue of one new customer over the first year, as gross margindollarsChapters 4 and 11
Hhours of staff or professional time released a monthhoursChapters 4 and 12
wvalue of one released hour in its next best usedollars an hourChapter 4
Bcustomers on the bookscustomersChapter 4
drchange in the monthly share of customers retainedchange in a shareChapter 4
Llifetime value of a customer, as gross margindollarsChapter 4
krun cost per contact the agent handlesdollarsChapter 5
ncontacts the agent handles a monthcontactsChapter 5
Ffixed monthly cost of the systemdollars a monthChapters 5 and 8
Eexpected monthly cost of error: the sum over kinds of error of probability times consequencedollars a monthChapter 9

Table 1. The symbols of the profit identity, with the meaning used in every chapter.

The demand line obeys a scope rule: a ledger that puts the plan's fee in F must put dN in the gain, because the fee buys those contacts, and a ledger for the agent alone leaves both out. The rule applies to every purchase. The plan is set against new clients, the agent against answered calls and released hours, and a compliance program against everything that runs on it, so that no single gain is asked to carry it. Every ledger in the earlier edition of this paper that charged a plan without crediting demand was wrong on this point, and this edition corrects it.

The identity is our own accounting model, not a law of economics. Given the definitions it is true by construction and has no empirical content of its own. The content sits in the assumptions a user makes about each term and in the records that would show whether those assumptions hold. Its use is to say where an agent can earn money, where it cannot, and what a firm must record to know which; it is not a forecast. A number placed on a symbol in this paper comes from a cited claim or from shown arithmetic on cited claims. A link that nobody has measured is shown across a range of values, and a quantity a firm must measure for itself is solved for: the count or share at which the gain covers the cost.

Every term is a difference against the same firm without the purchase: without the agent for the four mechanisms, and without the plan for demand. For that reason m is defined "without the agent", and it cannot be read off the contacts the agent now answers. The identity is monthly but V and L are values over a year and over a lifetime. It credits each month's new customers with their first-year margin when they arrive, not when it is earned, which overstates profit in a month of growth and understates it later. Over a steady flow the two roughly offset, and a firm reading cash should spread V across the months.

The terms are written as if they add, but they interact through capacity, as the boundary-conditions section shows. A contact the plan adds belongs to dN and is not counted again in N. A change in retention also changes lifetime value, since L depends on how long customers stay; the identity holds L at its old value and ignores that second-order effect. A firm should also avoid counting a customer's first year twice, once in V and again in L.

Table 2 shows the capture term at work on the dental call chain of our worked cases, in which every link but one is a cited figure. Patient Prism† (call scoring) analyzed 11,552,668 patient calls at 8,280 dental locations in 2025, about 116 calls a month per location (11,552,668 / 8,280 / 12)34. In Patient Prism†'s data, 31 of every 100 callers hung up before reaching anyone, so N x m is about 36 calls a month. Of the 69 who got through, 34 were real chances to book and 35 were not bookable, and 21 of the 34 ended in a booking34. If callers who hang up resemble those who get through, answering every one would give about 11.0 bookings a month (36 x 21 / 69).

The callers include new and existing patients, so a booking is valued at two ends. The low end is one visit, about $282: average gross billings per general dentist of $965,660 in 2025, divided by Statista's 3,429.5 visits per dentist53,54. Statista does not name the source of its count, so the figure is rough54. The high end is a new patient's first year, $887, the average dental spending per person with a visit in 2021 in AHRQ's household survey55. The value that hangs up each month is therefore $3,089 to $9,730 (calc.), and a practice's own phone log says where in that range it sits.

The open link is a, the share of those callers that an agent turns into a conversation. We estimate it lies between the 36.6% of calls that an AI voice assistant at Mount Sinai handled with no human intervention and the 86.4% it completed17 (from the nearest peer-reviewed study; likely high). Those were outbound calls before a procedure, to patients who expected them, unlike a caller who was about to hang up, and Table 2 keeps four values of a instead of leaning on it.

The agent costs $299 a month, its published price, with 250 call minutes included and $0.39 a minute after that18. Suppose each call runs four minutes: the included minutes are about 62 calls (250 / 4), and the busiest row of Table 2, 27 calls a month, is about 108 minutes (27 x 4). Every row stays inside the included minutes, so the fee in Table 2 is $299 throughout.

A dental practice also needs the HIPAA-ready Essentials program around the agent on patient calls, $824 a month in the first year ($699 + $1,500 / 12) and $699 after56. The program is a compliance layer and does not include the agent. The site assistant, texts to patients and the front-desk add-ons (reminders, recall, waitlist backfill, intake) run inside it too, at their own prices, so its fee is shared by everything on it and Table 2 does not charge it to answered calls alone56. The agent and the program together come to $1,123 in the first year ($299 + $824), which is the narrow case of a practice that buys the program for the agent and runs nothing else on it.

Share of hang-ups turned into conversations (a)Calls a monthBookingsGain, one visit eachGain, a new patient eachLess the agent ($299)Less the agent and a program bought for it alone ($1,123)
10%3.61.1$309$973$10 to $674-$814 to -$150
25%9.02.7$772$2,433$473 to $2,134-$351 to $1,310
50%18.05.5$1,544$4,865$1,245 to $4,566$421 to $3,742
75%27.08.2$2,317$7,298$2,018 to $6,999$1,194 to $6,175

Table 2. The capture term on the dental call chain of our worked cases, set against the agent's own fee and, in the last column, against the narrow case of a program bought for the agent alone, with the open link at four chosen values and each booking valued at one visit and at a new patient's first year.

In Table 2 the agent covers its own fee when it turns 3% to 10% of hang-ups into conversations: 0.3 bookings a month if each is a new patient, 1.1 if each is one visit ($299 / $887 and $299 / $282). At 10% the agent is ahead at both ends of the range, by $10 to $674 (Table 2), though the margin at the one-visit end is only $10. In the narrow case the shares are 12% to 36% (1.3 and 4.0 bookings). Henry Schein One† (software) puts the average dental practice at 30 new patients a month13. At the high end, then, the line between a loss and a profit for the agent is about 0.3 of a new patient a month in 30. A difference of that size is hard to see in a month's bookings, as Chapter 7 argues.

The plan is not charged in Table 2, because answered calls are not its product. Under the scope rule it is set against new patients, and 1.1 a month cover its $999 ($999 / $887), which is 3.8% more than the average practice's 30 ($999 / $887 / 30)18.

The split of connected calls into chances to book and calls not bookable is Patient Prism†'s classification34. The chain assumes that callers who hang up resemble callers who get through, which the source does not show, and it nets out none who ring back; counting them would lower every figure, most of all at the low end, where the callers are existing patients. A booking's value is revenue, not gross margin, and Chapter 5 shows the dental count at the share of spending a practice keeps. Conversion, capacity and retention are left out of the table, which is not the same as setting them to zero. Errors are not priced: the margin in each row is what they would have to consume before the agent lost money.

Capture: speed to lead and the contacts a business loses

Capture is the gross margin from contacts that would otherwise have been lost: N x m x a x c x V. In words, it is the inbound contacts a month (N), times the share that go unanswered or unanswered in time without the agent (m), times the share of those missed contacts that the agent answers and holds (a), times the share of answered contacts that become booked customers (c), times the value of one new customer over the first year as gross margin (V). It is a product of five factors, so it is zero when any one is zero. N is the calls, forms and chats that arrive. Its size is largely set before any agent acts, by search, maps, reviews and referrals, and the part that a plan adds is dN; Chapter 6 treats both.

The best-known evidence on the contacts a business loses is about the speed of response. Oldroyd, McElheran and Elkington wrote in Harvard Business Review in 2011 that their research shows most companies are not responding nearly fast enough57 (secondary, through a news report that reproduces the article). The authors audited 2,241 US companies, measuring how long each took to respond to a web-generated test lead58. Of those companies, 37% responded within an hour, 16% within one to 24 hours, 24% took more than 24 hours and 23% never responded at all58. The average response time among companies that responded within 30 days was 42 hours58.

A separate study cited in the same article involved 1.25 million sales leads received by 29 B2C and 13 B2B US companies58. In that study, firms that tried to contact potential customers within an hour of receiving a query were nearly seven times as likely to qualify the lead as those that tried to contact the customer even an hour later58, and more than 60 times as likely as companies that waited 24 hours or longer58.

The article defines qualifying as having a meaningful conversation with a key decision maker58. That outcome is a conversation, not a booked customer. The comparison is between firms that chose different speeds, and firms that answer fast may differ from the rest in other ways. The report also does not break the companies down by sector or size, so the figures cannot be read as figures for a dental, medical, legal or accounting practice.

Access studies of professional practices measure something different. AMN Healthcare's 2025 survey found an average new-patient wait for a physician appointment of 31 days across six medical specialties in 15 large metro areas59. Its research specialists called 1,391 physician offices between January 15 and February 24, 2025 to schedule a new-patient appointment59. The average wait for family medicine was 23.5 days, with a range from one day to 207 days59. The survey covers physician offices, not dentists. A long wait may be a symptom of a full schedule.

The telephone audits in our evidence measure access by insurance status, not how often an office answers. In a simulated-patient telephone study of primary care practices in 10 states, from November 2012 to April 2013, 84.7% of privately insured callers and 57.9% of Medicaid callers received a new-patient appointment60. Uninsured callers got appointments 78.8% of the time when full cash payment was possible, and only 15.4% of the time if payment required at the visit was restricted to $75 or less60.

In a telephone audit of dental offices eligible to bill Medicaid, which by its title concerned children in Illinois, privately insured patients had 5.9 times greater odds of obtaining an appointment than Medicaid patients61. We think these audits speak to c and to the firm's rules about payment and insurance more than to m: a call can end without an appointment, and how often it does differed with the caller's insurance and with the payment required at the visit.

In a secret-shopper phone survey of Arizona Medicaid pediatric dental practices, 86% of 185 cataloged practices were reached on behalf of a Medicaid-enrolled child and 81% on behalf of a commercially insured child62. The cited claim does not define reached or say how many attempts were made, so we do not read it as an answer rate.

The secret-shopper study of law firms by Clio† (practice-management software) comes closest to measuring whether the phone is answered. Clio† reported that in its 2024 test of 500 law firms only 40% answered the phone call, down from 56% in 201963. Counting callbacks, 52% either answered or called back, so Clio† says 48% were essentially unreachable by phone63. Clio† also reported that 33% responded to an emailed client inquiry, down from 40% in 201963. The 2019 figures are as Clio† gives them in its 2024 report.

Table 3 sets these studies against the quantity that the identity needs.

SourceWhat was measuredResultWhy it is not mClaim
Oldroyd, McElheran and Elkington (secondary)Response to a web-generated test lead, 2,241 US companies37% within an hour; 23% neverA test lead, not a firm's own traffic; sectors not given58
AMN Healthcare, 2025New-patient appointment wait, 1,391 physician offices, 15 metros31 days on averageMeasures the wait for an appointment, not whether a call was answered59
Rhodes and co-authorsSimulated-patient calls to primary care practices, 10 states84.7% private against 57.9% Medicaid got an appointmentMeasures access by insurance status60
LeHew and co-authorsTelephone audit of dental offices billing Medicaid5.9 times greater odds for private insuranceMeasures access by insurance status, for children in Illinois61
Reddy and co-authorsSecret-shopper calls to 185 cataloged Arizona Medicaid pediatric dental practices86% and 81% reachedOur claim does not define reached; one state62
Clio†, 2024Test calls to 500 law firms40% answered; 52% answered or called backTest of law firms; sample frame not given63

Table 3. Studies of speed and access set against m. None measures m, the share of a firm's own contacts that go unanswered or unanswered in time.

Each study in Table 3 is a test by outsiders: researchers sent test leads or placed test calls to firms. The quantity m belongs to a firm's own real traffic across its whole week, including evenings, lunch hours and the moments when the one receptionist is on another call. Counts of real traffic exist, and each counts a different event. In Patient Prism†'s 2025 data, 31 of every 100 callers hung up before reaching anyone34. In 2026 data from Invoca† (call tracking) on six health care fields, 54% of inbound calls were answered by a person64. Across Veterans Health Administration call centers, a mean 12.0% of calls ended before a staff member answered in October 201465, and 8.3% by September 201665.

These measures come from different settings and must not be averaged. The VA figure is a government system's call centers, not a private office. None of them is m, and none says how often a given firm loses a contact. Together they say that a practice whose every call is answered in time is not the usual case. "In time" is not defined the same way across the test studies either. The lead-response audit cuts at an hour58, and Clio†'s email result reports replies within eight hours66. So m is the quantity a firm must measure for itself, with a threshold it chooses for each door, and Chapter 7 shows how.

For a, the share of missed contacts that an agent answers and holds, the evidence is thinner. Relatient† (patient scheduling) says its voice product answered 62% of inbound calls, of which 72% were handled or passed to staff, which is not a containment rate39. A peer-reviewed containment rate on inbound calls is NF; the nearest evidence is a voice-based AI system that replaced a telecommunications company's phone menu and temporarily increased customers' demand for human service50. We think that is one reason a may fall short of 1, because some contacts the agent receives will ask for a person.

Conversion (c) depends in part on the firm's rules about insurance and payment, and it may differ for contacts an agent holds. In Invoca†'s 2026 health care data, 43% of calls answered by a person were leads and 45% of those leads converted on the call, by its classification64. Gartner reports that in its survey of 5,728 customers conducted in December 2023, 64% said they would prefer that companies did not use AI in customer service67. V must come from the firm's own books: a new customer's first-year revenue less the direct cost of serving that customer. Our worked cases give cited first-year payments for six kinds of client (Chapter 5), and they are revenue, not margin.

For a service business, capture is zero in four cases: the firm has no inbound contacts, it already answers every contact in time, the agent holds none of the contacts that were missed, or no caller can be booked, as when a practice takes no new clients at all. A full schedule is not among them. It changes what a captured caller is worth to the practice, as the section on capacity shows, and does not make the caller worth nothing. A firm that buys an agent without first counting its contacts and its misses is buying a product whose value it has not yet looked for.

Conversion: what turns a contact into a customer

Conversion is the gross margin from raising the booking rate on contacts that the firm was already answering: N x (1 - m) x dc x V. The factor N x (1 - m) counts the inbound contacts that were answered in time without the agent, and dc is the change in c on those contacts. The term differs from capture because it concerns contacts the firm already had. It is zero when the agent does not touch them, or touches them without changing what the customer experiences.

In Patient Prism†'s data behind Table 2, 69 of every 100 calls connected with the front desk34, about 80 calls a month at 116. A rise of one point in c on those calls is 0.8 more bookings a month, worth $226 at one visit each and $710 at a new patient's first year (116 x 0.69 x 0.01 = 0.8, then 0.8 x $282 and 0.8 x $887). A change that small is hard to see in a month's bookings. The sign of dc is open. A fall of 3.4 points would cancel the whole capture gain at 25% of hang-ups held, whichever value a booking carries (2.7 / 0.8). An agent that is better than silence and worse than the receptionist it replaces can lose more on conversion than it gains on capture.

The service-quality literature in our evidence says what customers use to judge a service, which gives a vocabulary for what dc depends on. The 1985 model of Parasuraman, Zeithaml and Berry rested on an extensive exploratory investigation of quality in four service businesses68. The SERVQUAL scale, in its 1986 working paper, measures five dimensions: tangibles, reliability, responsiveness, assurance and empathy69.

A 2013 literature review (secondary) defines responsiveness as the willingness of the firm to help customers and perform the service promptly70. It defines assurance as the competence and politeness of personnel and their capability to inspire confidence70. It defines reliability as the capability of the firm to perform the promised service in a careful and accurate manner70. The same review describes the customer gap as the discrepancy between expectations and the customer's perception of the service70. In the 1986 study, average perceptions sat about one point below average expectations on a 7-point scale across four services69. These sources measure how customers rate a service they have used. None studies why a contact books, and applying them to conversion is our inference.

The economics of professional services gives some support for the inference. Day and Barksdale (secondary) argue that purchasing professional services differs from buying physical products because services are typically intangible and customized, so that evaluation before purchase is difficult71. Darby and Karni (secondary, through a citation and a later summary) introduced credence qualities, which a consumer cannot evaluate even after purchase and normal use72. When a customer cannot judge the technical quality of a dentist, a lawyer or a tax preparer, the cues available at first contact carry weight: how promptly someone answers, whether the answer shows competence and inspires confidence, and whether what is promised is accurate. Those are the three dimensions above.

Clio†'s 2024 test of law firms separates responsiveness from the other two. Among the firms that replied to emails, 84% did so within eight hours, but just 18% provided clear next steps or cost information66. Clio†'s press release adds that 73% of the shoppers were unlikely to recommend the firms they contacted66. Replying within eight hours speaks to responsiveness. Giving clear next steps and a cost speaks to assurance and reliability. The design consequence we draw for an agent is that speed alone is unlikely to move dc. The agent must also state the next step, quote what the firm has decided to quote and keep what it promises, and Chapter 8 treats those as the content of the hand-off.

Capacity: AI automation and the value of released time

Capacity is the value of staff or professional time that the agent releases: H x w. The first factor, H, is the hours released a month, net of the hours spent supervising the agent, reading its summaries and correcting its mistakes. It can be negative. The second, w, is the value of one released hour in its next best use, and it is the factor that decides. An hour released from a task is worth what it is then put to: it may be resold to a paying client, it may avoid overtime or a hire, or it may sit idle, and an idle hour is worth nothing. An hour that avoids overtime or a hire is worth up to what the employer pays for it, which for office and administrative support occupations in health care was $34.59 an hour worked in June 2026, benefits included73.

The studies of time saved do not close the gap between a saved hour and a dollar. As Chapter 2 showed, a 14% rise in issues resolved per hour frees about 12% of the hours for a fixed workload (1 / 1.14 = 0.877)2. In an April 2025 labor-market working paper, Humlum and Vestergaard report that AI chatbots had no significant impact on earnings or recorded hours in any occupation, with confidence intervals ruling out effects larger than 1%31. They report modest productivity gains, with average time savings of 3%, combined with weak wage pass-through31. A later NBER version of the paper rules out effects larger than 2% two years after the launch of ChatGPT32, so its bound differs, and the figures above are from the April 2025 version only. Saved time that does not reach earnings or hours is what we would expect if the time was not resold. But the abstract names the modest gains and the weak wage pass-through as part of its own explanation for the limited effects, so we cannot separate the two explanations.

Whether time is resold depends on whether anyone is waiting for it: a practice with a queue can resell the hours it releases, and a practice with open slots cannot. In the AMN Healthcare survey described earlier, the average new-patient wait for a physician appointment was 31 days59. We infer a full schedule from a wait of weeks, though a long wait can have other causes and the survey says nothing about any one office.

Other cited figures show that many practices are not full. Nearly two in five private-practice dentists told the ADA Health Policy Institute in the third quarter of 2026 that they were not busy enough and could have treated more patients74. Owner dentists who were accepting new patients reported an average wait of 12 days for a first appointment74. In Clio†'s data, the average law firm utilization rate, the share of an eight-hour day put toward billable work, was 38%75. Booked medical visits are also missed: 18.8% across the ten main clinics of one VA medical center, and at the order of 23% in a review of 105 studies76,77, so even a booked day has gaps.

A full schedule changes the form of the gain from more demand. It does not make demand worthless. A practice with a queue turns more inquiries into four things: gaps filled when someone cancels or misses, better-paying work chosen among more requests, a higher minimum fee, and the case for adding a provider. A practice with open hours gains from inquiries directly, and gains little from released time until it has work to put in it. Where a practice sits between the two decides which of the mechanisms to look at first, and Chapter 11 does this for the four professions. Professionals who bill for their time face a further question, which is whether an hour saved may be billed as if it had been spent. That is for the practice's own counsel or compliance lead to decide, and Chapter 11 gives the cited rules.

Capacity is zero when no hours are released, when the released hours are idle or too fragmented to use, when they cannot be reassigned to anyone able to do the work, or when supervision uses as many hours as the agent saves. Practice leaders in a 2026 MGMA poll, asked which phone tasks take staff the most time, named eligibility and prior authorization in 45% of responses and scheduling in 31%78. H itself has to be measured from time records before and after the agent, not taken from a vendor's estimate of time saved.

Until a practice has those records, the calls an agent could take off a medical front desk can only be estimated from two cited neighbors. Call Box† (phone service) counted 299 to 505 calls a month per health care office location in 202033, and the Mount Sinai assistant handled 36.6% of its calls with no human intervention17. Together they give ~109 to 185 calls a month, or ~5.5 to 9.2 hours at the three-minute call that Chapter 5 supposes (299 to 505 calls x 36.6%; likely high). A front desk's inbound calls include more that need a person than a scripted outbound check does.

Covering the agent's $299 fee on hours alone would take 8.6 hours a month ($299 / $34.59), about 25 minutes a working day (8.6 x 60 / 21 working days). That is inside the talk-time estimate above, below its top of ~9.2 hours at the busier office and above its low end of ~5.5, so on the flat fee hours would cover the agent at the busier office and not at the quieter one.

The flat fee understates the cost at those volumes, because the calls in the estimate carry more than the 250 included minutes. ~5.5 to 9.2 hours of talk is ~330 to 552 minutes (5.5 x 60 and 9.2 x 60), and the fee at that volume is ~$330 to $417 ($299 + 80 x $0.39 and $299 + 302 x $0.39). That talk time is worth ~$190 to $318 at $34.59 an hour (5.5 x $34.59 and 9.2 x $34.59), below the fee at both offices, so hours are one gain among several and do not carry the agent by themselves. The program's fee is not set against hours at all, because it is shared by everything that runs on it.

Retention: customer lifetime value as the multiplier

Retention is the gross margin from keeping customers who would otherwise have left: B x dr x L. In words, it is the customers on the books (B), times the change in the monthly share retained (dr), times the lifetime value of a customer as gross margin (L). The product B x dr is the number of additional customers kept in a month, and each is valued at lifetime value, which gives retention a large multiplier on a small count.

In the medical case of our worked cases, a person makes 1.61 primary care visits a year, worth $300 at the 2016 mean payment of $186 a visit (calc.)79,80. That is an average over everyone, including people who made no visit, so a patient who seeks a practice out likely brings more. Among older patients of 23 Oklahoma physicians around 2000, the median relationship lasted 8 years81, which makes L $2,400 at payments. Keeping 0.12 patients a month who would otherwise have left would cover the agent's $299 a month ($299 / $2,400).

The mean is pulled up by costly visits. At the 2016 median of $107 a visit a patient pays $173 a year, so L is $1,382 and the count is 0.2280; at the 2023 mean of $347 for any doctor visit it is 0.07 (calc.)82. In the same study 14% of the patients changed physician in the first year, and 87% of those who changed did so involuntarily81, so most leaving lies beyond an agent's reach.

Gupta, Lehmann and Stuart define the value of a customer as the expected sum of discounted future earnings83. Valuing five firms from public data, they found that a 1% improvement in retention, margin or acquisition cost improves firm value by 5%, 1% and 0.1%, respectively83. They also found that a 1% improvement in retention has almost five times greater impact on firm value than a 1% change in the discount rate or cost of capital83. Another claim in our evidence records the paper's statement that improving retention by 1% is likely to improve customer and firm value by 3% to 7%83. It sets that against a margin elasticity of about 1 and an acquisition elasticity of 0.02 to 0.383. The 5% comes from that fitted model and depends on its assumptions83.

Reichheld and Sasser's 1990 argument is older and stronger in tone. In the summary of their Harvard Business Review article, as republished by Bain & Company (secondary), the authors state that companies can boost profits by almost 100% by retaining just 5% more of their customers84. That summary states what companies can do and does not measure what any did. A second claim records the article's examples in these words: "Reducing the defection rate just 5% generates 85% more profits in one bank's branch system, 50% more in an insurance brokerage, and 30% more in an auto-service chain"85. The quoted sentence does not say whether the 5% is a relative or an absolute change85.

Neither source concerns a dental, medical, legal or accounting practice, and neither concerns an AI agent. The identity needs only the direction and the multiplier: because L covers a customer's future years, a small change in the share kept is worth a large sum when the relationship is long. We think L is large for a practice with recurring clients, such as a dental practice with regular recall or an accounting firm with annual returns, and small for a practice whose clients usually come once. Each practice has to measure its own L.

An agent can act on dr by answering existing customers quickly, by rebooking and by reminding. The evidence on reminders concerns attendance, a step toward retention and not retention itself. A 2013 Cochrane review found moderate-quality evidence from seven trials with 5,841 participants that mobile text message reminders improved attendance at healthcare appointments compared with no reminders (risk ratio 1.14, 95% confidence interval 1.03 to 1.26)77. Attendance in the pooled trials was 67.8% with no reminders, 78.6% with mobile phone messaging reminders and 80.3% with phone call reminders77. The authors concluded that the evidence remains insufficient to conclusively inform policy decisions77. These are text and call reminders in healthcare settings, not reminders sent by an AI voice agent.

A randomized quality-improvement initiative at one safety-net primary care clinic studied appointments with a predicted no-show risk of 15% or higher86. Adding model-driven live reminder calls to standard automated reminders lowered the no-show rate to 33% against 36% in the standard-care group86. A 2018 systematic review of 105 studies put the average no-show rate in healthcare appointment scheduling at the order of 23%, with wide differences by continent77. Reminders sit between terms: a kept appointment is revenue that would otherwise have been lost, which the firm may count in c if it defines a booked customer as one who attends, or in dr if the kept appointment keeps the customer. The model counts each kept appointment once.

Retention is zero when the agent does not touch existing customers, or when something else sets retention, such as price, a move or the outcome of the work. A queue does not make it zero. A departing customer's slot can be refilled, but the customer's later years are lost, and the replacement has to be found, booked and brought through a first visit, which is the work and the spending a plan exists to do.

Costs: run cost, fixed cost and the expected cost of error

Run cost is k x n, where k is what the agent costs each time it handles one contact and n is the number of contacts it handles in a month. The factor n is not N: an agent that takes only the after-hours calls handles fewer contacts than arrive. As Chapter 5 sets out, a call consumes model tokens, speech, telephony and a platform fee, and that chapter prices each from published prices. Run cost rises with volume and is zero only when the agent handles no contacts.

The fixed cost, F, is paid whether or not anyone calls. It covers the agent's monthly fee and the cost of the framing and plumbing around it, which Chapter 8 describes. Our own AI voice agent is priced at $299 a month with 250 call minutes included and $0.39 a minute after that, which falls under k x n past the included minutes18, and our plans are priced at $999, $1,999 and $2,999 a month per location18. A plan carries no patient information, so a health care practice can buy it with no compliance program, while the agent on patient calls needs one56. A HIPAA-ready program is a compliance layer added to a plan and does not include the agent56. These are our prices, which we sell. They are a data point, not a market benchmark. A firm that builds the system for itself pays in other ways, in the owner's time and in tools, and Chapter 12 compares the options.

Expected error cost, E, is the sum over kinds of error of the probability of the error times its consequence. The nearest peer-reviewed count of a voice agent's errors is for outbound calls before a procedure at one hospital, with no control group17. A hallucination left 0.7% of the first phase's calls incomplete17. A Harvard news summary of a Lancet study reports that supervising physicians found one hallucination across 98 completed encounters with a text-chat intake assistant87. We estimate from those two that a fabricated statement turns up in about one contact in a hundred (likely low). The first counts only hallucinations that stopped a call, and both settings were closely watched. A field study of an agent answering inbound calls in these four professions is NF; the nearest evidence is the two settings above.

The consequence side is the harder half. Chapter 9 collects it from cited rules, orders and cases, and presents a framework, not an estimate. We know of no agent that handles contacts for which errors are impossible, so E is not zero, and this paper does not price it. A ledger that leaves E out should show the margin that errors would have to consume before the agent lost money, as Table 2 does.

Boundary conditions: when an AI agent earns nothing

Each term has a condition under which it is zero. Table 4 states them. None requires the agent to fail at its task: in every case the cause lies in the firm, the demand or the schedule.

TermExpressionCondition under which the term is zero
DemanddN x c x VThe plan adds no contacts (dN = 0), or the practice wants no new clients and would not choose among more of them
CaptureN x m x a x c x VNo inbound contacts (N = 0); every contact already answered in time (m = 0); the agent holds none of the missed contacts (a = 0); or no caller can be booked (c = 0). A full schedule is not a zero condition: it changes the form of the gain, as Table 5 shows
ConversionN x (1 - m) x dc x VThe agent does not touch answered contacts, or changes nothing the customer experiences (dc = 0). It is negative when customers do worse with the agent
CapacityH x wNo hours are released; the released hours are idle, fragmented or cannot be reassigned; or no one is waiting to buy them (w = 0)
RetentionB x dr x LThe agent does not touch existing customers, or something else sets retention. A slot refilled from a queue is not a zero condition: the customer's later years are lost and a replacement has to be found and booked
Run costk x nThe agent handles no contacts (n = 0)
Fixed costFNever zero for a deployed system
Expected cost of errorEZero only if the agent handles no contacts or no error can occur; in practice positive

Table 4. The condition under which each term of the identity is zero.

For a firm with no flow of contacts, the agent handles none, so n is zero, no gain arises, nothing can go wrong, and monthly profit is -F (by the identity). The schedule decides the form a gain takes, not whether there is one. Table 5 sets out three states of the schedule. It prices nothing, because the dollar value of each form depends on the practice.

State of the scheduleCited base rateForm the gain takesTerms that carry it
Open hoursNearly two in five private-practice dentists not busy enough, Q3 202674; average utilization of 38% of an eight-hour day in Clio†'s data, 25% for solo lawyers75,88New clients fill empty hours; released hours have no buyer yetDemand, capture, retention
Booked, with gapsMean no-show rate of 18.8% across ten main clinics of one VA medical center; about 23% in a review of 105 studies76,77Visits kept, and slots refilled when someone cancels or missesKept appointments (in c or dr), capture, demand for the waiting list
Full, with a queue12% of dentists too busy to treat all patients requesting care, and 18% treating all but overworked, Q4 202589; a 31-day average new-patient wait in AMN's survey of physician offices59Better-paying work chosen, a higher minimum fee, a provider added; released hours resoldDemand, as choice and price; capacity

Table 5. The form the gain takes in three states of the schedule, with cited base rates and no supposed dollar value.

In a booked day with gaps, the gain runs through reminders, rebooking and a waiting list. No row makes demand worthless. The agent and the plan are the same in each row, and only the schedule differs, which is the sense in which an agent is not profitable or unprofitable in itself. The identity says where to look, and the answer depends on where the practice sits.

Tables 4 and 5 together locate the one corner where the system loses money: a practice that is not growing and whose intake and scheduling are already sufficiently optimized. There every gain is small and F is still paid. The corner is real, but it is the exception: the cited base rates put a typical practice well away from it. Intake that is already sufficiently optimized is not what the call data show. In Patient Prism†'s data, 31 of every 100 callers hung up before reaching anyone34, and in Clio†'s 2024 test only 40% of 500 law firms answered the call63. Scheduling that is already sufficiently optimized, with a closed door to growth, is not what the no-show, busyness and utilization rates in Table 5 show. Only a practice's own records can show that it sits in the corner, which is the subject of the last section.

The terms can also interact. If captured customers use the hours that capacity released, the two terms are no longer independent: a firm that is nearly full may find that capture fills the hours that capacity freed, so that the two gains partly offset. We do not model that interaction. A user of the identity should check, for each practice, that one term is not being paid for twice with the same hour.

Why each term needs a technical marketing system to be observed

Every term in the identity is a difference against the firm without the agent, and a difference can be seen only through records. To the firm, a term with no record is unknown. Table 6 sets out, term by term, what a firm must record.

TermWhat must be recordedWhere the system records itChapter
dNNew contacts by source after the plan starts, against the months beforeSource-tagged phone numbers and forms, listing and review insights6 and 7
NEvery contact, by door and datePhone log, form inbox and chat log, with each door tagged with its source6 and 8
mWhether and when each contact was answered, before the agentCarrier call log with answer times, form timestamps, a first-response clock, or a holdout period7
aWhich missed contacts the agent answered, held, handed off or lostAgent log and summaries joined to the contact record7 and 8
c and dcWhich contacts booked, by source and by who answeredCalendar entries linked to the contact record7
VRevenue and direct cost of a new customer's first yearCustomer record and the firm's books7 and 11
H and wHours on each task before and after; what the released hours were used forTime records, the schedule, billing records7 and 12
B, dr and LStart date, visits, last visit and margin of each customerCustomer record7
k and nContacts handled and the cost of eachVendor invoices and usage logs5 and 12
FMonthly cost of the systemInvoices5 and 8
EErrors found, their cost and how they were foundIncident log and review of calls9 and 12

Table 6. What each term of the identity requires a firm to record.

An agent's own logs record what it did, not what it earned: the transcript ends at the call, and whether the caller booked, came, paid and returned sits in the calendar and the customer record. Most of these records must be made at the time. A firm cannot reconstruct from memory how many calls went unanswered last spring, so m cannot be measured after the fact unless a log exists. Every term needs a comparison as well as a count. The count of bookings from the agent says nothing about profit until it is set against what the firm would have booked anyway, through a baseline or a holdout, and Chapter 7 gives the designs a small firm can run.

The thesis of this paper maps onto the table. Demand is N and the part of it the plan adds, dN, the factors that nothing inside the agent can change. Process is what happens to each contact, which shapes a, c, dc and the hours in H. Measurement is every row of Table 6. A firm with demand and process and no measurement has an agent that may be earning and cannot show it. A firm with measurement and no demand can show that its agent has nothing to do. For that reason the paper treats the technical marketing system as the condition under which the identity can be read at all.

Chapter 5

The unit economics of an AI agent

AI agent cost per contact: what one call consumes

Under this chapter's assumptions, a three-minute call handled by an AI voice agent costs between $0.18 and $0.42 in the three configurations priced here (calc., Table 4), and up to $0.96 at the top of one vendor's printed range (calc., below). The same three minutes at the median receptionist wage cost $0.91 (calc., Table 5). Both sums rest on prices anyone can look up and on assumptions anyone can change, and neither describes a particular firm.

Run cost per contact (k) is what the agent costs each time it handles one contact. Contacts handled a month (n) is how many it handles. The fixed monthly cost of the system (F) is paid whether or not anyone calls. The expected monthly cost of error (E) is the sum, over kinds of error, of probability times consequence, and Chapter 9 treats it. This chapter does not price it, so each break-even below covers the priced costs only. Whatever a practice earns above break-even is the margin that errors would have to consume.

On the benefit side, V is the value of one new customer over the first year. In the identity it is gross margin; the worked cases below use payments and fees, which run higher than margin. Chapter 4 adds a demand line, dN x c x V, for the contacts a plan brings, and its scope rule decides which gain each purchase is set against. Captured customers come from inbound contacts a month (N); the share of them unanswered, or not answered in time, without the agent (m); the share of those the agent answers and holds (a); and the share of answered contacts that become booked customers (c).

A voice call consumes four things. Model tokens are the text the language model reads and writes. Speech is the conversion of the caller's voice to text and of the agent's reply back to voice. Telephony is the phone number and the minutes carried over it. The platform is the software that joins the other three and runs the call, and it may charge by the minute, by the month or both. Some vendors quote two or three of these as one per-minute rate, so the chapter follows whichever grouping each vendor publishes.

A price is given as published, with its date and unit. What a client is worth comes from the cited figures of our worked cases and is never supposed. The rest, such as how long a call lasts and how many tokens it uses, is an assumption, introduced by "suppose", listed in a table and varied afterward. Vendors mostly measure the contacts a practice receives, the share that go unanswered and the share that book. In the 2025 data of Patient Prism† (dental call analysis), 31 of every 100 callers to dental locations hung up before reaching anyone34. Chapter 4 sets out the other measures with their owners. The share of missed callers an agent holds appears only as a range, beside an estimate from the nearest study. Vendors are named as sources of published prices and of their own data. The paper compares no vendor's quality and recommends none.

AI voice agent cost: the inputs, published and supposed

Table 1 gives the model prices and Table 2 the voice, speech, telephony and platform prices. Each was read on 2026-11-10 from the vendor's own pricing page and is given as published. Prices change, and two model prices carry dates on which they step up, so a figure built from them needs its date.

ModelInput priceOutput priceAs publishedClaim
Claude Haiku 5.5$0.10$0.50per million tokens, prompts up to 100,000 tokens90
Claude Sonnet 5.5$2$10per million tokens; cache reads $0.1090
Claude Opus 5.5$4$20per million tokens90
gpt-6-luna$0.10$0.50per 1M tokens, Standard tier91
gpt-6.1-sol$2.00$10.00per 1M tokens, Standard tier91
Gemini 3.1 Flash-Lite$0.25$1.50per 1M tokens, text input, paid tier92
Gemini 3.8 Flash$0.75 through December 31, 2026; $1.50 starting January 1, 2027$3.75 through December 31, 2026; $7.50 starting January 1, 2027per 1M tokens, paid tier92
Gemini 4 Carbon$2.00 through November 30, 2026; $4.00 starting December 1, 2026$10.00 through November 30, 2026; $20.00 starting December 1, 2026per 1M tokens, paid tier92

Table 1. Model prices as published on 2026-11-10, for input and for output. Dated steps are shown as printed.

ItemPrice as publishedUnit and what it coversClaim
Retell AI, voice agents on pay as you go$0.07 to $0.31per minute93
Retell AI, example agent$0.11: model $0.04, voice infrastructure $0.055, text-to-speech $0.015, telephony $0.00per minute; the calculator's default state, not a bill93
Retell AI, chat agentsfrom $0.002per message93
Vapi, calculator at 1,000 minutes$82 to $129 in total: hosting $50, transcription $10, OpenAI model $8 to $45, ElevenLabs voice $15 to $24, telephony $0per month at 1,000 minutes; the calculator's default state, not a bill94
ElevenAgents$0.08, or $0.16 above the concurrency limit; model usage billed separately; text messages $0.003 eachper additional call minute95
OpenAI gpt-live-1$0.05per minute, billed per second; backend model and tool usage billed separately91
Deepgram Nova-3, monolingual speech-to-text$0.0077 on pay as you go; $0.0065 on the Growth planper minute96
Twilio, local number, receiving calls$0.0085 per minute, plus $1.15 a month for the numberprices dated October 202697

Table 2. Voice, speech, telephony and platform prices as published on 2026-11-10 (Twilio's prices are dated October 2026).

The comparison with a person uses two wages. The Bureau of Labor Statistics reports a median hourly wage of $18.27 for receptionists98 and $21.53 for customer service representatives99, both for May 2025. Each is a median wage, not the full cost of an employee or what any one county pays. Our own price for the voice agent is $299 a month with 250 call minutes included, about 62 four-minute calls (250 / 4), and $0.39 a minute after that18. The worked cases in Tables 6 to 8 stay inside the included minutes, so the agent's fee in them is $29918.

No.SupposeValueUsed for
1A handled contact is one inbound call of3 minutes of connected time (1.5 and 6 minutes tested later)every per-minute price
2Over the call the model is run12 times, each reading about 2,500 input tokens (instructions plus the conversation so far) and writing about 125 output tokens, so 30,000 input and 1,500 output tokensthe model line
3Calls arrive on a local number and telephony is billed separatelyat the receive rate, for each connected minutethe telephony line
4Vendors bill connected timewith no rounding upevery per-minute price
5The agent handles n contacts a month100, 300 or 1,000monthly run cost
6The calls in the worked cases stay inside the agent's included minutes, so its fee is$299 a month, the published price with 250 minutes includedTables 6 to 8
7The expected monthly cost of error, Enot priced; the margin above break-even is what it would have to consumeChapter 9
8A paid receptionist hour is spent on calls for100%, 50% or 25% of the timethe human comparison

Table 3. Assumptions used in this chapter. Each is a supposition or a limit.

A model that runs once per exchange re-reads its instructions and the conversation so far every time, so the input side is far larger than the output side. The 2,500-token figure in Table 3 supposes a short set of instructions and a short conversation; a longer instruction set or a longer call raises it, and the sensitivity section tests that. The example agent on Retell AI's calculator prices its model line at $0.04 a minute93, which is $0.12 for three minutes. Vapi's calculator prints $8 to $45 for an OpenAI model at 1,000 minutes94, which is $0.008 to $0.045 a minute, or $0.024 to $0.135 for three minutes. Our supposed tokens, priced at the Claude Sonnet 5.5 rates, cost $0.075 (30,000 x $2 / 1,000,000 = $0.06, plus 1,500 x $10 / 1,000,000 = $0.015), which lies inside that span. It is a consistency check and does not validate the assumption.

AI receptionist cost per call in three configurations

Three configurations show how far the answer moves with the choice of vendor and model. Configuration A takes a vendor's bundled per-minute price. Configuration B assembles a lean stack from a voice model and a low-priced language model, and Configuration C a premium stack with a top-priced model. All three use the supposed three-minute call, 30,000 input tokens and 1,500 output tokens from Table 3, and the same telephony line.

Configuration A uses the example agent on Retell AI's calculator: $0.04 a minute for the model, $0.055 for voice infrastructure and $0.015 for text-to-speech, which is $0.11 a minute93. The page shows telephony at $0.00 for that selection, so suppose instead that calls arrive on a local number billed at Twilio's receive rate of $0.0085 a minute97. Three minutes at $0.11 is $0.33 for the model, speech and platform, and three minutes at $0.0085 is $0.0255 for telephony, a total of $0.3555 ($0.33 + $0.0255).

In Configuration B the voice layer is OpenAI's voice model at $0.05 a minute, billed per second, with backend model and tool usage charged separately91. Suppose the backend is the low-priced model gpt-6-luna at $0.10 per million input tokens and $0.50 per million output tokens91. Three minutes at $0.05 comes to $0.15. The tokens cost $0.00375 (30,000 x $0.10 / 1,000,000 = $0.003, plus 1,500 x $0.50 / 1,000,000 = $0.00075). Adding $0.0255 for telephony gives $0.17925 ($0.15 + $0.00375 + $0.0255).

Configuration C puts ElevenLabs' ElevenAgents at $0.08 per additional call minute in the voice layer, with the model billed separately95. Suppose telephony is also billed separately, at the same Twilio receive rate, and the model is Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens90. Three minutes at $0.08 comes to $0.24. The tokens cost $0.15 (30,000 x $4 / 1,000,000 = $0.12, plus 1,500 x $20 / 1,000,000 = $0.03). Adding $0.0255 for telephony gives $0.4155 ($0.24 + $0.15 + $0.0255). The page prices additional call minutes and prints no plan fee or included minutes, so C's fixed cost may be understated.

ComponentA: bundled platformB: lean assembledC: premium assembled
Model tokens3 x $0.04 = $0.12$0.00375$0.15
Speech and platform3 x ($0.055 + $0.015) = $0.213 x $0.05 = $0.153 x $0.08 = $0.24
Telephony3 x $0.0085 = $0.0255$0.0255$0.0255
Run cost per call, k$0.3555$0.17925$0.4155
Model share of k34%2%36%
Speech and platform share of k59%84%58%

Table 4. Run cost of one three-minute call in three configurations (calc. from the prices in Tables 1 and 2 and the assumptions in Table 3).

The answer moves by a factor of 2.3 across the three configurations, from $0.17925 in B to $0.4155 in C (0.4155 / 0.17925 = 2.32). The model line alone moves by a factor of 40 between those two, from $0.00375 to $0.15 ($0.15 / $0.00375), so the model does not drive the spread. Speech and platform are the larger part of k in every configuration, and the model is the larger part in none. Hold C's speech and telephony fixed and swap Opus 5.5 for Haiku 5.590: the model line falls 40-fold, yet k falls only from $0.4155 to $0.26925, a factor of 1.5 (0.4155 / 0.26925). On these assumptions, minutes drive cost more than tokens do.

Retell AI prints $0.07 to $0.31 a minute for pay-as-you-go voice agents93, which is $0.21 to $0.93 for three minutes. With telephony added it is $0.2355 to $0.9555 ($0.21 + $0.0255 to $0.93 + $0.0255). Vapi's calculator estimates $82 to $129 for 1,000 minutes94, which is $0.246 to $0.387 for three minutes. The page's own parts add to $83 at the low end against the $82 it prints, a difference of $1 that we do not explain. The same calculator prints $10 for transcription at 1,000 minutes, whereas Deepgram's page prints $0.0077 a minute96, which is $7.70 for 1,000 minutes. Pages that do not reconcile to the cent are a reason to read prices as ranges. Across everything priced here, k runs from $0.17925 to $0.9555 for a three-minute call, a factor of 5.3 ($0.9555 / $0.17925).

Text contacts are cheaper per contact than calls, because they carry no speech minutes and no telephony. Retell AI lists chat agents from $0.002 a message93 and ElevenLabs lists text messages at $0.003 each, with model usage billed on top95. Suppose a chat contact is 10 messages. The message charge is then $0.02 to $0.03 (10 x $0.002 to 10 x $0.003), before the model line. At the Claude Sonnet 5.5 rates the model line for our supposed tokens comes to $0.075, so it would be the larger cost.

AI customer service against a receptionist's wage

On wage alone, a person answering the same call costs more per minute than any of the three configurations, but the comparison is weaker than it looks. At a median hourly wage of $18.27 for receptionists98, a minute of paid time costs $18.27 divided by 60, which is $0.3045. For customer service representatives at $21.5399, it is $21.53 divided by 60, or $0.3588 a minute. Three minutes are $0.9135 and $1.0765 (3 x $18.27 / 60 and 3 x $21.53 / 60). Against Table 4, the receptionist's three minutes cost 2.6 times Configuration A, 5.1 times B and 2.2 times C ($0.9135 / each total). The ordering is not universal: the top of Retell AI's printed range93, with telephony, is $0.3185 a minute (0.31 + 0.0085), which is above the receptionist's $0.3045.

A wage leaves out what an employer pays on top of it, such as payroll taxes, benefits and paid leave, plus equipment, space, recruiting, training and supervision, and the paid time not spent on calls. The first of these is published. Across private industry in June 2026, benefits were 30.0% of employer compensation costs, $14.07 of $46.89 an hour worked100. For front-office staff our worked cases use the row for office and administrative support occupations in health care73. Its employer cost was $34.59 an hour worked, of which wages were $23.50 and benefits $11.0873. That row covers the whole sector, hospitals included, and a broad occupation group.

Three minutes at the employer's cost of $34.59 are $1.73, which is 1.9 times the receptionist's bare-wage $0.9135 (3 x $34.59 / 60; $1.73 / $0.9135). The ordering against Configurations A to C does not depend on the loading, because added cost on the human side only widens the gap.

Occupancy matters more than the loading. A receptionist is paid while the phone is quiet. Suppose the share of paid time spent on calls is 100%, 50% or 25%, as in Table 3. Each call minute then carries one, two or four paid minutes, and the three-minute call costs the amounts in Table 5.

Occupation and hourly cost100% on calls50% on calls25% on calls
Receptionist, median wage ($18.27, May 2025)$0.91$1.83$3.65
Customer service representative, median wage ($21.53, May 2025)$1.08$2.15$4.31
Office and administrative support in health care, employer's cost with benefits ($34.59, June 2026)$1.73$3.46$6.92

Table 5. Cost of one three-minute call at three supposed levels of occupancy (calc. from the Bureau of Labor Statistics figures)73,98,99. The first two rows are wages only; the third adds benefits and still leaves out equipment, space and supervision.

The comparison does not show that an agent is cheaper than a person, because the two do not do the same work. A person handles unusual calls, uses judgment and does other tasks between calls, while an agent hands off what it cannot do, so some contacts cost both. How many cost both is not yet settled for a front desk: a peer-reviewed containment rate on inbound calls is NF. The nearest peer-reviewed study covers outbound calls before a procedure at one hospital, with no control group17. In its first phase 36.6% of calls were fully automated and needed no human intervention, and about a third needed a nurse's callback17. We estimate that share as an upper guide for a front desk (from the Mount Sinai figure; likely high, because front-desk callers ask for more than a scripted check).

The vendor reports are not containment either. Relatient† (scheduling) says its voice product answered 62% of inbound calls, of which 72% were handled or passed to staff39. Notable† (AI voice assistants) tells the story of one hospital system in which its assistant handled 40% to 60% of inbound insurance inquiries, with a saving of $4 a call projected at full scale rather than measured101. Field experiments show outcomes other than cost. At a large telecommunications company, introducing a voice-based AI system to replace the phone menu temporarily increased machine-service duration and customers' demand for human service, and persistently reduced customer complaints50. On Alibaba's Taobao platform, a randomized field experiment in customer chats found that an agentic AI system reduced average chat duration and had limited effects on retrial rates, but substantially lowered ratings for AI-eligible chats51. That experiment is a 2026 arXiv preprint51. Neither experiment reports a containment rate or a cost per call.

The right comparison also depends on the mechanism. For capture, the contacts at issue are ones no one was going to answer, so the comparison is between the agent and nothing, and the receptionist's wage is beside the point. For capacity, in which the agent takes work that staff would otherwise do, the comparison with the wage is the relevant one, and the occupancy of the staff decides it. Chapter 12 sets out the hire, buy or build comparison with these wages as inputs.

In a Gartner survey of 5,728 customers conducted in December 2023, 64% said they would prefer that companies did not use AI in customer service67. In the same survey, 53% said they would consider switching to a competitor if they found out a company was going to67. The release is the analyst firm's own account of its survey and does not give the sampling frame in the text we read. The survey does not show whether that preference carries over to a firm's own callers; if it does, c could be lower for contacts an agent handles than for contacts a person handles.

AI ROI: break-even in captured customers

Break-even, in this chapter, is the number of new clients at which their first-year value equals the cost of what is bought. For the agent alone they are the capture term's count, N times m times a times c. For a purchase that includes the plan they come from demand and capture together, dN times c plus the capture count, as the scope rule of Chapter 4 requires. With errors unpriced, a purchase breaks even when that count times V equals its cost, so the number needed a year is 12 times the monthly cost divided by V. Conversion lift, released hours, kept appointments and retention would lower the number needed, so the figures below are a test that new clients alone must pass, and a demanding one.

Table 6 takes the cost from our published prices18,56. The least any practice can add is the agent on a plan it already has, at $299 a month in every profession. A dental or medical practice also needs the Essentials program around the agent on patient calls56, $824 a month in the first year ($699 + $1,500 / 12) and $699 after. The program is shared by every tool that handles patient information, so Table 6 sets it against no single gain. The narrow case is a practice that buys it for the agent alone and runs nothing else on it, which costs $1,123 a month in the first year ($299 + $824), and Table 6 shows that case beside the agent's own.

The values of V are the first-year values of our worked cases. They are drawn from AHRQ and CDC for a primary care patient79,80, AHRQ for a dental patient55, Washington's Office of Public Defense for a DUI matter102, and the Insurance Research Council and the State Bar of California for an injury case103,104. For an advisory client at a firm that sells advisory services they come from the CPA.com and AICPA benchmark of 206 self-selected practices105, and for a bookkeeping client from the benchmark of Ignition† (accounting-firm software)106. They are payments and fees, not gross margin.

ClientFirst-year value, from our worked casesThe least a firm can add: agent on a plan ($299)The narrow case in health care: agent and Essentials on a plan ($1,123)With the lowest plan ($1,298, or $2,122 in health care)
Primary care patient$300: 1.61 visits a year at $18612.044.984.9
Dental patient$887: spending per person with a visit, 20214.015.228.7
DUI defense matter$3,500: average attorney fee, Washington, 20201.0none needed4.5
Auto injury case$6,667: one-third of the average claim payment0.5none needed2.3
Advisory client at a CPA firm that sells advisory services$17,867: median revenue per client, 2024, 206 self-selected practices0.2none needed0.9
Bookkeeping client, most common vendor band$3,000: low end of $250 to $499 a month1.2none needed5.2

Table 6. New clients a year needed to cover each purchase on first-year value alone, from our worked cases (calc. from cited values and our prices; not a forecast). The Essentials column is the narrow case of a program bought for the agent alone; where other tools share the program, its fee is not asked of new clients.

Clients arrive whole, so a fraction is a rate over time. At a CPA firm that sells advisory services, 0.2 advisory clients a year cover the agent, which is one every five years. The benchmark's 10% median turnover implies that such a client stays about ten years (1 / 10%)105. The benchmark's 206 practices chose to answer a survey about a service they sell, so the figure is the high end for a small firm that does mostly tax returns. Table 6's bookkeeping row brackets it from below. The count depends steeply on what a client is worth: across the rows it runs from 0.2 to 12.0 for the agent, which is the case for marketing to the clients a firm wants.

The health care rows are the most demanding test because a patient is worth less there than a client in law or accounting (Table 6), and not because the agent costs more: it costs $299 in every profession18. They are also the least complete, because the program around the agent is shared by everything that runs on it and is set against no single gain, and Table 7 shows that side.

Solo primary care is the tightest case in our worked cases, and it is not typical of medicine. One physician has 221 scheduled appointments a month, of which 171 to 180 are kept once 18.8% to 23% are missed76,77,107. New patients make 12.5% of primary care visits, so the practice sees 21 to 22 new-patient visits a month (12.5% of 171 to 180)108. The 84.9 new patients a year that cover the full system, 7.1 a month, would be 33.2% more, and the system costs 6.3% to 6.7% of one physician's payments. Counting three years of payments cuts the count to 2.4 a month, and counting the median 8-year relationship of older patients cuts it to 0.981. A four-physician practice needs 8.3% more on first-year payments (calc.).

The dental case shows what margin does to the count. On the full system at $2,122, 2.4 new patients a month cover it if the practice keeps all of a patient's spending (28.7 a year / 12, Table 6), and 4.8 if it keeps half (2 x 2.4). At 22.3%, the average general dentist's net income as a share of gross billings in 2025, it takes 10.753. The 22.3% is a floor on the margin of one more patient, since rent and salaries are already paid when that patient is seen, so 10.7 is the most it would take. Nearly two in five dentists told the ADA Health Policy Institute in 2026 that they could have treated more patients74.

Our plans are $999, $1,999 and $2,999 a month per location18. On a 12-month term paid in full two months are free, which makes them $9,990, $19,990 and $29,990 a year. The $2,999 plan carries a $3,000 setup fee and the two lower plans a one-time onboarding fee of $499 and $999, all three waived on a 12-month term18. The HIPAA-ready programs are $1,500 setup plus $699 a month (Essentials) and $3,500 setup plus $1,199 a month (Integrated), per location, after a compliance consultation56. A program is a compliance layer added to a plan, and it does not include the voice agent56. A plan carries no patient information, so a dental or medical practice can buy the marketing with no program at all56. A program is needed for tools that can tell a named person is a patient, such as the agent on patient calls, appointment forms and texts to patients56. Four front-desk add-ons run inside the program: reminders and rebooking, recall and reactivation, waitlist backfill, and intake with insurance capture56.

Under the scope rule of Chapter 4, Table 7 sets each purchase against its own product: the plan against new clients, the agent against answered calls and hours, and the program against everything that runs on it, so that no single gain is asked to carry it. Where Table 7 does set it against one gain, kept visits, it adds the reminders add-on, because reminders are what the two no-show studies measured. Charging a plan's whole fee against the agent would count the fee and none of the demand it buys, and charging the program's whole fee against the agent would count the fee and none of the other tools on it.

The costs in Table 7 are those of our worked cases, with setup spread over 12 months18,56. What covers each comes from the same cases (calc. from the cited figures named in Table 6 and five more). Henry Schein One† (practice software) counts 30 new patients a month at the average dental practice13, and the call funnel of Patient Prism† covers dental locations34. A dental visit is valued at about $282 from ADA billings and Statista's visit count, which names no source and so makes the figure rough53,54. The no-show rates of 18.8% to 23% are cited in Chapter 476,77, and front-office staff cost the Bureau's $34.59 an hour73. Collected lawyer hours use the Michigan median rate of $300 after the realization and collection rates of Clio† (legal practice software)75,109.

Way of buying FF a monthSet againstWhat covers it, from our worked cases
A plan alone: the marketing, with no patient information, so no program is needed$999New clients1.1 new dental patients a month, 3.8% more than an average practice's 30; 3.3 new primary care patients a month on first-year payments, 15.6% more for one physician and 3.9% more for four
Agent added to a plan the firm already has$299Answered calls and hoursLaw and accounting: 1.2 collected lawyer hours a month; 0.2 new advisory clients a year at a CPA firm that sells advisory services. Health care: the fee is what 8.6 front-desk hours a month cost ($299 / $34.59); in dentistry, 3% to 10% of callers who hang up turned into conversations; in medicine, 2% to 33% of missed callers across the four rows of the medical call chain
Lowest plan plus agent (law, accounting)$1,298New clients, answered calls and hours5.3 collected lawyer hours a month, or 4.5 DUI matters or 2.3 injury cases a year; 0.9 advisory clients a year at such a firm
Essentials and the reminders add-on, first year (dental, medical)$973 ($824 + $149)Kept visits, because reminders are what the two no-show studies measured; the program is shared by everything that runs on it565.2 kept visits a month, 2.4 points off the no-show rate for one physician or 0.6 for four; for reminders alone, where a program is already in place, 0.8 kept visits ($149 / $186)
Agent plus Essentials added to a plan, first year: the narrow case of a practice whose only patient-facing tool is the agent (dental, medical)$1,123The agent's gains only: answered calls and front-desk hours32.5 front-desk hours a month ($1,123 / $34.59); in dentistry, 12% to 36% of callers who hang up turned into conversations
Lowest plan plus agent plus Essentials, first year (dental, medical)$2,122All three2.4 new dental patients a month at full spending; for one physician, 22.4% to 27.4% of missed visits kept
The same after the first year$1,997All threeThe setup fee is paid

Table 7. Each purchase set against its own product under the scope rule, with costs and coverage from our worked cases (calc. from cited inputs and our prices; not a forecast). The program's fee is shared by every tool that runs on it, so no single gain is asked to carry it; the narrow case is the only one in which the agent's gains alone are set against it.

Read by scope, each purchase is covered by a small count beside its base rate. The agent at a law firm needs about 1.2 collected hours a month, in a profession whose average firm in Clio†'s data puts 38% of the day toward billable work75. The agent's own $299 in health care is the cost of 8.6 front-desk hours a month (calc., Table 7), and the answered calls in the call chain are what cover it: on talk time alone the fee is covered at about 18 hours of calls a month, because minutes past the 250 included are billed (1,080 minutes cost $622.70, $299 + 830 x $0.39, and release $622.62, 18 x $34.59). The program and the reminders add-on are the larger fee, and a four-physician practice covers them with 0.6 points off its no-show rate. A solo physician needs 2.4 points, toward the top of the 0.5 to 3 points found in two reminder studies76,86. One was a before-and-after comparison of phone reminders at one VA medical center and the other a randomized initiative with live calls at one clinic76,86. A solo office therefore covers the program by running more than one tool on it: the agent's answered calls, reminders, recall, waitlist backfill and intake share the same layer.

The medical rows use the 2016 mean primary care payment of $186 a visit80, which a few costly visits pull above the typical one. At the 2016 median of $107, the plan needs 5.8 new patients a month ($999 / (1.61 x $107)) and the program with reminders 9.1 kept visits ($973 / $107); at the 2023 mean of $347 for any doctor visit, they need 1.8 and 2.8 (the same sums)80,82. A kept visit in a slot that is already staffed adds nearly all of its payment, so the kept-visit counts sit closer to a practice's margin than the new-patient counts do.

A self-built agent priced from the vendors' own pages costs far less, with one term left out. The pages cited in Table 2 print no monthly platform fee for the three configurations. Vapi's calculator prints a $50 hosting fee at 1,000 minutes, which we do not treat as fixed because the page gives no rate per minute. The one fixed monthly charge in Table 2 that we use is $1.15 for the telephone number97. At 300 three-minute calls on Configuration C, where k times n is $124.65, a self-built agent costs $125.80 a month (1.15 + 124.65), which is 0.51 collected lawyer hours. Our agent at the same 300 three-minute calls takes 900 minutes, past its 250 included, so its published fee is $552.50 ($299 + 650 x $0.39) and the gap equals $426.70 a month ($552.50 - $125.80).

What a firm pays us for, in our description, is the work around the software: the rules, the calendar link, test calls with the hardest scenarios and a monthly look at the call summaries. A firm that builds its own stack pays for that work in its own time, which Table 2 does not price.

Cost per booked customer, not cost per contact

Cost per contact is the wrong headline because most handled contacts add nothing to the capture term. What decides whether the agent pays is the number of clients it adds, and the cost per booked client is the monthly cost divided by that number. The dental call chain of Chapter 4's Table 2 shows how far apart the two figures sit. The agent costs $299 a month with 250 minutes included. The totals of Patient Prism† come to about 116 patient calls a month per dental location (calc.), which is not a measured average per office34. If the agent sat on every one of those calls, at the three minutes of Table 3 that is 348 minutes (116 x 3), so its fee would be $337.22 ($299 + 98 x $0.39) and its cost per call about $2.91 ($337.22 / 116), roughly 8 times the $0.3555 run cost of Configuration A ($2.91 / $0.3555).

Per booking, the cost depends on the open link, the share of hang-ups the agent turns into conversations. Table 8 shows it at the four values of Chapter 4, which are not findings, beside the two values a booking can carry: one visit at about $282, from ADA billings and Statista's visit count53,54, and a new patient's first year at $88755. As in Table 6, it sets the agent's own fee against bookings, and it sets the program against none of them, because the program is shared by everything that runs on it.

Share of hang-ups turned into conversationsBookings a monthCost per booking, the agent ($299)That cost as a share of one visit ($282)That cost as a share of a new patient's first year ($887)
10%1.1$27397%31%
25%2.7$10939%12%
50%5.5$5519%6%
75%8.2$3613%4%

Table 8. Cost per booking on the dental call chain of Chapter 4, at four values of the open link (calc. on the chain before rounding). Each cost is the agent's $299 divided by that row's bookings before rounding: the 10.97 a month that answering every hang-up would book (116.27 calls x 31% x 21 / 69), times the share. At 10%, that is $272.56 ($299 / 1.097), shown as $273. In the narrow case of a program bought for the agent alone, every cost per booking is 3.8 times as high (the ratio of the fees in Table 6, $1,123 / $299).

The run cost per call can be looked up in Table 4 and the vendor ranges. The cost per booking cannot. It moves 7.5-fold across the four values of the open link (75% / 10%), and it falls below the value of a booking at about 3% of hang-ups turned into conversations if each booking is a new patient, or about 10% if each is one visit (calc., Table 8). It depends on N, m, a and c. N, m and c are Patient Prism†'s measures, and a is the open link, for which Chapter 4 gives an estimate and its likely error. A firm learns its own values only by keeping a record that links each contact to its outcome, which is the measurement problem of Chapter 7.

Sensitivity: long calls, usage spikes and model price changes

Table 9 changes one input at a time on Configuration C at 300 contacts a month. It shows the effect on what the calls cost to run, on our published fee at that volume, and on the collected lawyer hours a month that would cover the fee, at $245.52 for an hour billed at the Michigan median of $300 after Clio†'s realization and collection rates (calc.), as in our law case75,109. Our fee is $299 for the first 250 minutes and $0.39 a minute after18, so it follows minutes and not a vendor's price. The base case is 900 minutes (300 x 3), a fee of $552.50 ($299 + 650 x $0.39) and 2.25 collected hours ($552.50 / $245.52).

Longer calls raise k faster than linearly, because each later exchange re-reads a longer history. Suppose a six-minute call has 24 model runs of about 3,500 input tokens and 125 output tokens, which is 84,000 and 3,000 tokens. At the Opus 5.5 rates the model line is $0.396 (84,000 x $4 / 1,000,000 = $0.336, plus 3,000 x $20 / 1,000,000 = $0.06). Speech comes to $0.48 (6 x $0.08) and telephony to $0.051 (6 x $0.0085). The call therefore costs $0.927 ($0.396 + $0.48 + $0.051), which is 2.2 times the three-minute figure for twice the length ($0.927 / $0.4155). Prompt caching is left out. Cache reads on Claude Sonnet 5.5 cost $0.10 per million tokens against $2 for input90. Whether caching helps depends on what a cache write costs and on how often the same instructions are re-read within the cache's life.

Price changes arrive on printed dates. Gemini 4 Carbon doubles from $2.00 and $10.00 to $4.00 and $20.00 per million tokens starting December 1, 202692, and Gemini 3.8 Flash doubles starting January 1, 202792. On our supposed tokens the first raises the model line from $0.075 to $0.15 a call, or $22.50 more a month at 300 calls (300 x ($0.15 - $0.075)). The second raises it from $0.028125 to $0.05625, or $8.44 more a month. Swapping the Opus 5.5 model90 for Haiku 5.590 cuts the model line from $0.15 to $0.00375 a call. Discounts for batch processing exist90,92, but a live call cannot wait for a batch. We think they could apply to work done after the call, such as summaries and the monthly reading of transcripts; we have checked no specific call flow.

Change to Configuration C at 300 contacts a monthRun cost a monthChange in run costOur published feeCollected lawyer hours a month to cover the fee, at $245.52
Base case$124.65none$552.50 ($299 + 650 x $0.39)2.25
Calls twice as long (six minutes)$278.10+$153.45$903.50 ($299 + 1,550 x $0.39)3.68
Contacts triple (900 a month)$373.95+$249.30$1,254.50 ($299 + 2,450 x $0.39)5.11
Model line rises by $0.075 a call, the step Gemini 4 Carbon takes starting December 1, 2026 on our supposed tokens$147.15+$22.50$552.50, no change2.25
Every minute billed at the burst rate of $0.16$196.65+$72.00$552.50, no change2.25
Model swapped from Opus 5.5 to Haiku 5.5$80.78-$43.88$552.50, no change2.25

Table 9. One-at-a-time changes to what 300 contacts a month cost to run and to our published fee at that volume (calc. from90,92,95 and Table 3), with collected hours from our law case. The fee is $299 + (minutes - 250) x $0.39 once minutes pass the 250 included, so only the rows that change minutes move it. The sums rest on assumptions and are not a forecast.

Spikes and abuse are the cases in which the per-minute meter has no ceiling. ElevenLabs prints a burst price of $0.16 a minute, double the standard rate, when a firm exceeds its concurrency limit95. The burst row of Table 9 prices every minute at that rate, which is a worst case for a busy month. Another case is a call left open. Suppose one call stays connected for 60 minutes through a misdial, a prank or a loop. In Configuration A that call costs $7.11 to run (60 x ($0.11 + $0.0085)), and 100 such calls cost $711, which is 2.4 times our $299 fee ($711 / $299). At our published $0.39 a minute, the same 100 calls are 6,000 minutes and add $2,340 to a month's bill (6,000 x $0.39). OWASP names the general risk Unbounded Consumption: an application that allows excessive and uncontrolled inferences, leading to risks that include economic losses110. We think the control is a cap on call length and on daily spend, set below the point at which a month becomes a loss. It belongs in the agent's configuration from the first day.

No change in Table 9 moves the collected hours needed above 5.11 a month, which is where tripled contacts put it, and the supposed case of 100 runaway calls adds $2,340 for a month, which is 9.5 collected hours ($2,340 / $245.52). On these assumptions the cost side is bounded and, for a firm that sets caps, small. We think a firm that cannot afford it has clients worth little, high volume and an agent that captures nothing, and it has a demand and measurement problem before it has a price problem.

What the cost arithmetic does not settle

Beside the value of a client, the cost per contact is small. At the k of Configuration C, the dearest of the three configurations at $0.4155, the $300 a primary care patient pays in a year equals the run cost of 722 calls ($300 / $0.4155)80. The $887 a dental patient spends equals 2,13555. With fixed costs added, Table 6 asks law and accounting firms for fewer than five new clients a year in every cell but one: 5.2 bookkeeping clients for the lowest plan and the agent. In health care, five of the six cells ask for more than five. The main reason is that a patient is worth less, $300 or $887, and in the full-system column the program adds to it, which Table 7 sets against everything that runs on it. The deciding variable is therefore the number of clients a purchase adds, which is N times m times a times c for the agent and dN times c for the plan. A firm can learn it only by recording what each contact became, which is why the paper counts measurement among the complements.

The cost side can still mislead, as the case of a company that cut cost per contact shows. Klarna said in February 2024 that its AI assistant, a month after launch, was doing the equivalent work of 700 full-time agents38. It was estimated to drive a $40 million improvement in profit in 202438. These are the company's own unaudited statements and the release states no method. Entrepreneur (secondary) later reported that Klarna's chief executive told Bloomberg in May 2025 that the company was hiring human customer service workers again. The report adds that it was critical to be clear to the customer that a human would always be available if wanted111. If it is accurate, it fits this chapter's point that a lower cost per contact does not settle what the contact is worth. Chapter 10 reads the case.

The contacts, missed calls and booking shares behind Table 8 are Patient Prism†'s figures. The share of missed callers an agent holds, containment and error rates rest on Chapter 4's estimates, built mainly on one outbound hospital study; the first two likely err high for a front desk and the third low. F here is the price list, which excludes the time of the person who reads the calls and keeps the record (Chapter 12 prices it).

A practice that handles protected health information may not be able to use the published price. In our check of 28 tools, 15 sign a business associate agreement only on certain plans, a list that includes Twilio, ElevenLabs and Vapi, and 6 of those 15 do not publish the price of the qualifying plan20.

Chapter 6

Demand comes first: AI needs a funnel to stand in

AI lead generation starts upstream: the capture term is zero when N is zero

The system's ledger carries a line above the agent's four mechanisms: demand, written dN x c x V. Here dN is the contacts a month that the system adds, c is the share of those contacts that become booked customers, and V is the value of one new customer over the first year, as gross margin. Chapter 4 sets out the line. The line comes before the agent because of the form of the agent's own capture term, N x m x a x c x V. N is the inbound contacts a month, m the share of them that go unanswered, or unanswered in time, without the agent, and a the share of those otherwise-missed contacts that the agent answers and holds; c and V are as above. A product is zero when any factor is zero, and the factor that nothing inside the agent can change is N.

An agent attached to a line that rings twice a month has almost nothing to capture, however well it speaks, and an agent with no flow of contacts captures nothing. Once the agent works, only two things can raise its term: more contacts, or better ones, meaning contacts more likely to book and worth more when they do. Both are set upstream of the agent, by who is brought to the door and for what. The agent can change a and c on the contacts it receives but cannot summon them. The part of N that the system summons is dN, and the plan does the summoning.

How many added contacts does the plan need to cover its own fee? The scope rule sets the plan against new clients, and the answer comes from cases built for this paper from cited figures. Marketing alone carries no patient information, so a dental or medical practice can buy a plan, at $999 a month, with no HIPAA program18,56. A primary care patient is worth $300 in the first year on the payments data, which is 1.61 visits at the $186 mean payment for a primary care office visit79,80. The $300 averages over everyone, including people who made no visit, so a person who seeks a practice out is likely to bring more79. The plan is covered by 3.3 new patients a month, counting the first year only ($999 / $300).

A physician has 51.1 appointments a week107, or 221 scheduled a month (calc.), and 18.8% to 23% of appointments are missed76,77. Of the kept visits, 12.5% are with new patients108, which is 21 to 22 new-patient visits a month now (calc.). The plan's 3.3 is then 15.6% more for one physician, and 3.9% more than 85 to 90 for four physicians (3.3 / 21 to 22 and 3.3 / 85 to 90).

Counted over three years of a patient's visits, the figure falls to 1.1 new patients a month (3.3 / 3), and over 8 years, the median length of the relationship in one study of patients aged 65 and over81, to 0.4 (3.3 / 8). Solo primary care is the tightest case in the model: the whole system, with the agent and a HIPAA-ready program, needs 7.1 new patients a month on first-year payments, which is 33.2% more than now for one physician and 8.3% more for four. For a general dental practice the plan is covered by 1.1 new patients a month at the $887 a year that AHRQ reports for spending per person with a dental visit ($999 / $887)55. That is 3.8% more (1.13 / 30) than the 30 new patients a month that Henry Schein One† reports as the average among its customer practices of one to seven locations13. Its page does not define a new patient.

The dental count takes all of a patient's spending as the practice's gain. For the whole system at $2,122 a month in the first year, the same count runs from 2.4 new patients a month at a 100% share to 10.7 at 22.3% of billings, the average general dentist's net income share of billings and a floor on what a practice keeps, because rent and salaries are already paid53. The counts are also of new customers, which is dN x c, so the contacts the plan must add are those counts divided by c.

A measure of c for the contacts a plan adds is NF. The nearest evidence is two neighbors that give a figure each, and we do not blend them. Patient Prism† (call analysis) reports that of every 100 dental patient calls, 69 connected and 21 ended in a booked appointment, existing patients included34. That is 21 bookings in 69 connected calls (21 / 69), about 30%, and it likely errs high for new inquiries, because existing patients book at high rates. Invoca† (call tracking) reports in its health care data that 43% of calls answered by a person were leads and 45% of those converted on the call64, or about 19% of answered calls (43% x 45%). It likely errs low, because it counts only bookings made on the call.

At Patient Prism†'s 30%, the plan's 3.3 new patients a month for one physician would take ~11 added contacts, and its 1.1 for a dental practice ~4 (each count divided by 30%; likely low, because the 30% likely errs high). At Invoca†'s 19% the figures are ~17 and ~6 (each count divided by 19%; likely high, because the 19% likely errs low). Each comes from one vendor's measure, and the rest of the chapter relies on neither.

What produces dN is the work of the first three stages of the funnel in Table 3: being found, being chosen and being easy to contact.

On the cost side, Technical Marketing's AI voice agent is priced at $299 a month with 250 call minutes included and $0.39 a minute after that18. The included minutes are about 62 four-minute calls (250 / 4), and every case in this chapter that sets a gain against the agent stays inside them, so the agent costs $299 a month. A dental or medical practice also needs a HIPAA-ready program around the agent, Essentials, at $699 a month after $1,500 setup, which is $824 a month over the first year ($699 + $1,500 / 12)56. The program is a compliance layer and does not include the agent. The site assistant, texts to patients and the front-desk add-ons run inside it too, so its fee is shared by everything on it56. The agent and the program together cost $1,123 a month in the first year ($299 + $824), which is the narrow case of a practice that runs nothing else on the program, and $2,122 with the lowest plan ($299 + $824 + $999).

A dental location in Patient Prism†'s data has about 116 patient calls a month, of which 31% hang up before reaching anyone, which is 36 a month (31% x 116)34.

Because those callers are new and existing patients together, the model values a booking at two ends: one visit at about $282, a rough figure from the ADA's billings per dentist and a visit count whose source is not named, and a new patient's first year at $88753,54,55. To cover its own $299, the agent would have to turn 3% of those 36 calls into conversations if every booking were a new patient, and 10% if every booking were one visit (the 36 hang-ups hold 11.0 bookings, 36 x 21 / 69, worth $9,730 or $3,089 a month; $299 / $9,730 and $299 / $3,089), against 12% and 36% in the narrow case of a program bought for the agent alone. The chain is set against the agent's own fee only. The program is shared by everything that runs on it and is not charged to answered calls alone. The plan's product is new demand, counted above, and setting answered calls against the plan too would charge a fee with nothing credited for it.

Those Patient Prism† figures count hang-ups, not voicemail or unanswered rings, and the chain assumes that callers who hang up resemble callers who get through. It also counts none of the callers who ring back, which would lower the gain from every conversion and so raise every share needed, most of all at the one-visit end where the callers are existing patients. The fixed fee does not fall when the line goes quiet, so a line with fewer calls needs a larger share of its fewer hang-ups, and a firm cannot judge whether the fee is worth paying without knowing N. Chapter 5 puts the cost side together, and Chapter 11 sets the full ledger for each profession.

Demand reaches the other three mechanisms less directly. Conversion, N x (1 - m) x dc x V, is proportional to N like capture. Capacity, H x w, is the value of hours released by handling contacts, and a firm with few contacts has few hours to release. Retention, B x dr x L, multiplies the customers on the books, and B is the stock that past demand built. So we infer, from the structure of the identity, that every term in it traces back to a flow of customers who wanted the service and found the door. The argument has an omission: the identity has no term for referrals, so a customer served well who sends a friend raises later N by a path the identity does not model. Surveys count referral as a route, and the referral section below reports them. They count how often a route is named, not how many new clients one served client sends, and this paper does not size that path.

One objection is that every firm with a telephone already has demand, so the point is idle. But demand has a size and a source, and a firm that can state neither is guessing. A firm that says its problem is missed calls holds a hypothesis about N and about m, and the agent's value rests on both. Chapter 7 covers how the firm can find out.

A second objection comes from the busy practice: when the schedule is full, more demand is worth nothing. The premise fits some practices and not others, and the conclusion does not follow either way. The American Dental Association's Health Policy Institute reported for the fourth quarter of 2025 that one-third of dentists were not busy enough, 12% were too busy to treat all patients requesting care and 18% treated all of them but were overworked89. For the third quarter of 2026 it reported that nearly two in five were not busy enough74. In Clio†'s data, the average law firm puts 38% of an eight-hour day toward billable work75. Neither profession is uniformly full.

A full schedule changes the form of the gain from demand but does not remove it. Added inquiries let the practice fill the slots that cancellations and no-shows empty, and those are common: a VA medical center's ten main clinics averaged an 18.8% no-show rate, and a review of 105 studies found about 23% on average76,77. They also let the practice choose better-paying work over the work it takes now, raise its minimum fee, and make the case for another provider when the waiting list stays long. A count of empty hours would miss each of these gains from dN.

AI receptionist, AI agent and AI voice agent: what search volume shows about demand for the tool

Table 1 sets the phrases that name an AI tool beside those that name an outcome the tool serves or a part of the system around it, from Google Ads keyword data for the United States collected in November 2026.

PhraseWhat it namesSearches a monthCost per clickSource
ai receptionistA tool74,000$43.665
ai agentA tool49,500$18.255
ai voice agentA tool33,100$49.075
ai answering serviceA tool4,400$58.485
ai customer serviceA tool4,400$181.525
ai lead generationA tool1,600$63.025
ai sales agentA tool1,300$69.945
marketing funnelAn outcome or system part9,900$19.045
conversion rate optimizationAn outcome or system part2,400$25.505
call trackingAn outcome or system part1,600$49.175
speed to leadAn outcome or system part880$18.535
lead managementAn outcome or system part880$58.045
marketing attributionAn outcome or system part720$46.415
lead response timeAn outcome or system part50$68.025
how to get more patientsAn outcome or system part70$34.115
legal lead generationAn outcome or system part70$67.605

Table 1. Search phrases that name an AI tool and phrases that name an outcome or part of the system around it: US searches a month and average cost per click, Google Ads keyword data, November 2026.

The table has limits. The phrases are ones an owner, a practice manager or a vendor would type, and none is a phrase a customer would type to find a service, so the table describes demand for tools, not for services. Each volume is an average monthly figure reported in rounded steps, so two phrases one step apart are not distinguishable, and phrases the keyword tool treats as variants share one figure, so we never add them. Volume shows interest, not purchase: the searcher may be an owner ready to buy, a student, a journalist or a vendor, and the data cannot say which. The phrases are also the ones we chose to look up, so a phrase missing from Table 1 may be searched more than any phrase in it.

The phrase "ai receptionist" is 7.5 times "marketing funnel" (74,000 / 9,900), 84 times "speed to lead" (74,000 / 880) and 1,480 times "lead response time" (74,000 / 50). Below the top the picture is mixed. "Marketing funnel" is searched more than "ai answering service", and "conversion rate optimization" more than "ai lead generation" and "ai sales agent"5. What holds across the table is narrower. The three most-searched phrases, "ai receptionist", "ai agent" and "ai voice agent", name tools, and the three smallest are "lead response time" at 50, "how to get more patients" at 70 and "legal lead generation" at 705.

"Ai customer service" draws the highest price in the table, $181.52, against $68.02 for the most expensive phrase on the outcome side, "lead response time"5. An advertiser's price per click shows what advertisers on that phrase will pay for a visitor, not what a tool earns its buyer. Chapter 1 read the tilt of this data the same way: we think far more people are looking to buy AI than are asking whether it pays. Dentistry is the exception, documented in Chapter 5 of the companion paper, Technical marketing for professional practices. There "dental ai" is searched 880 times a month against 1,300 for "dental marketing"5, and the front-desk phrases are small: "ai dental receptionist" at 170, "dental answering service" at 110 and "how to get more dental patients" at 505.

Tool demand is the wish of people who run businesses to buy something that answers calls or writes copy. Service demand is the wish of customers for the heating repair, the cleaning or the dental visit, expressed as a call, a form or a message at a firm's door. Our keyword data measure tool demand and do not measure service demand for any firm. The capture term needs service demand, because N counts the contacts that arrive and service demand is what produces the ones worth capturing. Heavy tool demand in the market tells a single firm nothing about whether its own service demand is enough for a tool to act on.

How customers reach a service business: search, maps, reviews and AI answers

Customers reach a service business by several routes at once, and no single instrument counts them all. Table 2 lists nine instruments that touch the question from different sides. The instruments sample different populations, in different years, with different definitions of use.

InstrumentWho or what is countedWhat it measuresFigureClaim
Our study of assistants' answers, 2026-11-10160 local questions about ten kinds of business in 16 US cities, each put once to three assistants through their APIsShare of businesses named that all three assistants named3.5% (84 of 2,414 names)19
Pew Research Center, 2025Google searches in the studyShare of searches that produced an AI summary18%112
SparkToro† and Datos clickstream analysis, 2024 (secondary)Google searches in a clickstream datasetShare of US searches with zero clicks58.5%113
Pew Research Center, February 20265,119 US adultsEver use AI chatbots such as ChatGPT, Gemini or Copilot49%114
KFF Tracking Poll, early 20261,343 US adultsTurned to AI chatbots for health information in the past year32%115
Chatterji and co-authors, study of ChatGPT use, 2025Adults worldwideAdoption of one consumer product by July 2025About 10%9
BrightLocal† consumer surveys, 2025 and 2026 (vendor surveys)Two separate panels of US adultsUsed AI tools to find local reviews (2025); used generative AI for local recommendations (2026)6%; 45%116,117
Clio† consumer survey, 2019 (secondary)Clients seeking legal representationSought a referral; searched on their own59%; 57%118
Patient Prism†, dental calls in 2025Attributable patient calls to dental locationsShare of those calls that came from the Google ecosystem90%34

Table 2. Nine instruments about how people reach local services, and what each one counts. The figures are not comparable with one another.

Our own study is the one instrument built around the customer's question. Each question had the form "Who is a good dentist in Atlanta, GA? Name specific businesses.", with the city and the kind of business changed19. The assistants mostly named different businesses19, so we think there was, in this run, no single AI answer for a firm to get into. The study asked each question once, matched names by spelling, read up to eight names per answer and used answers obtained through an API, which are a baseline, not what every signed-in user sees19. It does not show that any assistant is wrong, that a named business is good, or that any answer produced a call. The companion paper reports it in full, by profession, in its Chapter 4.

Clicks when an AI summary appears

In Pew's study, 18% of all Google searches generated an AI summary, and 88% of those summaries cited three or more sources while 1% cited a single source112. A firm's own page is therefore one candidate among several in most summaries.

In Pew's analysis of US adults' browsing data, those who met an AI summary clicked a traditional search result link in 8% of visits, against 15% of visits without one112, and they ended the browsing session after 26% of pages with a summary, against 16% of pages with only traditional results112. Pew's search-page dataset held 68,879 unique Google searches, of which 12,593 produced an AI summary112. Users who met a summary clicked a link inside the summary itself in just 1% of visits to pages with one112. These are differences between visits and pages with and without a summary, not a test of what a summary causes. The claims describe all visits and all searches and do not isolate searches for a local service. We read the figures as the direction of the difference and put no click rate into the profit identity.

A second click figure is of a different kind: a clickstream analysis by SparkToro† and Datos (secondary) reports that 58.5% of US Google searches in 2024 ended with no click113. It covers 2024, a year before Pew's study, its unit is the search where Pew's click figure counts visits, and it is not an AI figure. The claim does not say how a tap on a phone number or a map listing was counted, which matters for a local service business because a search can end in a call without a visit to the website.

Vendors have measured what an AI Overview does to clicks, in datasets they chose. Ahrefs† (SEO software) compared 300,000 keywords, December 2023 against December 2025, and found that the presence of an AI Overview correlated with a 58% lower average clickthrough rate for the top-ranking page119. Seer Interactive† (search marketing) found, on 3,119 informational search terms, that organic clickthrough on queries with an AI Overview fell 61%, from 1.76% in June 2024 to a low of 0.57% in July 2025, and stood at 0.61% in September 2025120. Overviews may sit on queries that already drew fewer clicks, so these correlations are not the effect of an Overview on a click. Seer† also reports that brands cited in an Overview got 35% more organic clicks than brands not cited, an association that brand strength could explain120.

A third measure comes from a preprint. Khosravi and Yoganarasimhan compared English Wikipedia with its German and French editions around Google's staggered rollout of AI Overviews, and estimate that default availability reduced English search traffic by 5.45% and 4.82%, depending on the comparison language121. It is one site of informational articles, in a paper now in its seventh version. Google's head of Search wrote in August 2025 that total organic click volume from Google Search to websites had been relatively stable year over year, a statement with no data attached122. In the vendor rows of our evidence, none measures calls, bookings or clients at a dental, medical, legal or accounting practice.

Whitespark† (local SEO) ran 540 queries by hand in Houston, Phoenix and Denver across six industries, including dentists, medical clinics and personal injury lawyers, and found AI Overviews on 15% of local-intent queries against 92% of informational-intent queries, with local packs on 93% of the local-intent queries123. The authors chose the queries, and the snapshot is from May 2025. Conductor† (SEO platform) found Overviews on 48.7% of the page-one Health Care queries it tracks, in a four-week window ending October 12, 2025124. SE Ranking† (SEO software) found them on 233 of 300 legal keywords (77.67%) and on 65.33% of 300 health keywords in September 2024125. Those are mostly informational queries, such as legal definitions, not searches for a lawyer near the searcher.

How often Overviews appear on accounting or tax searches is NF; the nearest evidence is finance, where Conductor† found Overviews on 25.8% of the Financials queries it analyzed, a sector of banks, insurers and financial services, and SE Ranking† found them on 125 of 300 finance keywords (41.67%)125,126. The two used different queries and dates. We estimate that informational tax questions sit near the finance figures. The estimate likely errs high for a search for an accountant near the searcher, because local-intent searches drew Overviews far less often than informational ones in Whitespark†'s snapshot.

We think the figures, read together, fit a division of the work. Informational questions are where Overviews are common and where the clicks they take were measured; a firm's pages meet the searcher there, before any provider has been chosen. Local-intent searches are where Overviews were rarer and the map pack nearly always present in Whitespark†'s snapshot, so the profile, the reviews and the distance decide. Both feed dN, by different routes. Google changes when Overviews appear, so each figure is a dated snapshot.

What AI answers send to a site

AI answers send few visits so far, in the vendor datasets that count them. Conductor† found AI referral traffic at 1.08% of all website traffic across the ten industries it analyzed, and .64% for its Health Care domains124,127. Microsoft Clarity† found AI referrals still under 1% of overall traffic on more than 1,200 publisher and news sites, despite 155.6% growth in AI-driven traffic over eight months128. Neither dataset is a sample of small practices, and a click count misses answers nobody clicks. OpenAI's help page says that for a local business a mention counts when ChatGPT shows a result for that location, even if the customer does not open the listing or visit the website, and that its Insights report mentions and website clicks129.

AI referral share on a dental, medical, legal or accounting practice's site is NF. The nearest evidence is the .64% that Conductor† reports for its Health Care domains and the .48% for Financials, and we estimate the share is under 1% of visits124,126. It likely errs high for the clicks themselves, because those are large brands' sites, and low for the influence of answers, because a mention with no click is not counted. We form no estimate of conversion, because the studies we found cover retail and publishers and their gaps point both ways. Adobe reported that in September 2026 AI-referred visits to US retail sites converted 90% better than other visits, just over a year after they had converted 9% worse130, so the sign of that gap reversed. Clarity†'s publisher data show sign-up conversion of 1.66% for traffic from language models against 0.15% for search128.

Who asks assistants

Surveys of adults describe use, not calls. In Pew's February 2026 survey of 5,119 US adults, 49% said they ever use AI chatbots, 42% ever use them to search for information, the most common use Pew asked about, and 20% to get medical advice114. A June 2026 Pew survey of 3,488 adults found that 34% ever use chatbots for at least one of eight health reasons, and 15% cited helping decide whether to see a doctor, a share whose base Pew's sentence does not state131. KFF's poll of 1,343 adults, fielded February 24 to March 2, 2026, found that about a third (32%) had turned to a chatbot for health information in the past year, and that 42% of those who asked about physical health did not follow up with a doctor or other health professional afterward115. The follow-up figure cuts two ways: some who asked never needed a visit, and for others the answer may have stood in for one.

Vendors measure local recommendations. BrightLocal†'s 2026 survey of 1,002 US adults reports that use of ChatGPT and other generative AI tools for local recommendations rose from 6% in 2025 to 45%, making them the third most popular source of business recommendations117. The question wording is not printed, the two years are separate panels, and a jump of more than seven times may partly reflect a change in wording. The 2025 survey (secondary source) had reported that 6% of consumers used ChatGPT or other AI tools to find local reviews, down from 9% in 2024116. A survey by the consulting firm Bain of 1,500 consumers in September 2025 found 56% mostly or always using search engines as their default search tool and 16% using chatbots132.

By July 2025 around 10% of the world's adult population had adopted ChatGPT's consumer product, and its three most common conversation topics, "Practical Guidance", "Seeking Information" and "Writing", accounted for nearly 80% of all conversations9. Those claims do not isolate questions about local businesses. Annual surveys need not share a sample, so we read no trend into any of these figures.

Referral, and where calls come from

Clio†'s 2019 survey in the legal market (secondary) found that 59% of clients sought a referral to an attorney and 57% searched on their own118. The two figures add to 116% (59% + 57%), so if both are shares of the same clients, at least 16% did both, and in that survey referral was a route about as common as search. It is one profession in one year, but it shows that a firm's contacts do not all begin in a search box.

Later surveys point the same way without measuring it. FindLaw† (legal marketing) reports in its 2024 survey that 48% of respondents who learned about the attorney they contacted used referrals from family, friends or trusted coworkers, the most popular offline resource, which is a way of researching, not proof of how they found the lawyer133. MouthWatch† (teledentistry) reports that 65.6% of 448 dental practices named word of mouth and patient referrals as their top source of new patients in 2026, down from 77.5% in its 2025 survey, and 15.6% named Google Business Profile134. Those are practices' opinions of their sources, not measured attribution. In Accounting Today's survey of firms, 79% named unsolicited client referrals a top source of new clients, in a multiple-choice question135. Press Ganey† (patient experience) reports that 83.5% of consumers with a referral in hand at least occasionally turn to online reviews to validate it136.

The one count of where calls come from is Patient Prism†'s. In its 2025 dental data, 90% of attributable patient calls came from the Google ecosystem, and Google Business Profile alone drove 54% of attributable inquiries34. Those are the calls it could trace. A share of new patients or clients by source, taken from practice records, in any of the four professions is NF. The two nearest measures count different things, traced calls in one case and practices naming a top source in the other, so we form no estimate from them.

The routes overlap, and the mix for one firm is unknown to this paper, so each firm has to measure its own. We infer the design consequence: a firm should be readable on the surfaces every route passes through, its own pages and the third-party profiles, and each surface needs a way to make contact.

How much control a firm has

The sources say a firm's control over its place in AI answers is limited. Google states that a page must be indexed and eligible to be shown in Search with a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional technical requirements137, and that existing SEO fundamentals remain worthwhile137. The paper that introduced generative engine optimization reports that its methods can raise a source's visibility in generative engine responses by up to 40%, and that content creators have little to no control over when and how their content is displayed138. The 40% is a best case in the authors' own setting138, and the authors add that efficacy varies across domains138. Our study tests none of those methods, and we promise no place in any assistant's answer.

The makers say little more. OpenAI's help pages, read on November 10, 2026, say that ChatGPT ranks search results using multiple factors it does not list and that placement is not guaranteed139. They also say that ChatGPT may rewrite a prompt such as "good restaurants near me" into a search for the general area it infers from the user's IP address139. A statement from the makers of Gemini, Claude or Perplexity on how they choose local businesses is NF; the nearest evidence is OpenAI's help pages above.

Assistants also vary from one asking to the next. SparkToro† and Gumshoe.ai† had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI, 2,961 runs in November and December 2025. They report less than a 1 in 100 chance that ChatGPT or Google's AI, asked 100 times, would give the same list of brands in any two responses, and less than 1 in 1,000 for the same list in the same order140. The study is not peer reviewed, its author discloses a possible conflict of interest, volunteers pasted the answers by hand, and the prompts did not concern dentists, doctors, lawyers or accountants.

A repeat-asking study of those four professions is NF; the nearest evidence is SparkToro†'s under 1 in 1,000 for the same order even among limited options such as Los Angeles Volvo dealers140. We estimate that the same list is rarely repeated there either, because a city has more dentists than that. The estimate likely errs toward too much variation for a small town with few practices of a kind, where lists may repeat more. In our own study, 79.7% of the businesses named were named by only one of the three assistants19. Together the two say that a name in one answer is a draw from a changing list, not a position, so a firm cannot count on the assistant route and has reason to measure what arrives.

Google as a funnel: marketing funnel stages and where an AI agent sits

Traditional web design treats Google as a phonebook: the business is listed, the customer finds the number, and whoever picks up the phone is the rest of the process. A technical marketing system treats Google as the top of a funnel, which we define here as a sequence of stages a customer passes through, with a count at each stage, so that loss can be located. The search result, the map listing, the business profile and the AI answer are the first stages of a process the firm designs and measures from the customer's first sight of the firm to the customer's return. Table 3 gives the stages we use for a service business and marks where an agent can act.

StageWhat the customer is doingWhat the firm controlsWhere an agent can actWhat leaks if the stage fails
1. FoundSearches, asks an assistant, scans a map listing or a directoryIts profile, its pages, the facts on third-party sitesRarely: the agent is not visible at this stageCustomers who never learn the firm exists
2. ChosenCompares ratings, reviews, prices and answers, then picks one firmThe reviews it earns, the answers on its pagesA chat that answers a comparison questionCustomers who picked another firm
3. ContactedCalls, submits a form, opens a chat or sends a textThe doors and how easy each one is to reachThe agent can be the door for calls and chatCustomers who tried and gave up
4. Answered and qualifiedAsks, is asked, is sorted by kind of job and urgencyWho or what answers, the rules, the hand-offThe main work of the agentContacts unanswered, or answered wrongly or slowly
5. BookedTakes a slot or agrees a next stepThe calendar and the rules for opening itThe agent books on the real calendarContacts that end without a next step
6. Served and keptReceives the work, pays, may return or referDelivery, reminders, follow-upReminders and follow-upCustomers who do not return

Table 3. Six stages of a service business's funnel, and where an AI agent can act. The stages are our framework.

The notation of the identity maps onto the table. N counts the contacts that pass from stage 3 into stage 4, door by door, and dN is the change in that count that stages 1 to 3 produce when the firm is found, chosen and easy to reach. The share m is leakage at stage 4, the contacts that the firm fails to answer or answers too late. The shares a and c describe what an agent adds at stages 4 and 5, and B and dr describe stage 6.

First, a funnel with counts at each stage lets a firm see where customers leave, and a phonebook does not, because it records nothing after the listing. Second, an agent at stage 4 cannot repair a leak at stages 1 to 3: if the firm is not found, or is found and not chosen, the agent receives no contacts. Third, weakness at stage 2 can reach the agent as poor demand, not as no demand. Contacts that arrive for the wrong service, in the wrong area or with the wrong expectation lower c, and the firm may then blame the agent for a loss that the upstream stages caused.

At stage 1, position on the results page and closeness to the searcher appear to matter. A 2013 study of mobile internet search by Ghose, Goldfarb and Han reports that ranking effects are higher on mobile phones, so links that appear at the top of the screen are especially likely to be clicked, and that stores located in close proximity to a user's home are much more likely to be clicked on mobile phones141. The study predates AI summaries and its claims do not say which businesses it covered, so we cite it for a mechanism, not a magnitude.

At stages 2 and 3, the work is to give each surface a job, a next step and a record. In our practice, each page that targets a question answers it in its opening lines in a form a person or an assistant could quote, and calls, forms, chats and bookings are tied to the search or ad that sent them. In the design we describe, a profile also carries the facts a customer needs and a way to call, and each page has one primary door, tagged so that the contact's source is recorded when it arrives. None of this guarantees a position. The funnel therefore rests on the surfaces the firm controls, and every contact, from whichever route, lands in the same record.

The agent works in the middle of the funnel, at stages 4 and 5, with a reach into stage 6 through reminders and follow-up. It is not the top of the funnel or a replacement for one. We think this is the main reason to describe an agent as a rep, not as a source of leads: it works the contacts that arrive, and the phrase "AI lead generation", searched about 1,600 times a month5, promises something that belongs to stages 1 to 3.

The website as a working rep: what changes when the site can answer, quote, book and hand off

A billboard shows the business and stops there. A working rep does five jobs: it answers, quotes, books, qualifies and hands off. That is the paper's working idea of what a site is for, and it depends on what exists behind the page. A rep needs a source of facts the firm keeps current to answer, a rule about prices to quote, a calendar to book on, a record to qualify against and a person to hand off to. A page with none of these can look the same as a page with all of them; the difference shows when a visitor does something, not when the page loads.

Suppose a heating and cooling company has a page on furnace repair, built first as a billboard. It states the company's services and shows a phone number and a form. The form sends an email to an inbox that someone reads in the morning. A visitor at 9:40 in the evening who has no heat reads the page, finds nothing that answers the actual question, and either calls a number that rings out or leaves. The rep version of the same page can state whether the firm serves the visitor's town, ask what the system is doing, and offer the first real opening on the calendar the staff use. When the answers put the case on a list of urgent calls that the firm wrote, the rep rings the named on-call person, and the booking and the reason reach the office as a summary. The visitor's source, the conversation and the booking land in one record. We describe a design here and report no result; nothing in this paragraph is a statistic about how often sites do any of it.

Each of the five jobs has a limit that the design must state. Answering is limited to facts the firm supplied and the page contains. Quoting is the job most likely to go wrong, because a number given in the firm's name is the firm's number. A rep can publish a fixed price the firm approved, give a range with the conditions attached, or book the visit that produces the quote, and it must not invent a price. Chapter 9 treats the cost of a wrong statement as a term in the identity. Booking is limited to openings that are real, which means the rep and the staff use one calendar. Qualifying is limited to the questions that decide who should handle the contact. Handing off is limited by the existence of a person who will receive it.

The effect on demand is indirect. A working rep leaves N unchanged on the day it is switched on and changes the shares m, a and c, the things the agent is for. We expect it to change what the upstream stages receive over time, because a site that answers specific questions in plain, quotable text gives search and assistants more to send people to. That is how SEO, web design and technical systems pay when they are built together: the page that wins the click and the rep that works the contact are one design. Bought separately, from vendors who share no map of the contacts, they risk a page that wins clicks it cannot route and an agent that never learns where its callers came from.

Online reviews and third-party profiles: demand infrastructure for AI answers

Reviews and third-party profiles belong in a chapter on demand because they are part of how a customer gets from stage 1 to stage 3 and part of what an assistant reads. The studies of reviews in our evidence come from other markets, so Table 4 lists what each one measured, with one study of a physician practice's new patients. Restaurants and books are not dental practices, so after the table we say what transfers.

StudyMarketWhat was measuredResultClaim
Luca, 2016Restaurants, Yelp ratingsRevenue against star ratingA one-star increase led to a 5 to 9 percent rise in revenue; the effect was driven by independent restaurants, and ratings did not affect chain-affiliated ones142
Anderson and Magruder, 2012Restaurants, Yelp ratingsSell-out frequency at peak reservation times against a half-star ratingAn extra half-star made restaurants sell out 19 percentage points (49%) more frequently, with larger impacts when other information was scarcer143
Chevalier and Mayzlin, 2006Books sold at two online sitesRelative sales at a site against reviewsBetter reviews raised relative sales at that site, and for most samples 1-star reviews had more impact than 5-star reviews144
Gao and co-authors, 2012PhysiciansPatients' online ratings on RateMDsMean rating 3.93 of 5; 45.80% of 112,024 physicians had a perfect 5145
BrightLocal†, 2023 (secondary)Financial and legal sectorStated importance of reviews81% of consumers called reviews important or very important146
Block, Reddy and Corkum, 2023One large single-specialty practice, New York and New Jersey areaNew patient volume after a 1-star physician reviewMean new patient volume decreased 2.3% to 2.6%, with effects lasting at least 16 weeks147

Table 4. What the review studies in our evidence measured, and in which market.

The first three rows report effects of ratings or reviews on sales, for restaurant meals and books, which are small purchases next to a course of dental treatment or a legal matter. Luca's result uses Washington State Department of Revenue data on restaurants148. A Berkeley press release (secondary) says Anderson and Magruder analyzed 148,000 Yelp reviews for 328 restaurants in the San Francisco Bay Area149. The book study compares relative sales at Amazon.com and Barnesandnoble.com144.

The fourth row describes what physicians' ratings look like, not what they do to demand, and the claim itself reports that online reviews were generally quite positive145. A high rating was therefore common among the physicians rated, and the claims behind Table 4 do not say which features of a profile separate one practice from another.

The fifth row is a survey (secondary) of stated attitudes, which tells what people say about reviews, not what they did. In BrightLocal†'s 2023 survey the 81% was up from 66% in 2022146, and we do not treat the two years as a trend. Later vendor surveys say the same in stated terms. BrightLocal†'s 2026 survey of 1,002 US adults reports that 97% of consumers read reviews for local businesses and that 41% always do, against 29% the year before (separate panels)117. FindLaw† reports that 82% of respondents who contacted an attorney after learning about them online used online reviews in their decision133.

The sixth row is the nearest to a physician's front door, and it is an association. The result comes from a retrospective analysis of 1.12 million new patient appointments and 12,882 physician reviews147. It is a before-and-after comparison, not a trial, at one practice in one metro area and one specialty, and only 6.7% of its reviews were 1-star147. We take it as a measured size for the review channel in one practice, not as a rate for primary care.

The mechanism that the Anderson and Magruder result points at may transfer: ratings had larger impacts when other information was scarce143. We think a customer choosing a heating contractor, a dentist or a lawyer usually has little other information. Day and Barksdale (secondary) argue that professional services are typically intangible and customized, which makes evaluation before purchase difficult71. Darby and Karni (secondary) distinguish credence qualities, which are expensive to judge even after purchase, from search and experience qualities72. On that reasoning reviews may carry more weight for such services. So may the pattern in Luca's result, that ratings mattered for independent restaurants and not for chains142. We infer that most small service businesses resemble the independent, with a name that few customers know in advance.

Other differences may stop the result from carrying over. The first is the size and kind of purchase: none of the three effect studies concerns one made once a year or once a decade, and the book study measures relative sales between two sites, not a firm's bookings144. It also found that for most samples 1-star reviews had more impact than 5-star reviews, which suggests that poor reviews may weigh more than good ones, in books. The second is who can judge. A diner can judge a meal, and a reviewer can rarely judge the clinical or legal quality of what was done, so reviews of professionals may reward manner and convenience over competence. That is a hypothesis this paper does not test. The third is rules: how a practice may ask for reviews, respond to them and use them in its own advertising is covered in the companion paper's chapters on dentists, doctors and lawyers, and Chapter 9 of this paper covers the FTC's rule on fake and false reviews.

Our own study adds the other half: what the assistants read. The 25 most-cited websites in our study drew 29% of all citations, and each was cited in at least 9 of the 16 cities. They are directories, review platforms, list sites, publishers and a forum, not any one local business's site19. Websites cited in only one city, where a local business's own site would fall, drew 39% of citations19. The study counts names and cited sites separately, so it does not show that a cited profile caused a name to appear19. It does show that the pages an assistant cites for a local question are often third-party pages.

This is why we call profiles and reviews demand infrastructure. They are not the firm's own property, yet they supply facts such as name, address, telephone number, hours and services to the readers that stand between a customer and the firm. Keeping them accurate and consistent is work the firm controls, and its own pages and its agent should give the same facts. A telephone number on a profile is also a door, and a door that is not part of the system is a contact the system cannot route or count.

The failure pattern: an AI agent bought before the demand system exists

We think one way a small firm's AI project can fail is a failure of order, though this paper reports no case of a firm that bought an agent before building the demand system, and no figure for the share of failed projects that fit the description. The adoption polls count adopters without asking what came first. MGMA's poll of September 30, 2025 found that nearly 7 in 10 medical groups (68% of 351 applicable responses) reported adding or expanding use of AI tools in 2025150, and its February 10, 2026 poll found scheduling (31%) and calls (27%) the top areas where practice leaders were turning to AI to improve access36. Those are self-selected members' answers. The reading rests on the structure of the identity and on the evidence in Chapters 1 and 2.

An owner sees a problem at the front desk, such as missed calls or a full inbox, and searches for the tool that seems to fix it, a step that the search volume in Table 1 would fit but does not show. The owner buys an agent. The agent answers the contacts that arrive, but too few arrive, or the wrong ones, because the profile, the pages and the reviews that bring customers to the door were never built. The returns are therefore thin, and with no record of source, handling and outcome, the owner cannot tell a thin return from a hidden one. The project is then judged by feel, and a firm that cannot see what the agent earned concludes that AI does not work for its business.

Chapters 1 and 2 hold evidence that is consistent with that pattern and does not show it. Chapter 1 reported the MIT Project NANDA finding that 95% of organizations were getting zero return, a figure whose definition of return we could not recover3. It also reported Gartner's 2024 prediction (secondary) that at least 30% of generative AI projects would be abandoned after proof of concept, citing poor data quality, inadequate risk controls, escalating costs or unclear business value12. As Chapter 2 noted, none of those four reasons says the model could not do the task. Chapter 2's reading was that gains appear where the task is frequent, the output is checkable and the work sits inside a measured process. N is the frequency of the task for a small service business. A firm with a thin flow of contacts fails the first condition, and a firm with no record fails the third.

Our reading makes a prediction that could fail. Among small service firms that install an agent, those with a thin flow of contacts, as measured by N, should show less demonstrable return than those with a steady flow, other things equal. The reading would be wrong if agents in low-N firms paid through capacity or retention at rates similar to those in high-N firms, or if most disappointing deployments turned out to have failed on errors in what the agent said rather than on a lack of contacts.

A randomized or field study of an AI agent answering inbound calls or scheduling in dental, medical, legal or accounting settings that tests either branch is NF, and so is a peer-reviewed containment rate for inbound calls. The nearest evidence is Mount Sinai's study of outbound pre-procedure calls, where the voice assistant completed 86.4% of calls in Phase I17, and the vendors' own rates that Chapter 8 sets out. Two kinds of evidence would settle it: a field experiment that assigns agents to firms with different measured flows, and open reports of containment and booking rates tied to contact counts.

The consequence is an order of work. The firm should first make itself findable and chosen, give every door a record, and count the contacts by source. Then it can ask what an agent adds to a flow it can see. Chapter 8 describes the architecture that supplies the doors and the record, and Chapter 13 sets out the sequence. A firm that has done the first part may still find that its N is small and that an agent does not pay, and it then learns that before paying for the agent rather than after.

Chapter 7

The measurement system: knowing what the AI earned

Marketing attribution is hard even at scale: what the large advertising experiments show

Lewis and Rao (2015) report twenty-five large field experiments with major US retailers and brokerages, most reaching millions of customers, and find that measuring the returns to advertising is difficult48. The median confidence interval on return on investment across their experiments was over 100 percentage points wide48. They attribute this to noise. Individual-level sales are very volatile relative to the per capita cost of advertising, and a coefficient of variation (the spread of sales relative to their average) of 10 is common48. Informative experiments can easily require more than 10 million person-weeks (one customer observed for one week), which makes them costly and potentially infeasible for many firms48. The setting is digital advertising by large firms, not a small firm's telephone line.

The cheaper alternative is to compare outcomes without assigning anything at random, and the same authors warn against it. They call selection bias, due to the targeted nature of advertising, "a crippling concern for widely employed observational methods"48. When advertisements are aimed at people who are already likely to buy, a comparison between people who saw one and people who did not mixes the effect of the advertisement with the difference between the two groups. Gordon, Zettelmeyer, Bhargava and Chapsky used 15 US advertising experiments at Facebook, with 500 million user-experiment observations and 1.6 billion ad impressions, to set experimental results beside observational ones49. The observational methods often failed to produce the same effects as the randomized experiments, even after conditioning on extensive demographic and behavioral variables49.

Two more studies set observational estimates beside randomized ones and show how large the gap can be. In Lewis, Rao and Reiley's Yahoo! experiment, the randomized comparison gave a 5.4% increase in users searching the advertiser's brand keywords151. Matching exposed to unexposed users on their behavior in the prior week gave estimates of search lift from 871% to 1198%151. It was one display campaign for one advertiser, and the paper dates from 2011. In 2023, Gordon, Moakler and Zettelmeyer analyzed 663 Facebook experiments with access to over 5,000 user-level features152. They found median randomized lifts of 29%, 18% and 5% for upper, middle and lower funnel outcomes, against 83%, 58% and 24% from a double machine learning method applied to non-experimental data152. They concluded that they could not reliably estimate an ad campaign's causal effect from such data.

Neither study estimates the error of last-click attribution. The methods they test are observational matching, a before-and-after comparison and machine-learning models fitted to logged user data. Such methods can overstate an advertising effect, in these settings by multiples.

Blake, Nosko and Tadelis give the sharpest example, from eBay alone. For eBay's paid search, typical OLS methods gave a return on investment of over 4,100% without time and geographic controls and over 1,400% with them153. Experimental methods found a return of -63%, with a 95% confidence interval of -124% to -3%153. When eBay turned off brand-keyword paid search, which we take to mean ads on searches that already name the advertiser, 99.5% of the forgone click traffic was immediately captured by natural search traffic153. For non-brand keywords, new and infrequent users were positively influenced by the ads, but frequent users, whose purchasing was not influenced, accounted for most of the expense153.

The logic transfers to a small firm and the magnitudes do not. A booking that a system touched is not necessarily a booking the system caused, and the customer who would have come anyway is the hardest case to see. Berman adds a point about credit. In a game-theoretic model of multi-publisher advertising, he shows that the popular last-touch attribution method over-incentivizes ad exposures, often resulting in lower advertiser profits, and that the Shapley value achieves higher profits154. The result holds under the model's assumptions. Berman's paper also refers to what it calls a recent survey, in which 54% of advertisers indicated that they used a last-touch method154.

A measurement of how far last-click or platform-reported attribution misstates a firm's results is NF; the nearest evidence is the comparisons above. Gordon and co-authors' median observational lifts were 2.9, 3.2 and 4.8 times their randomized lifts (83 / 29, 58 / 18 and 24 / 5). We do not carry those ratios over to last-click, a different method on different data, and form no number for a small firm. We think any tag that credits a channel the customer would have used anyway is likelier to overstate than understate, as the brand-keyword case shows.

A firm that books a few dozen customers a month should not expect a precise return, given intervals that wide, and should ask for a sign and a plausible range. The device that Gordon and co-authors set against observational methods is the randomized experiment, and a small firm can randomize hours, weeks or lines, as the designs below do. And a figure from a dashboard or a simple before-and-after count describes what happened and does not measure what the agent caused.

How to measure AI ROI: what an agent's log records and what it lacks

An AI agent is easier to measure than an advertisement in one respect. Every contact it handles passes through systems the firm pays for. Chapter 5 priced those systems by the connected minute, so the count and length of calls exist on the invoice whether or not anyone reads them. A log can also hold a transcript, a summary, the door the contact used and what the agent did, including whether it booked or handed off. An advertisement is seen by people the advertiser cannot list, while the agent's contacts can be listed one by one, with no sample or modeled audience.

The agent is harder to measure in another respect, and that respect decides profit. The log records what happened during the contact and nothing about what the contact became or would have become. It lacks the outcome: whether the person who booked came, paid, left a margin and returned, all of which sit in the calendar, the books and the customer record. It also lacks the counterfactual: what would have happened to the same contact without the agent. The voicemail may have been returned the next morning, the caller may have rung again, or the caller may have rung a competitor.

Chapter 4's identity builds in both gaps. Its terms, in words, are the inbound contacts a month (N); the share of them that go unanswered, or unanswered in time, without the agent (m); the share of those otherwise-missed contacts that the agent answers and holds (a); the share of answered contacts that become booked customers (c); and the change in that share on contacts that were already being answered (dc). Value enters as the value of one new customer over the first year, as gross margin (V), the hours of staff or professional time released a month (H) and the value of one released hour in its next best use (w). Retention adds the customers on the books (B), the change in the monthly share retained (dr) and lifetime value (L). The costs are the run cost per contact (k), the contacts the agent handles a month (n), the fixed monthly cost of the system (F) and the expected monthly cost of error (E).

Only some of these terms are in the log. The count N, the agent's handling and the cost terms k and n are in the log or on the invoices. The term m is not. It is defined as the share unanswered without the agent, so once the agent is on, it describes a situation that no longer exists: the contacts that would have gone unanswered are answered, and nothing in the log marks them. To recover m, the firm needs the telephone system's record of calls that rang out, if the system keeps one, a flag that says whether a person was available when each contact arrived, or a period in which the agent is withheld. Chapter 12's third job, which reconciles the carrier's log with the agent's summaries, checks the first of these.

A short illustration from the dental chain in this paper's model shows the gap between what the log records and what the agent added. A location in the data of Patient Prism† (call analysis) has about 36 calls a month that hang up before reaching anyone34. Of every 100 calls, 69 connect, 34 of those are a real chance to book and 21 book34. If the hang-ups resembled callers who get through and every one were answered, they would yield 11.0 bookings a month (36 x 21 / 69). At 25% of them turned into conversations, that is 9.0 calls and 2.7 bookings a month. The source counts new and existing patients together, so the model values a booking at one visit or at a new patient's first year. The log would show the 2.7. It cannot show how many of those callers would have rung again and booked with the office anyway, and the chain counts none of the callers who ring back.

That share is the one link nobody has measured, so it is shown as a range. If a quarter of the bookings would have happened anyway, the agent added about 2 of the 2.7, and if half, about 1.4. The log reads the same in every case, and only a comparison can tell them apart. The identity's capture term counts every contact in m as lost, so it is an upper bound. A firm reads the term net of recovery, and the holdout described below is how it learns the recovery.

The log is also shaped by selection, since it holds only the contacts that reached the agent. A caller who ends the call on hearing that it is an AI assistant, as every caller in our practice is told, shows up only as a short call. A caller who never connected appears only in the carrier's log. Both records are needed, and a disagreement between them is a finding. The difficulty of seeing a benefit is an old one: Brynjolfsson and Hitt wrote in 2000 that the benefits of IT investment are often intangible and disproportionately difficult to measure14. Chapter 3 sets out why that matters for the timing of returns.

Closing the loop: contact, source, conversation, booking, revenue

A technical marketing system differs from a billboard in one way that matters here: every door writes to the same record. The loop has five links. The contact is that someone reached the firm, when, and through which door. The source is what brought them. The conversation is what was said and done, by the agent or by a person. The booking is whether a next step was fixed and when. The revenue is what the customer turned out to be worth, counted as gross margin. A sixth link, the customer's return, matters for the retention mechanism. The loop is closed when a contact can be followed from the first link to the last without guessing.

Chapter 8 showed which parts of the plumbing write the first four links and that the customer record holds the join. Table 1 sets out the measurement record.

FieldWhere it comes fromTerm it lets you observe
Contact number, date, time and door (main line, tracked number, form, chat, text)Telephony, form and chat toolsN, counted by door and by hour
Source tag: the number dialed, the page of origin, and the answer to "how did you hear about us" when askedA distinct number or code at each door; the form; the agent's questionWhere N comes from, so that a change in demand is not read as a change in the agent
Staff availability when the contact arrived (open, line busy, closed)The firm's hours and the phone system's statusm: contacts that arrive when no person is available approximate those that would go unanswered without the agent
Who answered first (agent, person, nobody) and after how longCarrier log and agent logm, and speed to lead
Agent disposition: held to the end, dropped, handed off, caller asked for a personAgent platforma
Kind of contact (inquiry, existing customer, wrong number, sales pitch)The map of Chapter 8, applied by the agent and checked in readingN and m, cleaned of contacts that were never customers
Transcript and summary, with a reviewed flag and the kind of any error foundAgent platform; the reader's notesE, as counts by kind of error
Booking: type, time, and whether the agent or a person made itCalendarc, and dc where the same field is kept for contacts answered by people
Customer number, matched on phone number or emailCustomer recordThe join from a contact to everything after it; B
First-year revenue and gross marginInvoices and the firm's booksV
Later visits, repeat purchases, cancellationsCustomer record and billingdr, and the data from which L is forecast
Staff minutes spent on the contactA sample of staff time, or the phone logH; w is the owner's judgment and is in no log
Connected minutes and chargeCarrier and platform invoicesk and n; F comes from the invoices and the owner's time
Arm: whether the agent was on or off for this contact by designThe schedule of the test in the next sectionThe counterfactual for capture and conversion

Table 1. Fields of the measurement record, where each comes from, and the term each lets a firm observe. The list is ours.

A firm does not need all fourteen fields on the first day. The smallest useful record has the contact and its door, the source, who answered, whether it booked and, later, whether the customer paid. With those, N, the booked share of contacts and a rough V can be read. The other fields add the terms that separate one mechanism from another. The join is the fragile field. A booking in the calendar with no source and a payment in the books with no contact are two halves of a loop that nobody has closed. We think a phone number used as the key, in a spreadsheet if nothing better exists, closes it for a small firm.

V is gross margin, and the books hold revenue. Margin needs the cost of delivering the work, which a firm may know by job type and rarely by customer, so the practical course is a stated margin per job type, labeled as an assumption. The source field also needs care. We think a single source field credits one source and behaves like a last-touch rule: a customer who read a review, asked an assistant and then tapped a map listing has three touches and one field. Recording both the tag and the customer's own answer keeps the disagreement visible. A distinct tracked number at each door is what the phrase "call tracking" names, and it draws about 1,600 searches a month in the US5.

Chapter 6 argued that each firm has to measure its own mix of routes, and this field is where it does. A visit or a call that comes from an assistant's answer may arrive with no recorded origin. OpenAI's own business dashboard counts a mention when ChatGPT shows a result for a local business, even if the customer does not open the listing or visit the website129. It reports mentions and website clicks, not calls or bookings129. This paper does not measure how often that happens. The nearest published count is an AI referral share of about 1% of site traffic from Conductor† (SEO platform), which counts clicks only127. The safeguard is the question put to the customer.

The first four links are visible within days. First-year revenue takes a year to complete, though early invoices give a partial reading. Lifetime value takes longer, because Gupta, Lehmann and Stuart define the value of a customer as the expected sum of discounted future earnings83, which makes L a forecast until the customers have aged. The phrase "customer lifetime value" draws about 3,600 searches a month in the US5.

Incrementality testing and marketing mix modeling: four designs a small firm can run

Incrementality is the part of an outcome that would not have happened without the agent. In Google Ads keyword data for the United States, collected in November 2026, "marketing mix modeling" is searched about 2,900 times a month, "marketing attribution" 720 times and "incrementality testing" 480 times5. Those are phrases an owner or a vendor would type, and volume shows interest, not purchase. Marketing mix modeling, as we use the term, fits a statistical model to sales over time and to spending across channels. We think it needs many periods with enough variation to tell channels apart, which a firm that books a few dozen customers a month lacks; this paper does not test that.

The four designs below need no model. Each compares the agent's presence with its absence.

Before and after in the same season

A before-and-after comparison sets the weeks before launch against the same number of weeks after, using the same weeks of the previous year as a guide to the season. It costs nothing in lost contacts and uses the record the firm already has. It can show the direction and rough size of a large change, and for a firm with no other option it is the starting point.

Its threat is that anything else that changed at launch is credited to the agent. A new competitor, the weather, a campaign, a price change, a staff change or a run of reviews can coincide with the launch unless the firm makes sure it does not. The launch also changes the record, because tracking added at launch counts bookings that the earlier period missed. A launch timed to a bad stretch will be followed by a better one with no agent at all, since unusually poor months tend to be followed by ordinary ones.

Lewis, Rao and Reiley tested this comparison directly in a second experiment, with participants recruited through Amazon Mechanical Turk and shown a video ad. Looking at the treatment group before and after exposure, with no control group, would have overstated the ad's effect by 350%151. That was a small online setting in 2011, so the size belongs to it. The direction is the point: a count taken without a control credits the change with whatever else moved. The mitigations are modest. List every other change in the window, start the record before the launch, and read the result as a description.

Alternating weeks

In the alternating-weeks design, the firm turns the agent on for some weeks and off for others, with the order set before the test by a coin flip for each pair of weeks (random assignment). A regular on-and-off pattern is worse, because it can line up with a rhythm in the calendar such as a payday or a billing cycle. The firm then compares bookings, and the booked share of contacts, across the two kinds of week. Both fall in the same season, with the same news and the same advertising, so a difference is easier to attribute to the agent than in a before-and-after count.

Carry-over is one threat: a caller who reached voicemail in an off week and rang again in an on week is counted for the agent. The direction in which this blurs the comparison depends on how callers behave, which we cannot say in advance. A season holds only a limited number of pairs of weeks, so the noise of the next section remains. The design also has a price: the off weeks forgo the very captures the firm is trying to measure.

A holdout of hours or lines

A holdout keeps the agent on for most of the time and withholds it from a random slice. The slice may be a few randomly chosen after-hours blocks each week, or one of several lines, and it goes to the old routine of voicemail and callback. The held-out slice is the control group, and it runs at the same time as everything else, so season, advertising and news are the same for both. The sacrifice is smaller than in alternating weeks, because most contacts are still served. The design measures net capture directly: the firm compares the share of held-out contacts that booked after a callback with the share of agent-handled contacts in the same kind of window.

Size is the first threat. A small slice gives a noisy comparison, and a large slice gives a clearer one at a higher cost, so the firm trades sacrifice against clarity. The slice must be chosen by chance. If the owner holds out the line that carries the cheaper advertising, the comparison is unfair from the start. Callers who find voicemail may ring again at an hour when the agent is on, which contaminates the comparison. And hours differ in kind: a night-time call from someone with no heat is not like an afternoon call about a price, so the slice has to be drawn within a kind of window.

Matched call logs

Matched call logs suit a firm that cannot or will not withhold the agent. The firm compares contacts the agent handled with comparable contacts in its own records. Comparable means calls in the same hours of the week, through the same door, of the same kind and in the same season, taken by a person before launch or at the same time. The telephone system's log, if it keeps one, shows calls that rang out. The agent's transcripts show how many after-hours calls were real inquiries and how many were wrong numbers or sales pitches. The design can describe what the agent did with contacts of a given kind, and it can give a rough count of the contacts that were being missed before.

It cannot show cause, because matching uses only what is recorded. A caller who rings at 9:40 in the evening with no heat differs from one who rings at 10 in the morning to ask about a price, in ways the log does not hold. Callers who chose to ring while a person was available also differ from those who rang when none was. This is the observational problem again, which Lewis and Rao call "a crippling concern for widely employed observational methods"48. We rank matched logs below the randomized designs and use them to check those results.

Test contacts for the process

A fifth tool measures the process, not the outcome. The firm places its own test contacts through each door at randomly chosen hours and records how long the response took, what the agent said, and whether the hand-off rang the named person. The design has precedent: Oldroyd, McElheran and Elkington, as quoted from their Harvard Business Review article in a 2011 news report, audited 2,241 US companies, measuring how long each took to respond to a web-generated test lead58. Rhodes and co-authors ran a simulated-patient telephone study of primary care practices in 10 states60. Test contacts check the plumbing, meaning m, a and E, and say nothing about whether bookings were caused.

The four designs side by side

Table 2 sets the four designs against one another.

DesignWhat it can showWhat it cannot showMain threatCost to the firm
Before and after in the same seasonDirection and rough size of a large changeWhich of several simultaneous changes caused itOther changes at launch; better tracking after launch; the timing of the launchNone in lost contacts
Alternating weeks, order set by chanceThe agent's effect on contacts and bookings in the same seasonEffects that take months, such as retentionCarry-over of callers between weeks; few weeksThe captures lost in the off weeks
Holdout of hours or lines, chosen by chanceNet capture, including callers the firm would have recovered anywayA precise figure when the slice is smallCallers who switch hours or lines; lines of unequal kindA smaller loss than alternating weeks, in proportion to the slice
Matched call logsWhat the agent did with contacts of a given kind; a rough count of earlier missed contactsCause, because matching uses only what is recordedUnrecorded differences between callers and timesNone

Table 2. Four designs, what each can and cannot show, and the main threat to each. The assessment is ours.

These rules apply whichever design a firm chooses.

  1. Fix the outcome, the length of the test and the stop date in writing before it starts, and do not stop on a good week.
  2. Record the arm in the contact record, so that every contact carries the label of the condition it met.
  3. Keep urgent and safety contacts out of any arm that withholds the agent, and keep a person on the route for them, as the map of Chapter 8 specifies.
  4. Never withhold the disclosure that the caller is speaking with an AI assistant. A test of wording compares two disclosures and never a disclosure with none.

Small numbers: how long a practice must wait before a difference means anything

Chapter 1 took the average dental practice's 30 new patients a month and supposed that an agent adds two, a difference ordinary month-to-month variation could produce. This section measures how large ordinary variation is, using a cited base rate. Henry Schein One† (practice management) reports that its customer dental practices of one to seven locations average 30 new patients a month, against 81 for the top 10%13. Its page does not define a new patient or describe its sample.

Take 30 as the practice's rate, and assume that new patients arrive one at a time, independently of each other. That is an assumption. A count of events that arrive that way varies from month to month by about the square root of its average, which here is 5.5 (the square root of 30). Roughly two months in three would land within 5.5 of 30, between 24.5 and 35.5 (30 - 5.5 and 30 + 5.5). Real months also move with season, holidays and campaigns, which this arithmetic leaves out, so we think the real swing is larger.

The effects to be seen are small beside that swing. The model's range for the dental chain turns 10%, 25%, 50% or 75% of the 36 monthly hang-ups into conversations34. That adds 1.1, 2.7, 5.5 or 8.2 bookings a month (calc. from Patient Prism†'s figures)34. The range shows the one open link at four values and is not a forecast. The comparison below treats every booking as a new patient, the high end, because only new patients appear in a count of 30. At the low end the callers are existing patients and would not show in that count at all. A month at 32.7 against an average of 30 sits inside an ordinary month's swing.

Averaging helps, but it needs months. The average of n months varies by about the square root of 30 divided by n. The difference between a block of n months before the agent and a block of n months after varies by about the square root of 60 divided by n (calc.). Table 3 gives that swing by block length.

Months in each blockSwing of the difference between two blocks, in new patients a monthEffects in the model's range larger than the swing
34.55.5 and 8.2
63.25.5 and 8.2
122.22.7, 5.5 and 8.2
241.62.7, 5.5 and 8.2

Table 3. The swing of a before-and-after difference in monthly new patients when they arrive independently at an average of 30 a month (Henry Schein One†), set against the four effects of the dental chain, each booking read as a new patient (calc.).

An effect of 1.1 bookings a month is below the swing at every block length in the table. The effects the model asks for are of the same order as ordinary variation, so a before-and-after count of all new patients can need years. The swing falls only with the square root of the months, so quadrupling the wait halves it.

A firm can compute the gaps between comparable blocks of the length it plans to test and ask whether the effect it hopes to see is larger than those gaps. It can also move to a finer outcome that has more events, such as bookings from after-hours windows by the week instead of all new customers by the month. It can judge early on leading indicators and late on revenue, as the next section does. And it can accept that the honest answer after a short test is "not known yet", with a date on which to decide on cost alone.

The same arithmetic applies to a randomized holdout, and it shows why gross and net differ. The log credits the agent with every booking in its windows. The held-out windows show the booked share without the agent. Applying that share to the contacts in the agent's windows gives the bookings the firm would have had anyway, so the net addition is the log's count less that number, as in the earlier illustration.

The holdout's own count is small. A slice of one hour in ten holds a tenth of the contacts, so on the same chain it holds 3.6 of the 36 monthly hang-ups. At most, if held-out callers booked as often as the agent's did, a tenth of the 2.7 bookings is 0.27 a month, or about 3.3 in a year (calc.). A count of 3.3 swings by about its square root, 1.8, which is more than half of the count. The estimate of what the firm would have had anyway inherits that swing, and so does the net figure. The result of any one test should therefore be read as a range, and the slice should be sized with that swing in mind.

Leading and lagging indicators for capture, conversion, capacity and retention

A leading indicator moves soon after a change and gives an early sign of the outcome. A lagging indicator is the outcome itself and arrives late. Because the lagging indicators are slow and noisy, as the last section showed, a firm reads leading indicators first and checks them against the lagging ones over the months that follow. This is the order Chapter 12 uses in its monthly cadence. Table 4 lists both kinds for each mechanism and for the cost side.

MechanismLeading indicators (days to weeks)Lagging indicators (months)Terms
CaptureShare of contacts arriving with no person available that the agent answers; share of those that were real inquiries; time to first response; share held to the end of the conversationBookings from those contacts, net of the holdout; shows; first-year margin of those customersm, a, c, V
ConversionShare of answered contacts that reach a booking attempt; reasons for no booking, coded from transcripts; time from contact to booking; share of calls that end soon after the disclosure lineBooked share of contacts already being answered, agent against matched or held-out contacts; shows; marginc, dc, V
CapacityStaff minutes per contact; share of contacts that needed no staff time; share of hand-offs that arrive with a usable summaryHours actually redeployed to paid work, and changes in overtime or payrollH, w
RetentionReminder and confirmation completion; no-show rate; time to callback on complaints and warranty contactsRepeat visits, cancellations and lapses over the retention period; the forecast of lifetime valueB, dr, L
Cost and errorConnected minutes and charges against caps; wrong statements found in readingInvoice totals against contacts handled; incident counts by kindk, n, F, E

Table 4. Leading and lagging indicators for each mechanism, and the terms each informs. The assignment is ours.

A badly chosen leading indicator misleads, as two field experiments show. In the telecommunications experiment, introducing a voice-based AI system to replace the phone menu temporarily increased machine-service duration and customers' demand for human service, and persistently reduced customer complaints50. In the Alibaba preprint, which is not peer reviewed and concerns customer chats, an agentic AI system reduced average chat duration and had limited effects on retrial rates, but substantially lowered ratings for AI-eligible chats51. A firm that tracked only handling time would have scored the first as a setback and the second as a gain, the opposite of what complaints and ratings showed. Table 4 therefore keeps a customer-facing measure, such as shows and bookings, beside the process measures wherever a mechanism faces the customer.

Time to first response is what the phrase "speed to lead" names, searched about 880 times a month in the US5. For capture, the denominator is the trouble: the count of after-hours calls includes wrong numbers, sales pitches and repeat attempts, and the transcripts are the way to separate real inquiries from the rest. Vendors classify calls in their own ways. Patient Prism† puts 34 of every 100 dental patient calls as real booking opportunities and 35 as not bookable34. Invoca† (call tracking) counts 43% of the calls answered by a person in its healthcare data as leads64. The definitions differ, so neither is a benchmark for another firm; the firm's own transcripts are the comparator.

For conversion, the agent might lower c as well as raise it, since dc can be negative. A Gartner survey of 5,728 customers in December 2023 found that 64% would prefer that companies did not use AI in customer service67. A stated preference in a survey does not measure what callers do, which is why the booked share is the measure.

For capacity, the hours released can be counted, but their value depends on whether they are resold, as Chapter 1 argued, so the lagging indicator is a change in what the freed hours produced. Retention is the slowest signal, though it may carry weight: Gupta, Lehmann and Stuart fitted a model to five firms valued from public data and found that a 1% improvement in retention, margin or acquisition cost improves firm value by 5%, 1% and 0.1% respectively83. The result is model-based, covers five firms and does not estimate a service business. For a small firm dr is small and slow to see, so retention is measured last and with the least confidence, through the leading indicators in the table.

What published phone measures count

Published call data exist, and each figure counts something different. We do not average them or put them in one list, because "hung up", "went to voicemail", "abandoned" and "not answered by a person" are not the same event. Almost all come from vendors (marked †) that sell a fix to the practices they measure.

In dentistry, Patient Prism† reports that of every 100 patient calls to 8,280 dental locations in 2025, 31 hung up before reaching an agent34. PatientNews† (dental marketing) says its average client practice answers 75% of calls, against 94% for top practices155. A practice trainer reported in a 2015 trade column that 35% of test calls to dental practices went straight to voicemail during regular business hours156.

In health care more widely, Invoca† reports that 54% of inbound calls to healthcare organizations were answered by a person64. The figure ranged from 45% to 68% across six sub-industries64. Call Box† (call tracking) found that the rate of patient callers connected to a qualified person fell from 73% in 2019 to 63% in 202033. The data cover more than 1,700 practices, and 2020 was a pandemic year33. In the Veterans Health Administration, a government system and not a private practice, the mean abandonment rate was 12.0% in October 201465 and 8.3% by September 201665. That study found that speed of answer, not abandonment, was associated with patients' satisfaction65, which does not show that a missed call costs nothing. In mental health, Ohio State researchers calling listed providers in five states could not analyze more than 6 in 10 calls because they reached an automated system or an answering machine157. The numbers came from insurance directories that the release says were often wrong157.

In law, Clio†'s secret-shopper test, with one call to each of 500 firms, found that 40% of firms answered in 2024, down from 56% in 201963. Among small businesses of all kinds, 411 Locals†, a local-search firm, monitored the calls of 85 businesses for 30 days in 2016 and reported 37.8% answered, 37.8% forwarded to voicemail and 24.3% with no response158.

No study measures calls to accounting firms. We estimate that between about four and six calls in ten reach no person at a small accounting firm in a busy season (from Invoca†'s health care calls and Clio†'s law-firm test above; likely high for a firm with a full-time receptionist). The low end is the 46% of health care calls not answered by a person (100% less Invoca†'s 54%), and the high end the 60% of law firms that did not answer Clio†'s test call (100% less Clio†'s 40%). The two measures differ, one counting calls and the other firms, and neither is an accounting firm.

The share of calls that arrive after hours is a second gap. A measurement of it in any of the four professions is NF; the nearest evidence is the 35% of test calls that went to voicemail during regular business hours156, which covers the hours a practice is open, so we form no estimate of the after-hours share. These figures show that missed calls are common in several settings under several definitions; only a firm's own phone log gives its m.

Privacy limits on measurement in regulated settings

The record in Table 1 is data about people, and in a health, legal or tax setting some of it may be protected information. The companion paper, Technical marketing for professional practices, treats each profession's rules in its Chapter 10 on compliance architecture and its Chapter 11 on measurement and attribution, and this section does not repeat them.

First, each tool in the measurement chain is a vendor. In our check of 28 software tools that small practices commonly use, 4 offered no business associate agreement, and they include Google Analytics and Meta Pixel20. Second, HHS gives an example: a patient makes an appointment through a covered clinic's website, and the site uses third-party tracking technologies that send the appointment and the IP address to a vendor159. In that case the vendor is a business associate and a BAA is required159. A court vacated one element of HHS's bulletin and said the vacatur was not meant to limit other guidance in the document159,160. Third, recording calls has its own rules. Federal law lets a private person intercept a communication if the person is a party to it or one of the parties has given prior consent, unless the purpose is a criminal or tortious act161. Washington requires the consent of all participants in a private communication162. Fourth, whoever reads the transcripts sees whatever callers said, which in a regulated practice may include protected information.

The design response is to measure with fewer fields: the door, the source, the kind of contact, whether it booked and what the customer was later worth, with no health, legal or tax details in the measurement record. The transcripts stay in a system covered by whatever agreements the practice needs. This is a design choice and does not say that any set-up complies with any rule. The same care applies to test contacts, which carry no real patient or client information. The designs above need only counts and a label for the kind of contact, so a practice can run them without moving sensitive content into a spreadsheet.

What we have not measured

This paper reports no outcome data from agents that Technical Marketing has deployed, and none of the designs above has been run by us on a client in a way it reports. The figures in the worked examples come from cited sources and from a stated statistical assumption. Published figures for AI voice agents come from vendors or from one peer-reviewed study of a different task, and Chapter 8 gives them with their definitions. A reader who meets a benchmark for these quantities, in a vendor's material or elsewhere, should ask how the quantity was defined, what its denominator was and who counted it.

This paper states no threshold and no sample-size rule, and a firm that wants to pool results across many contacts or many firms should have a statistician plan the analysis before the data are collected. Chapter 14 proposes studies that would give small firms better guidance than a single firm's record can.

Chapter 8

System design: framing, plumbing, finish

AI implementation starts with an order: framing, plumbing, agent, then finish

A house is built in an order. The frame fixes where the rooms are, the plumbing runs through the frame, and the finish goes on last, fitted to the rooms that exist. A technical marketing system with an AI agent has the same order. Framing is the system map: who contacts the business, about what, through which door, and where each contact goes. Plumbing is the set of connections that carry a contact from the door to a person or a calendar and leave a record: telephony, forms, chat, the calendar, the customer record, tracking and the hand-offs. The agent sits inside that structure with defined facts, permissions and duties. Finish is the visible surface, the pages and their look, fitted to the business last.

Each layer constrains the next. The map decides which doors exist and where contacts go. The plumbing can carry only what the map names. The agent can act only through the connections the plumbing offers, and the finish can show only the doors that exist. Reverse the order and each layer's constraints arrive after its decisions are made. The chapter's central claim is that the hand-off, the moment a contact passes from the agent to a person, is where the system succeeds or fails, and so it is the center of the design.

Suppose a heating and cooling company in a mid-sized town. It has an owner who estimates new installations, an office manager who keeps the calendar and dispatches, and six technicians, one of whom is on call each night. It has a main phone line, a website with a contact form, a business profile that shows the phone number, and customers on maintenance plans. The firm is invented for illustration, not a client or a real business, and nothing said about it is a result. We carry it through the chapter because an architecture is easier to judge on a case.

Framing: the system map of who contacts a service business, about what, and through which door

The map is a table with one row for each kind of contact that reaches the firm. Table 1 draws the map for the heating and cooling company. The rows are the kinds of contact such a firm plausibly receives, and the cells describe a design, not an outcome.

Contact typeDoorWhat the agent doesHand-offWhat is recorded
New customer, no heat or no coolingMain line, profile call button, site chatAsks the address and what the system is doing, checks the safety line, offers the first real openingIf the call is on the urgent list, the on-call technician is rung; otherwise the office manager sees the bookingSource, address, symptom, booking, time and result of any ring
Gas smell or an alarm soundingMain line, site chatReads the safety line the owner approved, does not coach a fixThe on-call person is rung at once, before anything is bookedSource, time, line read, who was rung, whether they answered
Estimate for a new systemForm, chat, main lineCollects the home details, books the estimate visit, gives no priceThe owner receives a summary before the visitSource, home details, booking
Maintenance-plan member changing a visitText, main line, chatFinds the member in the record, offers other times, confirms by textNone, unless no slot fits, then the office managerMember, old and new time
Warranty or callback on past workMain line, formTakes the address and the problem, marks the contact as a callbackThe office manager, first on the morning listJob reference, problem, callback flag
Billing questionMain line, emailTakes a message and does not discuss amountsThe office manager calls backCustomer, invoice number as given, message
Supplier, inspector or builderMain lineTakes the company, the job and a numberThe owner or the office manager, by textCompany, job, number
Sales pitch or wrong numberMain lineEnds the call politelyNone; marked no actionCall logged, no action
Anyone who asks for a personAny doorOffers the named number, takes a message if no one answersThe named personRequest, and whether it was answered

Table 1. The system map for an invented heating and cooling company: contact type, door, what the agent does, hand-off and what is recorded.

First, the map's contact types come from what the firm receives, not from the software's menu. The firm builds the list from its phone log, its inbox and its form submissions. Where it keeps none, the first job is to start a log, because a count of contacts by type is the first measure of N. Second, every door is on the map, including third-party doors. A business profile with a call button, or a directory that prints the number, is a door, and a contact that arrives through it must be routable. Chapter 6 argued that such pages are part of the demand infrastructure; this is where they enter the plumbing.

Third, every row ends in a named place: a person, a calendar entry or a deliberate mark of no action. No row ends at the agent or in an inbox that nobody is assigned to read. Fourth, the firm writes the urgency column. It decides what goes on the urgent list, and for a safety call such as a gas smell the owner approves the wording that is read. Fifth, the map has a row for anything else, and that row routes to a person, so an unforeseen contact is not left to the agent's improvisation. The map is also a measurement design: contact type and door are the fields the record needs to connect contacts to outcomes.

Plumbing: telephony, forms, chat, calendar, customer record and tracking

Plumbing is the set of connections between the doors and the places where contacts end. Table 2 lists the parts and what breaks when one is missing or unconnected. The failures in the last column follow from the design; they are not statistics about how often they occur.

PartJobMust connect toWhat breaks if it is missing or unconnected
TelephonyReceives calls, applies the routing rule, rings out for transfersThe agent, the on-call and office phones, the recordCalls that no one routes; a transfer that rings a dead number
FormsCollect the fields the map needsThe record, the agent, the officeMessages that sit in an inbox; contacts that cannot be sorted by type
ChatAnswers on the site under the same rules as the voice agentThe record, a real route to a personA chat that apologizes and ends, with no one told
CalendarHolds the true state of openingsThe agent, the staff, the recordDouble bookings; promises the calendar does not support
Customer recordHolds one history for each customerEvery door, the agent, the calendarDuplicate customers; no source and no outcome to count
TrackingTags the source of each contactThe doors, the recordContacts of unknown origin; no basis for Chapter 7
Hand-off channelDelivers the summary and the transfer to a named personThe agent, the record, the peopleContacts passed on that nobody reads

Table 2. The parts of the plumbing for a service business with an AI agent.

Telephony comes first because we think the main line is the principal door for a business of this kind. The firm chooses a routing rule, such as the agent answering at night and when the line is busy, and the design needs a fallback so that a failure of the agent's line sends calls to a person or a voicemail, not to silence. Transfers need numbers that can be rung: the on-call phone and the office line. Whether calls are recorded is the firm's choice, and state law governs what callers must be told.

A form asks for the kind of job, the address or town, the urgency, a phone number and a good time, and writes to the record as well as to an email. A chat has a real route to a person, meaning a call button or a callback request, not an apology.

The calendar needs the most care, because it is where a promise becomes binding. One calendar should hold the truth: the one the staff use, with appointment types, durations and rules about where each type is offered. The agent should see only the open slots of the types it may book, and the booking should be checked again at confirmation, since a form and a call can offer the same slot within seconds. The customer record holds one history for each customer, matched on phone number and email. Every contact adds an event to it with the door, the type, the time, the source, a summary, the hand-off and, later, the outcome; Chapter 7 gives the data model. Tracking tags the source at each door, with a distinct number or code for every door that can be told apart and the page of origin on every form. A door without a tag still works, but its contacts arrive with no recorded origin, which can be recovered later only by asking each customer.

Every hop is a vendor

Each connection in Table 2 often runs through a vendor. The map of the plumbing should list every vendor in the path: the telephone carrier, the speech and model providers, the calendar, the record, the texting service and the analytics tools. For a heating and cooling company that list is a matter of cost and reliability. For a medical or dental practice it also carries legal weight. HHS describes a business associate as a person or organization that performs functions or activities on behalf of a covered entity that involve creating, receiving, maintaining or transmitting protected health information163. A subcontractor that does so on behalf of a business associate is itself a business associate163. Each hop in a practice's plumbing may therefore need its own agreement.

Our check of 28 software tools that small practices commonly use found that 9 will sign a business associate agreement without requiring a special tier20. Another 15 sign one only on certain plans and 4 offer none, according to the vendors' own pages20. The tools with none include two tracking tools, Google Analytics and Meta Pixel20. For design, that means tracking and recording choices are made against the vendor list before the pages are built. Whether a set-up meets the rules is for the practice's own counsel or compliance lead to decide. Chapter 9 and Chapter 12 return to the law and the agreements, and the companion paper sets out each profession's rules.

For a dental or medical practice, which hops touch patient information also decides whether a HIPAA-ready program is needed at all. Marketing alone carries none, so a plan can be bought without one. A program is required for tools that can tell a named person is a patient: the voice agent on patient calls, the site assistant, appointment forms, texts to patients, call tracking and recordings, and any connection to a scheduling system56. The program is the compliance layer for those tools. It is added to a plan and does not include the voice agent56. Seen as service, the layer lets the practice answer quickly and gives the front desk more hands, through the four add-ons that run on it: reminders and rebooking, recall and reactivation, waitlist backfill, and intake with insurance capture56.

The AI agent's place: what it is given, what it may do and what it must pass on

The agent is a component with a defined interface, not the center of the system. In the paper's definition it is a software system built on a language model, given a bounded job, access to facts and tools, and rules about what it may do and when it must hand over to a person. Table 3 states the interface for the heating and cooling company.

PartFor the heating and cooling company
Given: factsServices and equipment worked on, the towns served, hours, maintenance-plan terms, published prices such as a service-call fee
Given: rulesThe urgent list, the approved safety line, the words it may use, the questions to ask for each contact type
Given: toolsRead open slots of the allowed types, create a booking, read and write the customer record, send a text to listed numbers, ring a named number
May doSay at the start that it is an AI assistant, answer from the facts, ask the questions the map assigns, book a real opening, take a message, confirm by text
May not doDiagnose equipment, quote a price it was not given, promise an arrival time the calendar does not show, approve a warranty claim or a refund, tell a caller that a gas smell can wait
Must pass onAnything on the urgent list, any request for a person, any complaint, any question outside its facts, anything it cannot resolve within a set number of attempts

Table 3. The agent's interface in the heating and cooling example.

The first principle behind the table is that the tools, as well as the prompt, should enforce a limit. A rule written into instructions is a request to a model; a rule built into what the tools permit is a constraint. If the agent has no tool that issues a refund, no phrasing by a caller can make it issue one, and if the booking tool offers only open slots of allowed types, it cannot book a slot that does not exist. A tool limit bounds actions, not words. The agent could still say something about a refund or a price, so the statements it may not make are also written into its rules and checked in the summaries that a person reads.

The second is that the agent should be given less than the staff have. It reads open slots and creates bookings, but it does not edit other customers' records or send text to arbitrary numbers. The least access that does the job limits the harm of any single failure.

The row for what the agent may not do rests on a view about errors. In the identity, the expected monthly cost of error is E, the sum over kinds of error of probability times consequence. We think the errors with the largest consequences come from a few kinds of statement: a price, a promise of time, a safety instruction, an approval. The design therefore does not try to make the agent right about everything. It removes those statements from the agent's authority, keeps its answers to the facts it was given, and lets it say that it does not know. Chapter 9 treats E as a term in the identity; here E is a design variable, set when the interface is written.

Hand-off as the design center of AI customer service: summaries, transfers, the named person and speed to lead

Every contact that the agent may not or cannot finish ends in a hand-off, and how those contacts turn out depends on how well it works. One measure that could be chosen for an agent is containment, the share of contacts finished without a person. The design does not treat containment as the goal. A contact that should reach a person and reaches one quickly, with a clear summary, is a success. A contact the agent finishes that should have reached a person is a failure, and the kind that raises E.

No peer-reviewed study gives a containment rate for inbound calls. No randomized or field study has tested an AI agent answering inbound calls or scheduling in dental, medical, legal or accounting settings. The nearest evidence comes from other tasks. At Mount Sinai, an AI voice assistant making outbound pre-procedure calls at one cardiac catheterization lab completed 86.4% of its calls in Phase I17. In Phase I, 36.6% of calls were fully automated, and about a third needed a protocol-driven nurse callback17. The study had no randomized control group, and an AI hallucination ended a call early in 0.7% of Phase I calls, a count that omits wrong answers inside completed calls17. Notable† says its voice assistant handled 40% to 60% of inbound insurance inquiries at one hospital system, with a saving of $4 a call projected, not measured101.

We estimate that roughly a third to a half of inbound front-desk calls would finish without a person (from the Mount Sinai and Notable† figures; likely high for a practice whose calls are mostly scheduling changes and urgent matters). The Mount Sinai calls followed a script, and Notable†'s were insurance questions in a deployment chosen for that use. The design does not rely on it.

Relatient† (patient scheduling) says its Voice AI answered 62% of inbound calls, of which 72% were either autonomously handled or intelligently escalated39, a figure that counts calls passed to staff as handled. Fin† (customer-service agent) states an average resolution rate of 76% on text chat and email for software and e-commerce firms164. Neither defines its terms in full or reports an audit. The figures support a narrower point than any containment claim. Relatient†'s counts hand-offs in its numerator, Fin†'s is an average for text chat in other industries, and Mount Sinai's reports how often a person was still needed.

The evidence on customer responses to AI in service is mixed, and it bears on hand-off in particular. In a Gartner survey of 5,728 customers conducted in December 2023, 64% said they would prefer that companies did not use AI in their customer service67. In the same survey, 53% said they would consider switching to a competitor if they found out a company was going to use AI for it67. In the Gartner text we read, the sample is given but not the sampling frame, and the figures record stated preference, not behavior. The Consumer Financial Protection Bureau reported in 2023 that several customers had complained of getting stuck in "doom loops" with chatbots, often because the issue fell outside the chatbot's limited capabilities165. These are complaints filed with a consumer-finance regulator, not a measured rate and not small-business phone lines.

Two field studies pull in different directions. A natural field experiment at a large telecommunications company found that a voice-based AI system replacing the interactive voice response temporarily increased machine-service duration and customers' demand for human service50. It persistently reduced customer complaints50. A 2026 arXiv preprint, posted before peer review, concerns text chat on Alibaba's Taobao platform. It reports that an agentic AI system for AI-eligible chats reduced average chat duration and had limited effects on retrial rates, but substantially lowered ratings for those chats51.

Neither is a small service business. In the telecommunications study demand for a human rose when the system was introduced, so the route to a person has to be ready on the first day. In the preprint, ratings fell although chats got shorter, so handling time alone does not show that customers were served well. Entrepreneur (secondary) reported that Klarna's chief executive told Bloomberg in May 2025 that it was critical to be clear to customers that a human would always be available on request111. Chapter 10 reads the company's statements further.

The first part is the named person. Each hand-off class has one human assigned in advance, with a number the system can ring, a fallback if that person does not answer within a set time, and a rule for when no one answers at all. The second is the choice between a live transfer and a message with a text. A transfer suits a contact where someone is expected to answer, and it should never leave a caller on a line that rings out. A message with a text suits the rest. The third is the summary. It carries who the contact is and how to reach them, what they want, what the agent said and promised, the urgency, the source and the next action. It is short enough to read on a phone and ordered with the urgent contacts first, so the person does not have to ask the caller to start again.

Closing the loop is the fourth part. Every hand-off creates a task with an owner and a clock. A task nobody claims within its time escalates to the next name, and a task is closed with an outcome, so the record shows what each hand-off became. The fifth is speed. The time from contact to a human response is a design variable that the firm sets for each class, from minutes for the urgent list to the next morning for a billing question. "Speed to lead" is searched about 880 times a month in the US5, and Chapter 4 reviews the evidence on response time. Suppose the text for the furnace call reads: Alder Street, furnace will not start, no gas smell reported, booked tomorrow 8:00, on-call technician rung 9:47 and answered. It shows the fields a person needs in a few seconds. In our practice, the text sent for a regulated practice carries no diagnosis or treatment details.

Finish: the visible surface, fitted last

Finish is the part of the system that a customer sees: the layout, the type, the color, the photographs, the tone of the copy and the way pages are put together. It is where a customer decides whether to trust the firm enough to make contact, which is stage 2 of the funnel in Chapter 6. The claim here is about sequence: the finish is fitted last, to the business and to the map, because by then the doors, the questions and the records exist and the finish can be shaped around them.

The map hands the finish a short list of constraints. Each page type has one primary door, and the door can be reached in one tap on a phone from any page. Each question that a customer asks and the map has answered has a page, or a place on a page, where the answer is stated in text. Each form is the form the map specified, and each phone number is the one tagged for its door. The system, in turn, hands the finish its material: the facts, the photographs, the reviews and the service descriptions that the business actually has. A finish designed first has to guess at all of this; one designed last uses it. This paper does not measure how much a better look adds to contact rates, and the order argument does not need a figure.

Template-first web design inverts the order, and what that costs

Template-first design starts from a theme or template: a set of pages, sections, forms and widgets that someone designed in advance. The firm chooses the look it likes and fills the template with its content. This inverts the order of this chapter, because the look is chosen first and the framing and plumbing are accommodated to it afterward. The argument against it concerns what cannot be measured or routed when the visible surface decides the structure. It does not depend on templates being poorly built or on any statistic about templates or conversion.

A template carries default decisions that the firm has not noticed making. If the template's contact form asks only for a name, an email address and a message, it does not ask the kind of job or the urgency, so the system cannot sort contacts by type, and every message reaches the same inbox. A phone number placed as plain text in a design element gives no way to tell which page or source a call came from, unless someone plans for tagging it. The template's page set was drawn up in advance for a generic business, not from the questions this firm's customers ask, so there may be no page that answers whether the firm serves a given town. An embedded booking widget that keeps its own calendar may not write to the firm's record. Whether a form submission, a tap on the number and a booking are separate recorded events depends on how the site is instrumented when it is built.

The first cost is rework: changing a form field or moving a door after the look is approved means reopening the finish, so the plumbing is likely to be negotiated against the design instead of derived from the map. The second is a lost baseline. Measurement cannot begin before recording begins, and the months before the plumbing exists are months in which the firm cannot say where its customers came from. Chapter 3 argued that a firm that cannot say where its last ten customers came from cannot say what an AI agent added, and the same holds for any change to the system. The third is invisibility. A form that sends to an unread inbox looks identical to one that works, so a failure of framing does not show on the surface.

The fourth is that the finish can pass for the system. Brynjolfsson and Hitt found, from case studies and firm-level evidence, that organizational investments have a large influence on the value of IT investments14. They also found that the benefits of IT are often intangible and disproportionately difficult to measure14. We think a template is technology chosen before the organization of work around it, and a site built that way can look finished long before it works.

Retrofitting is the strongest objection. A template can host a custom form, a call-tracking number, a booking widget and an analytics tag, and a capable designer can accommodate all of it. The argument does not deny this. It says that when the look decides first, the plumbing is added as patches the look must absorb, and the omissions are the ones nobody sees. A template is a respectable choice if it is made after the map, as a finish. A second objection is that many firms already have a site and cannot start from the frame. The order can be recovered: draw the map from the existing site, list every door and where it leads, repair the leaks and add the tags, and redesign the surface only then. We make no claim about how many firms are in that position.

Standards a design should meet: NIST AI RMF, the Generative AI Profile, OWASP and ISO/IEC 42001

The NIST AI Risk Management Framework 1.0, dated January 2023, describes four functions, GOVERN, MAP, MEASURE and MANAGE, broken down into categories and subcategories166. It describes itself as voluntary, rights-preserving, non-sector-specific and use-case agnostic166, and treats governance as a cross-cutting function infused throughout the other three166. NIST's Generative AI Profile, NIST AI 600-1, is dated July 2024167. OWASP's Top 10 for LLM Applications 2025 lists risks to applications built on large language models168. ISO/IEC 42001:2023 is an international standard for AI management systems, not a law169. Table 4 states what each of the four says and how the design answers it. We claim neither compliance nor certification with any of them.

SourceWhat it saysHow the design answers it
NIST AI RMF 1.0, four functions166GOVERN, MAP, MEASURE and MANAGE, with governance cross-cutting166The system map records the agent's use and its limits (MAP), the record and the call review supply measurement (MEASURE), the hand-off and a switch-off rule manage the failures (MANAGE), and a named owner governs the rules
NIST AI 600-1, confabulation167Confident production of erroneous or false content by which users may be misledThe agent answers from supplied facts, says when it does not know, and has no authority over prices, dates or approvals
NIST AI 600-1, human-AI configuration167Arrangements between a human and an AI that can lead to anthropomorphizing, algorithmic aversion, automation bias or over-relianceThe agent says it is an AI assistant, a caller can ask for a person, and a human reads the summaries and spot-checks the calls
NIST AI 600-1, data privacy167Leakage and unauthorized use, disclosure or de-anonymization of biometric, health, location or other personal dataFew fields are collected, texts carry only what the recipient needs, and the vendor list is checked before anything is built
OWASP LLM01:2025, prompt injection170User prompts alter the model's behavior in unintended ways, even when imperceptible to humansEvery caller, chat and form text is untrusted, and limits are enforced in the tools
OWASP LLM06:2025, excessive agency171Damaging actions in response to unexpected, ambiguous or manipulated model outputThe agent is given the fewest tools and permissions that do the job
OWASP LLM10:2025, unbounded consumption110Excessive, uncontrolled inference leading to denial of service and economic lossesCall length, daily usage and form and chat rates are capped, with an alert
ISO/IEC 42001:2023169Requirements for an AI management system; a standard, not a lawThe roles, change record and review of Chapter 12 have the shape of such a system

Table 4. What four sources say and how the design answers each. The right-hand column is our reading.

Table 4 draws three risks from NIST's Generative AI Profile, the three that bear most directly on a front-desk agent; a fuller treatment would go through the whole profile. The human-AI configuration risk concerns the staff as much as the caller: a summary that is trusted without being read, or an agent assumed to be right because it sounds sure, is the automation bias and over-reliance the profile names.

Prompt injection is the first-listed risk in OWASP's 2025 edition168. The list position is not a measured frequency, and the page's ranking method is not in our evidence. From the definition in Table 4 we draw the practical consequence that a limit should live in the tools as well as in the prompt. For a firm that pays by usage, unbounded consumption has a financial meaning. OWASP says it can lead to economic losses110. The meter that bills the firm per minute or per message is also the thing an abuser can run up. Caps and alerts bound the term k times n in the identity, whose sensitivity to usage Chapter 5 treats.

Disclosure rules differ by state and by trigger. Utah's general disclosure duty for a supplier of a consumer transaction (Utah Code 13-77-103) applies when the individual asks or otherwise prompts the supplier about whether AI is in use172. The request must be clear and unambiguous172. A separate provision gives a safe harbor to generative AI that clearly and conspicuously discloses at the outset of the interaction and throughout it173. We do not set out the other states here. The design takes a single position: the agent says at the start, on every channel, that it is an AI assistant.

Our own practice, described as ours

Technical Marketing is the marketing practice of the Strata Modus team, and we sell the kind of system this chapter describes, including an AI voice agent priced at $299 a month with 250 call minutes included and $0.39 a minute after that18 and plans from $999 a month18. This paper reports no outcome data from agents we have deployed.

In our practice, callers and site visitors are always told that they are speaking with an AI assistant. The agent answers only from the client's website and the facts we collect with the client, and it says so when it does not know. It books on the client's real calendar, offers only real openings and confirms by text. The client names the person for urgent calls and writes the urgent list, and a call that matches the list rings that person. A caller who asks for a person is transferred to the number the client named, and the agent takes a message when no one answers. It does not quote a price it was not given, promise a date the calendar does not show, diagnose equipment, dispatch trucks or decide a warranty claim or a refund. After each call the client receives a text summary, so a wrong answer shows up there and the facts the agent uses can be corrected.

The set-up follows the order of this chapter. We write down the client's services, area and hours, and the questions customers ask most. We write the urgent list with the client, connect the calendar and the on-call phone, test the agent on the client's ordinary and awkward calls, and then read the first summaries together. For a regulated practice, our HIPAA-ready programs begin with a consultation and depend on signed agreements and on how each tool is set up. We do not call any tool, or ourselves, HIPAA compliant. Chapter 12 describes how the system is run after it is built, and Chapter 13 gives the sequence for a firm starting from a brochure site and an unanswered phone.

Chapter 9

Risk, law and the cost of error

Risk and the cost of error: the E term in AI profit

The profit identity of Chapter 4 ends with a subtraction, the expected monthly cost of error, written E. For each kind of error, such as a wrong statement of a price, multiply the probability that one handled contact produces it by the number of contacts n that the agent handles in a month and by what one occurrence costs. E is the sum of those products over all kinds of error. Chapter 2 noted that the studies reporting gains report no cost of errors, so a firm that leaves E at zero has computed only the part of profit that is easy to see.

Consequences appear in statutes, orders, decisions and enforcement actions, and this chapter collects them with their sources. The probabilities are harder to find. Studies now measure how often AI tools state something false, but they measure other tools on other tasks. A randomized or field study of an AI agent answering inbound calls in dental, medical, legal or accounting settings is NF; the nearest measurement is of scripted outbound calls at one hospital, where an AI hallucination ended 0.7% of calls in Phase I17, or 1 call in 143 (1 / 0.007). A bounded script at a small firm's front desk produces a false statement in ~1 call in 100 (from that study; likely low). The study counts only hallucinations that ended a call, not a wrong answer inside a call that finished, and its calls were outbound and scripted, not open questions from new callers.

This paper does not measure how often regulators or private parties enforce the rules below against firms of this size. The chapter therefore offers a framework and a table of consequences with their sources, not an estimate of E, and no ledger in the paper carries a dollar figure for E.

Six kinds of error organize the chapter. In the first, the agent states a fact the firm owns, such as a price, an opening hour or a policy, and states it wrongly. In the second, it goes beyond its brief and gives advice or makes a commitment that belongs to a person. In the third, it fails to tell someone that they are dealing with AI, or tells them too late. In the fourth, information goes where it should not, through the agent, its logs or its vendors. In the fifth, a call is placed or recorded without the consent the telephone statutes describe. In the sixth, a claim made about the AI, by its seller or by the firm that uses it, is false or unsupported. The classification is ours, and the law does not use it.

This chapter describes rules. It names each one and says what it says, and it does not say what a court or regulator would do in any case. The rules are the ones in our evidence: federal statutes and rules, a few state statutes, professional-conduct texts, and some enforcement actions and decisions. The absence of a rule from this chapter, including most states, is not evidence that none exists. Law also moves inside the span of our own sources, as the Rytr order, the Colorado statute and the California guidance below show. We read most primary texts on November 10, 2026.

The first body of law concerns the telephone. In Declaratory Ruling FCC 24-17, adopted on February 2, 2024 and released on February 8, 2024, the Federal Communications Commission confirmed that the restrictions of the Telephone Consumer Protection Act (TCPA) on an "artificial or prerecorded voice" encompass current AI technologies that generate human voices174. It said that calls using such technologies fall under the TCPA and its implementing rules and so require the prior express consent of the called party to initiate, absent an emergency purpose or exemption174. We saw no vacatur or withdrawal on the FCC's page on November 10, 2026, which describes the page and is not a legal check.

The ruling turns on the word "initiate". It concerns calls initiated using an AI-generated voice, which we read as outbound calls. This paper does not say that it extends to an AI agent that answers a call the customer placed. In a later proposal, the September 2024 notice of proposed rulemaking on AI-generated calls (FCC 24-84), the FCC said that the TCPA's requirements do not extend to technologies used to answer inbound calls175. The same notice proposed to define an AI-generated call as one made over an outbound telephone call175. That is the Commission's description of the statute in a proposed rule, not a holding or a binding rule, and we did not confirm whether any proposal was adopted. Whether any part of the statute reaches an inbound call is a legal question for the practice's counsel.

In the designs in this paper, an agent that returns a call after a form submission, or that places reminder or recall calls, is placing a call the business initiates. For those flows the FCC describes a consent rule.

The statute behind the ruling is 47 U.S.C. 227. Section 227(b)(1)(A) makes it unlawful to make a call to the categories of lines the paragraph lists using an automatic telephone dialing system or an artificial or prerecorded voice176. An emergency call and a call made with the prior express consent of the called party are excepted176. Section 227(b)(1)(B) separately covers calls to residential lines using such a voice to deliver a message without that consent176. It excepts a call for emergency purposes, a call that is solely a federal debt collection call, and a call exempted by the FCC176. Under 227(b)(3) a person may bring a state-court action to recover actual monetary loss or $500 in damages for each violation, whichever is greater176. A court that finds a willful or knowing violation may increase the award to not more than 3 times that amount176.

The measure is per violation, so this consequence is a per-contact cost and not a lump sum. For illustration, suppose a flow placed 200 AI-voice calls in a month to numbers for which consent was missing. At $500 each, the statutory alternative to actual loss comes to $100,000 (200 x $500), before any increase for a willful or knowing violation. The example shows how the measure scales with the number of calls; it does not say that any call is a violation.

Recording is the second telephone question, because logging, one of the controls in the last section, is a form of recording. Under 18 U.S.C. 2511(2)(d), a private person may intercept a wire, oral or electronic communication where the person is a party to it or one of the parties has given prior consent, unless the purpose is a criminal or tortious act161. That federal text needs the consent of one party, and the two state statutes that follow ask for more. Washington's RCW 9.73.030(1) makes it unlawful to intercept or record a private communication transmitted by telephone or other device without first obtaining the consent of all the participants162. California Penal Code 632(a) punishes recording a confidential communication, including one carried on by telephone, without the consent of all parties177.

Three more states appear in our evidence. Florida's section 934.03(2)(d) makes it lawful to intercept a wire, oral or electronic communication when all of the parties to it have given prior consent178. Maryland's Courts and Judicial Proceedings section 10-402(c)(3) makes it lawful for a person who is a party to a communication to intercept it where all of the parties have given prior consent, unless the interception is for a criminal or tortious purpose179. Massachusetts General Laws chapter 272 section 99 defines an interception as secretly hearing or secretly recording a wire or oral communication without the prior authority of all parties180. That statute turns on secrecy. We think a recording the caller has been told about is not secret, though whether the statute treats it so is for counsel.

These statutes were written for conversations between people. Whether a voice-agent platform that records and transcribes a call counts as a party to the call or as a third party is not a question our sources answer. Our practice of always telling callers and site visitors that they are talking to an AI assistant is a disclosure practice and not a statement about consent to record.

Disclosing AI to customers: Utah, California, Texas and Colorado

Four states in our evidence have rules on telling a person that they are dealing with AI. Table 1 sets them side by side as the texts read when we consulted them.

RuleWhom the duty addressesWhat triggers itDateClaim
Utah Code 13-77-103(1)A supplier using generative AI to interact with an individual in a consumer transactionThe individual asks or otherwise prompts, clearly and unambiguously, whether AI is in useNo date in the cited text172
Utah Code 13-77-103(2) and (3)An individual providing services in a regulated occupationProminent disclosure when the use is a high-risk interaction, with no request condition in the cited text; verbally at the start of a verbal interactionNo date in the cited text172
California Business and Professions Code 17941A person using a bot to interact with someone in California onlineIntent to mislead about artificial identity to incentivize a purchase or sale or influence a vote; no liability if disclosedOperative July 1, 2019177
California Health and Safety Code 1339.75A health facility, clinic, physician's office or group practice officeGenerative AI used for patient communications about clinical information; verbal disclaimer at start and end of audioOperative date not verified in our evidence177
Texas Business and Commerce Code 552.051 (H.B. 149)A governmental agency, in (b); the provider of a health care service in which AI is used, in (f)An AI system intended to interact with consumers; for health care, disclosure by the date the service is first providedEffective January 1, 2026181
Colorado S.B. 26-189 (part 17, article 1, title 6)A deployer using a covered automated decision-making technologyNotice before using it to materially influence a consequential decision in a covered domainTakes effect January 1, 2027, except for sections that take effect upon passage182

Table 1. Disclosure rules in the evidence, whom each addresses, what triggers each and the date the cited text prints, if any.

Utah's statute holds two duties, and a statement about it must say which is meant. The general duty is conditional. The supplier must disclose that the individual is interacting with generative AI and not a human if the individual asks or otherwise prompts the supplier about whether AI is being used172. The prompt must be a clear and unambiguous request172.

The second duty carries no ask-first condition in the cited text. An individual providing services in a regulated occupation must prominently disclose generative AI use in regulated services when the use is a high-risk interaction172. A high-risk interaction involves collecting sensitive personal information such as health, financial or biometric data183, or giving personalized advice on which a person could reasonably rely for significant decisions, including financial, legal and medical advice183. A regulated occupation is one regulated by the Department of Commerce that requires a license or state certification183. The cited definition does not list those professions, so we do not say whether any profession discussed here falls within it; that is for the practice's counsel.

Under Utah's safe harbor, a person is not subject to an enforcement action for violating 13-77-103 if its generative AI clearly and conspicuously discloses, at the outset and throughout the interaction, that it is generative AI, is not human, or is an AI assistant173. Separately, it is not a defense to the violation of a statute administered by the Division of Consumer Protection that generative AI made the violative statement or was used in furtherance of the violation184. The Artificial Intelligence Policy Act, Title 13, Chapter 72, has a repeal date of July 1, 2027185. The disclosure rules now sit in Chapter 77, for which we found no repeal date and make no claim185.

California's bot law makes it unlawful to use a bot to communicate or interact with another person in California online, with intent to mislead the person about its artificial identity in order to incentivize a purchase or sale or influence a vote177. A bot is an automated online account177. Our sources do not answer whether a website chat agent or a telephone voice agent falls within those words, and we do not say that either does. California's health provision is addressed to health facilities, clinics, physician's offices and group practice offices. As our claim states it, the provision lists those four and does not list dental offices177. Its duty attaches to AI-generated patient communications about patient clinical information, a term that excludes administrative matters such as scheduling and billing177. It does not apply to a communication read and reviewed by a human licensed or certified provider177. Whether either provision reaches a given practice is for the practice's counsel.

In Texas, H.B. 149's general consumer-facing disclosure, in section 552.051(b), is addressed to a governmental agency that makes available an AI system intended to interact with consumers181. The cited text does not make it a general duty on private businesses181. A further subsection addresses health care: the provider of a health care service in which an AI system is used must give the disclosure not later than the date the service is first provided, or as soon as reasonably possible in an emergency181. Whether a given practice is such a provider is for counsel. The Attorney General may not bring an action before the 60th day after giving written notice, or if the person cures and provides a written statement within that time181.

In 2026, Colorado S.B. 26-189 repealed and reenacted part 17 of article 1 of title 6 of the Colorado Revised Statutes under the heading Automated Decision-Making Technology in Consequential Decisions182. The earlier text of part 17 is therefore not the text to cite. The act takes effect January 1, 2027, except for sections that take effect on passage182. Its notice duty concerns decisions in covered domains, which are education, employment, the lease or purchase of residential real estate in Colorado, a financial or lending service, insurance, health-care services, and essential government services and public benefits182. Legal services are not on that list182. The duty is a notice that the deployer used or will use the technology in a consequential decision, not a general notice that a person is talking to an AI182. Whether an agent that screens contacts for a business in a covered domain influences a consequential decision is a question for counsel.

We think the texts share a theme, to disclose early and in clear terms, and their differences are a reason to adopt one disclosure practice for every contact, as our practice does, and not to decide mid-call which statute applies. That is a design decision. We do not claim that it satisfies any statute.

The FTC on AI claims, fake reviews and AI hype

On September 25, 2024 the Federal Trade Commission announced Operation AI Comply, five actions against operations that use AI hype or sell AI technology that can be used in deceptive and unfair ways186. We report each action as the FTC's own words describe it, an allegation, a proposed order or an approved order, and not as a finding after trial.

One action announced in September 2024 reached a final order. DoNotPay claimed to offer the world's first robot lawyer186. The FTC's complaint alleged that it did not test whether its chatbot's output was equal to the level of a human lawyer and did not hire or retain any attorneys186. On January 16, 2025 the Commission voted 5 to 0 to approve a final order requiring $193,000 in monetary relief and notice to subscribers187. The order bars advertising that the service performs like a real lawyer without sufficient evidence187. We cite it for the shape of the claim: a statement that an AI does a professional's work, which the FTC alleged was made without testing against the professional's standard.

Another action from September 2024 no longer stands. The 2024 proposed order would have barred Rytr from marketing any service dedicated to, or promoted as, generating consumer reviews or testimonials186. The order became final in 2024. On December 22, 2025 the FTC reopened it and set it aside, saying that the complaint failed to satisfy the legal requirements of the FTC Act and that the order unduly burdens AI innovation188. The Rytr order is not current. We include it because it shows how far an enforcement position in this field can move between September 2024 and December 2025, a reason to date every rule in this chapter.

The other actions in our evidence are allegations about sellers of AI-branded products and business opportunities. The FTC said its complaint against Ascend Ecom alleges consumers were defrauded of at least $25 million186. It said the FBA Machine scheme, which marketed AI-powered tools for online storefronts, cost consumers more than $15.9 million186. In an August 25, 2025 complaint it alleged that consumers, many of them small business owners, lose as much as $250,000 after relying on Air AI's false promises189. It also alleged that Air AI claimed its product can replace human customer service representatives189. These are allegations, and we did not check the outcome of the Air AI case. They concern the buyer's side of the table, which is why Chapter 13 gives questions to ask a vendor.

The FTC's Rule on the Use of Consumer Reviews and Testimonials, 16 CFR part 465, went into effect on October 21, 2024 and authorizes courts to impose civil penalties for knowing violations190. The Commission said the rule addresses reviews that misrepresent that they are by someone who does not exist, such as AI-generated fake reviews, and prohibits businesses from creating or selling them191. It also prohibits compensation or incentives conditioned on reviews expressing a particular sentiment191, and unfounded legal threats, intimidation or certain false public accusations used to prevent or remove a negative review191. The Commission said in December 2025 warning letters that violations can bring civil penalties of up to $53,088 per violation192, and in September 2026 that civil penalty amounts remain unchanged during 2026193. Requesting, drafting and answering reviews are tasks a marketing system can assign to an agent, and the rule is to be read against any such flow.

A business blog post of February 27, 2023, which we could read only in an archived copy because the live page returned an error, said that false or unsubstantiated claims about a product's efficacy are the agency's "bread and butter"194. It added that AI performance claims would be deceptive if they lack scientific support or apply only to certain types of users or under certain conditions194. We did not check whether the FTC withdrew the post and do not call it current guidance. This paper holds itself to the standard it states: as Chapter 1 said, it reports no outcome data from agents we have deployed and so makes no performance claims for them.

Health data and AI for dental and medical practices: HIPAA and business associates

The companion paper, Technical marketing for professional practices, treats each profession's rules in depth.

Under HIPAA, the main question when an agent or a marketing tool touches a practice's patients is whether the vendor is a business associate, and for a given vendor that question is for counsel. HHS describes one as a person or organization that performs functions or activities on behalf of a covered entity that involve creating, receiving, maintaining or transmitting protected health information (PHI)163. A covered entity may disclose PHI to a business associate if it obtains satisfactory assurance that the associate will appropriately safeguard it195, documented in a written contract that meets 45 CFR 164.504(e)195. HHS says no Security Rule standard requires a covered entity to certify its compliance196. It says it does not endorse private organizations' certifications, which do not absolve covered entities of their obligations196. That is one reason this paper describes no tool as certified, or as compliant with HIPAA. We say HIPAA-ready: compliance depends on a signed business associate agreement and on how the tool is set up. Chapter 12 treats vendors and agreements.

HHS reasons from function to association. In one example, a patient makes an appointment through a covered clinic's website, and the site uses third-party tracking technologies that send the appointment and the patient's IP address to a vendor159. HHS says the vendor is a business associate and a business associate agreement is required159. The example concerns tracking, not agents, and we cite it for HHS's reasoning only. HHS also says that banners asking users to accept or reject cookies do not constitute a valid HIPAA authorization159.

One element of its guidance was set aside. On June 20, 2024, in American Hospital Association v. Becerra, the Northern District of Texas declared unlawful and vacated a portion of the guidance159,160. The court described the portion, which it called the "Proscribed Combination", as circumstances where an online technology connects an individual's IP address with a visit to an unauthenticated public webpage addressing specific health conditions or healthcare providers160. The court said the vacatur was not meant to limit other guidance in the document160. It would be wrong to say the bulletin was struck down. The point matters to the measurement of Chapter 7, since tracking is how a firm attributes a contact to a source.

The FTC has acted on health data sent to advertisers. It announced a proposed order under which GoodRx would pay a $1.5 million civil penalty197, and one requiring BetterHelp to pay $7.8 million to consumers198. These are announcements of proposed orders, and the claims report no court finding. The Health Breach Notification Rule does not apply to HIPAA-covered entities, or to any other entity to the extent that it acts as a business associate of one199.

HHS has acted on replies to reviews. Its Office for Civil Rights announced that Elite Dental Associates agreed to pay $10,000 to settle potential violations200 on a complaint that it disclosed treatment and insurance details in reply to a Yelp review201. It imposed a $50,000 civil money penalty on a North Carolina dental practice that disclosed a patient's PHI in response to a negative review and did not respond to OCR's requests202. A review reply is a public channel, and an agent given the job of drafting or posting replies works in it.

A proposed rule bears on agents directly. In its notice of proposed rulemaking of January 6, 2025, HHS stated that electronic PHI in AI training data, prediction models and algorithm data maintained by a regulated entity for covered functions is protected by the HIPAA Rules203. It also stated that a risk analysis must consider the type and amount of electronic PHI an AI tool accesses, to whom the data is disclosed and to whom the output is provided203. It is a proposal and not a final rule: a Unified Agenda entry read on November 10, 2026 lists the proposed rule as of 01/06/2025 and the final action as 07/00/2027, which we read as July 2027204.

Every text in this section is a model rule, an opinion or guidance. The ABA texts are Model Rules and a formal opinion of the association205. This chapter does not say whether a given state has adopted a rule, or what weight its disciplinary body gives an opinion.

Three Model Rules set the frame. Rule 1.6(c) says that a lawyer shall make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client206. Rule 5.3 says that a partner or lawyer with comparable managerial authority shall make reasonable efforts to ensure that the firm has measures giving reasonable assurance that a nonlawyer employed, retained by or associated with the lawyer behaves compatibly with the lawyer's professional obligations207. Rule 1.18 treats a person who consults a lawyer about forming a client-lawyer relationship as a prospective client, and bars the lawyer from using or revealing what was learned even when no relationship ensues, subject to exceptions208. An intake agent meets the third rule first.

ABA Formal Opinion 512, dated July 29, 2024, applies those duties to generative AI. It says that submitting a tool's output without an appropriate degree of independent verification or review could violate the duty of competent representation under Rule 1.1205. It adds that lawyers have additional supervisory obligations when they rely on others outside the firm to employ such tools205. It says that because many of today's self-learning tools are designed so that their output could lead directly or indirectly to the disclosure of information relating to the representation of a client, a client's informed consent is required before such information is input205. It also says that merely adding general boilerplate provisions to engagement letters is not sufficient to obtain informed consent205.

Florida Ethics Opinion 24-1 is advisory and was approved by the Bar's Board of Governors on January 19, 2024209. It says that generative AI chatbots that communicate with clients or third parties must comply with restrictions on lawyer advertising and must include a disclaimer that the chatbot is an AI program and not a lawyer or employee of the firm210. It adds that intake chatbots present further risks, including that a prospective-client or even a lawyer-client relationship may be created without the lawyer's knowledge210. It recommends that a lawyer obtain the affected client's informed consent before using a third-party generative AI program if doing so would involve disclosing any confidential information210. We read the opinion from a copy hosted by a law school, and a reader should check the Bar's own version.

The State Bar of California's 2026 guidance, approved May 14, 2026, replaces the November 2023 version211. It says an attorney must not input confidential client information into a generative AI solution that may present material risks to confidentiality or security, absent informed client consent211. It also addresses agentic systems: attorneys must not deploy them in a manner that allows the system to make substantive legal determinations, communicate legal advice, prepare and file pleadings, or otherwise act in a representative capacity without meaningful attorney supervision and review211. The Committee on Professional Responsibility and Conduct issued the original under the title Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law.

The Professional Ethics Committee of the State Bar of Texas issued Opinion 705 in February 2025212. It says that where a lawyer intends to provide client confidential information to a generative AI program, Rule 1.05 on confidentiality will likely be implicated212. It also says that a lawyer should always verify the accuracy of any responses received from a generative AI tool212. The New York State Bar Association's Task Force on Artificial Intelligence issued a report in April 2024 that the House of Delegates approved on April 6, 2024213. It says that the duty of confidentiality under RPC 1.6 also extends to what client information a lawyer may share when using certain generative AI tools213. The report is a policy report, not an ethics opinion. All four state texts in this section are advisory or policy documents, not rules.

AI in accounting and tax: return information, safeguards and the diligence duty

Accounting and tax practice has a statute of its own on a client's tax return information. Under 26 U.S.C. 7216, a tax return preparer who knowingly or recklessly discloses or uses information furnished for a return, other than to prepare the return, is guilty of a misdemeanor214. The punishment is a fine of not more than $1,000 ($100,000 where section 6713(b) applies), imprisonment of up to 1 year, or both214. Section 6713 adds a civil penalty of $250 for each disclosure or use, capped at $10,000 for any calendar year215. The regulations say that, unless authorized, a preparer may not disclose or use tax return information before obtaining the taxpayer's written consent216. The definition of a preparer includes a person engaged in the business of providing auxiliary services in connection with return preparation, including one who develops software used to prepare or file a return217. Whether a vendor of an intake or marketing agent falls within it is a question for counsel.

IRS Publication 5708 says that tax and accounting professionals are considered financial institutions under the FTC Safeguards Rule regardless of size, and must have a written, accessible information security plan218. Under 16 CFR 314.4(j), effective May 13, 2024, a financial institution must notify the FTC no later than 30 days after discovery of a notification event involving the information of at least 500 consumers219,220. Circular 230 section 10.22(b) presumes due diligence when a practitioner relies on another person's work product and used reasonable care in engaging, supervising, training and evaluating that person221. It speaks of another person, and we do not say whether that reaches a software tool221.

The IRS Office of Professional Responsibility issued an alert of June 24, 2026, titled Introductory Guidelines for Responsible AI Use in Federal Tax Practice, which says that technology is a powerful tool and not a substitute for professional judgment222. We describe it as guidelines, not as a regulation. It says Circular 230 due diligence requires verifying the accuracy of facts, citations and calculations produced by AI and that practitioners cannot rely solely on AI222. It also says that outsourced or third-party AI tools should be vetted222.

The AICPA Code bars a member in public practice from disclosing confidential client information without the client's specific consent223. Its interpretation on third-party service providers says the member should first enter a contractual confidentiality agreement with the provider, with reasonable assurance that it has procedures to prevent unauthorized release, or obtain the client's specific consent223. Our claim for the interpretation does not mention AI.

The AICPA's own journal carries a staff-written article of August 5, 2026 that says the Code's core responsibilities of competence, due care, confidentiality and independence do not disappear because new technology is involved224. It says members should ask, before pasting client data into any AI system, whether it is confidential client information, whether consent is required and has been given, and whether the use could result in disclosure to third parties224. It is an explanatory article by ethics staff, not an interpretation of the Code. An interpretation on generative AI is NF in the AICPA's Code material; the nearest evidence is this article.

Accuracy: what is known about wrong answers in professional work

The other factor in E is probability, and what is known about it is narrow. NIST's Generative AI Profile defines confabulation as the production of confidently stated but erroneous or false content, known colloquially as hallucinations, by which users may be misled or deceived167. None of the studies below measures an agent that answers a small service business's inbound calls; the nearest measurement comes after the legal and tax tests.

The most direct measurements are in law, where a statement can be checked against a source. In a preprint (version 2), Dahl and co-authors report that legal hallucinations occurred between 58% of the time with ChatGPT 4 and 88% with Llama 2 when the models were asked specific, verifiable questions about random federal court cases26. The claim reports figures for those models and none for legal-specific products. Magesh and co-authors found that the research tools of LexisNexis and Thomson Reuters each hallucinate between 17% and 33% of the time, less than a general-purpose model but contrary to the providers' claims of being hallucination-free27. The same Stanford evaluation put the share of queries answered accurately at 65% for Lexis+ AI and 42% for Westlaw's AI-Assisted Research, on 202 queries that its authors call a small sample40.

A different test points the other way. Vals AI, an AI-evaluation company, reports an October 2025 benchmark of 200 legal research questions in which four products scored 74% to 78% on an average weighted score, against 69% for a lawyer baseline225. The two tests use different products, questions and scoring, so we do not average them or call either a current error rate.

In accounting, ChatGPT 3.5 could not pass any of the CPA, CMA, CIA or EA exams in one study, with an average score of 53.1%226. After moving to ChatGPT 4 and adding few-shot training and the ability to reason and act, such as using a calculator, it passed all of them with an average of 85.1%226. The scores moved with the model, which is a reason to treat every figure here as dated. A tax-question test shows the same movement. A 2025 paper in Artificial Intelligence and Law tested models on 390 IRS-endorsed Volunteer Income Tax Assistance certification questions, including perturbed versions227. GPT-4 and Claude 3.5 Haiku reached 83.33% accuracy on basic multiple-choice questions, against 54.17% for GPT-3.5227. Those are certification questions and not client returns.

Field evidence from customer-facing deployments measures other outcomes. In a natural field experiment at a large telecommunications company, Wang and co-authors found that replacing the IVR with a voice-based AI system temporarily increased machine-service duration and customers' demand for human service but persistently reduced customer complaints50. A 2026 arXiv preprint reports a randomized field experiment on Alibaba's Taobao platform51. Deploying an agentic AI system for AI-eligible chats reduced average chat duration but substantially lowered ratings for those chats51. The CFPB's 2023 report, which discusses users of a bank's chatbot165, said several customers had filed complaints about getting stuck in "doom loops" with chatbots, often because the issue fell outside the chatbot's limited capabilities165. Those are complaints and not a measured rate.

The closest measurement to a voice agent on a health care front desk is a prospective study in npj Digital Medicine at Mount Sinai, from January to July 2025, of 1,431 patients and 1,606 calls. An AI voice assistant making pre-procedure calls completed 86.4% of calls in Phase I17, and 36.6% of the calls in Phase I were fully automated and needed no human intervention17. An AI hallucination led to an incomplete call in 6 calls, or 0.7%, in Phase I, and the paper reports AI system errors in 48 calls, or 6.0%, mostly a temporary portal downtime17. The study covered outbound calls to scheduled patients at one cardiac catheterization lab and had no control group.

We expect, and the evidence does not test, that an agent confined to a short list of facts the firm supplies and a short list of actions faces a different problem from the open legal questions above, because each fact can be checked against the firm's own source and a question outside the list becomes a hand-off and not an answer. That expectation is the argument for scope limits in the last section.

Expected cost of error: a framework and a table of consequences

The consequences that sources price fall into four classes: amounts set by statute per violation, amounts a tribunal ordered in one dispute, settlements and proposed orders that regulators announced, and sanctions that are not money. Table 2 collects them, one row for each kind of error, with the caveat that belongs with it. Chapter 10 reads the last two rows at length.

Kind of errorSourceConsequence as writtenCaveatClaim
Call placed with an AI voice without consent47 U.S.C. 227(b)(3)Actual loss or $500 per violation, whichever is greater; up to 3 times that if willful or knowingState-court action; the FCC ruling concerns initiated calls, and its statement on inbound calls is in a proposed rule175,176
Call recorded without all-party consentCalifornia Penal Code 632(a)Fine not exceeding $2,500 per violation, or $10,000 after certain prior convictions; imprisonment or bothConfidential communications only; Washington's penalty is not in our evidence177
Disclosure duty not met, UtahUtah Code 13-77-105Administrative fine up to $2,500 per violation; up to $5,000 per violation of an orderSafe harbor in 13-77-104173,228
Violation of H.B. 149, TexasCivil penalties by category$10,000 to $12,000 per curable violation; $80,000 to $200,000 per uncurable; $2,000 to $40,000 a day if continuedAttorney General may not sue before the 60th day after written notice; disclosure duty addresses agencies and health care providers181
Health information disclosed, HIPAAHHS table of January 28, 2026 (2025 adjustment)Per violation by tier: $145 to $73,011; $1,461 to $73,011; $14,602 to $73,011; $73,011 to $2,190,294HHS's 2019 notice applies different annual limits of $25,000, $100,000, $250,000 and $1,500,000 before inflation; current figures not confirmed229,230
Patient information in a review replyHHS Office for Civil Rights$10,000 agreed to settle potential violations; $50,000 civil money penaltyThe $10,000 was a settlement; the $50,000 penalty came after the practice did not respond to OCR's data request200,202
Tax return information disclosed or used26 U.S.C. 6713 and 7216Civil: $250 per disclosure or use, capped at $10,000 a year; criminal: fine up to $1,000 ($100,000 where 6713(b) applies), up to 1 yearCriminal offense requires knowing or reckless conduct214,215
Fake or AI-generated reviews, review incentives16 CFR part 465Civil penalties up to $53,088 per violationCourts may impose them for knowing violations190,191,192
Unsupported claim that an AI does a professional's workFTC final order, DoNotPay$193,000 and notice to subscribersA settlement; the Rytr order was set aside187,188
Chatbot misstates a company policyMoffatt v. Air Canada, 2024 BCCRT 149$812.02 in total: $650.88 damages, $36.14 interest, $125 feesOne decision of the Civil Resolution Tribunal231
Unverified AI output in court filingsCourt sanction orders$5,000; $3,000 and $1,000 each; admission revoked; reprimand, disqualification, bar referralOne order carries no money amount in our evidence232,233,234

Table 2. Consequences of error with their sources. Amounts are as the sources state them and carry their dates.

The amounts run from the $812.02 ordered in Moffatt231 to the $2,190,294 top of the fourth HIPAA tier229, a ratio of about 2,700 ($2,190,294 / $812.02). The figures are not like for like. One is a tribunal award in a single dispute, the other a regulatory ceiling per violation for the fourth tier, and the table gives no sign of how often either occurs. We think the spread suggests that the kind of error matters more to E than small differences in its probability.

Statutory amounts per violation, such as $500 in the TCPA176 and up to $2,500 in California's recording statute177, scale with the number of contacts to which the error applies. In the identity they behave like a per-contact cost. Awards and sanctions in single disputes behave like incidents. A scope limit that removes a kind of contact removes the per-contact exposure, whereas a review step lowers the chance of an incident.

Several of the texts describe their own exits. Utah gives a safe harbor for clear and conspicuous disclosure173. California's bot rule says a person using a bot is not liable under that section if the person discloses that it is a bot177. Texas bars an Attorney General action before the 60th day after written notice, or if the person cures and provides a written statement within that time181. We do not say whether a given practice would qualify for any of them. In the terms of E, each exit lowers the consequence of a kind of error without lowering its probability, and the controls in the next section are tied to them for that reason.

The table leaves out consequences that none of its sources prices: the owner's time, legal fees and the customer's reaction. On the last, Gartner reported from its own survey of 5,728 customers in December 2023 that 64% would prefer companies not use AI in customer service67. It also reported that 53% would consider switching to a competitor on learning a company was going to use it67. These are stated preferences and not observed behavior, and the claims we cite give the sample size and month but not the sampling frame.

Insurance is the nearest thing to a price on the whole of E, and the evidence on it is thin and mixed. ALPS, a legal malpractice insurer, wrote on August 21, 2025 that lawyers' professional liability policies typically do not exclude coverage for negligence claims arising from the use of generative AI235. The same post says that if firm lawyers allowed an AI to make critical legal judgments without attorney oversight, that could be viewed as the unauthorized practice of law, which such policies typically exclude235. It is a risk manager's article and not policy wording, and "typically" is the author's generalization across the market.

On the general liability side, a trade-association summary of Verisk's announcement says that from January 2026 ISO endorsements let insurers exclude generative AI exposures from commercial general liability policies236. An insurer chooses whether to use the forms, and general liability is a different policy from professional liability. An insurer statement on AI errors in medical, dental or accounting professional liability, and a bar statement on malpractice coverage of AI errors, are NF; the nearest evidence is the ALPS post on legal professional liability and the Verisk summary on general liability above. A firm that wants to know what its own cover says must read the form and ask its broker. The ALPS point also supports the scope limits below, since the cover it describes depends on a person supervising what the AI decides.

Pricing errors needs two inputs, and neither has been measured for a small firm's front desk. One is a rate for each kind of error, which only the firm's own log can supply, since the studies above measured other tools. The other is what one occurrence costs after legal costs, fees and any insurance, for which Table 2 gives the amounts that sources print. An insurer's quotation would be a third route to a price. What can be done now is the reverse calculation: take the margin that errors would have to consume before a purchase stopped paying, and solve for the rate at which they would do it.

The dental call chain, built from published figures, gives a worked bound. Patient Prism† (call analysis) reports 11,552,668 patient calls in 2025 to 8,280 dental locations on its platform34, or 116 calls a month a location (11,552,668 / 8,280 / 12). Its funnel runs from 100 calls to 69 connected, because 31 hung up before reaching an agent34, and of every 100 calls 34 were real chances to book and 21 booked34. The 31 hang-ups are 36 calls a month. The chain assumes that hang-ups resemble callers who get through, which the source does not show. On that assumption, answering every one would book 11.0 a month (36 x 21 / 69). The share of those 36 that an agent would turn into conversations is the one link nobody has measured, so suppose it is 25% or 50%, a range and not a forecast.

Patient Prism† counts new and existing patients together, so a booking is valued at two ends. At the low end it is one visit, about $282, which is ADA gross billings of $965,660 per general dentist53 divided by 3,429.5 visits a year in a Statista series whose source is not named54, so it is rough. At the high end it is a new patient's first year, $88755. At 25% the 36 hang-ups give 9.0 calls and 2.7 bookings, worth $772 to $2,433 a month, and at 50% they give 18.0 calls and 5.5 bookings, worth $1,544 to $4,865 (calc.).

Our HIPAA-ready program is a compliance layer added to a plan, and it does not include the voice agent56. The agent costs $299 a month with 250 call minutes included, about 62 four-minute calls (250 / 4), and $0.39 a minute after that18. A dental or medical practice also needs the program around the agent on patient calls, $824 a month in the first year ($699 for Essentials and $125 a month of its $1,500 setup)56. Every tool that handles patient information runs inside it, so its fee is shared and is not charged to answered calls alone56. The two together come to $1,123 a month in the first year ($299 + $824), which is the narrow case of a practice that buys the program for the agent and runs nothing else on it.

Taking the agent's $299 from the values above leaves a range of $473 to $2,134 a month at 25% and $1,245 to $4,566 at 50% (each end less $299). The answered hang-ups are 9.0 and 18.0 calls, or 36 and 72 minutes at four minutes a call, inside the 250 included18, so the fee is $299 at both shares. Every end is positive, so at these two shares answered calls alone cover the agent's own fee at either value of a booking. In the narrow case, with the program's $824 taken out as well56, the same ranges are -$351 to $1,310 at 25% and $421 to $3,742 at 50%. A negative end there means that answered calls alone do not cover that purchase at that share, and errors then have no margin to consume. Billings and spending are not profit, so the positive ends are ceilings on the margin that errors could consume.

Now set the expected cost of error equal to the agent's margin. If the agent took all 116 calls and each kind of error cost $812.02, the Moffatt total231, the break-even rate is the margin divided by 94,194 (116 x 812.02). At the 50% row that is 1,245 divided by 94,194, about 0.013 or 1 in 76, at the low end, and 4,566 divided by 94,194, about 0.048 or 1 in 21, at the high end. The high end of the 25% row gives 1 in 44 (94,194 / 2,134). If one error cost $10,000, the amount HHS's Office for Civil Rights announced for Elite Dental Associates200, the divisor is 1,160,000. The 50% row then gives about 1 in 932 at the low end and 1 in 254 at the high end, and the high end of the 25% row gives 1 in 544 (each margin / 1,160,000). If the agent took only the 36 calls that hang up, each rate rises 3.2 times (116 / 36).

The divisor counts all 116 calls as handled by the agent, while the margins above are net of $299, the fee for the answered hang-ups alone. All 116 calls at four minutes are 464 minutes, 214 past the 250 included, which adds $83.46 at $0.39 a minute (214 x $0.39)18 and makes the fee $382.46. Taking that from the 50% low end leaves $1,161.54, or 1 in 81 at $812.02 (94,194 / 1,161.54) and 1 in 999 at $10,000 (1,160,000 / 1,161.54).

The nearest measured rate is Mount Sinai's: hallucinations ended 0.7% of Phase I calls there17, and the ~1 call in 100 for a bounded inbound script is of that order. Only some hallucinations would carry a cost like Moffatt's or a settlement like Elite Dental's. In the 50% row any share of hallucinations could cost $812.02 each, at the low end and at the high end (0.013 and 0.048 both exceed 0.007), while at most 15% could cost $10,000 each at the low end (0.00107 divided by 0.007) and at most 56% at the high end (0.00394 divided by 0.007). At the $382.46 fee both shares fall by about a point, to 14% and 55% (0.00100 and 0.00386, each divided by 0.007). Because the estimate likely errs low, read these shares as ceilings that are too generous, and read the low end as the one that matters for a practice whose callers are mostly existing patients.

The firm need not know the probability to use these figures. It asks whether it can believe that the error happens less often than the rate, and if it cannot, the kind of error needs a control before the agent takes that kind of contact. An expected value is also an average, and a small firm may care about one large loss more than about the average, which is a further reason to bound a consequence and not to rely on an estimate.

Controls that lower the cost of error

Five controls act on E: scope limits, disclosure, hand-off, logging and review. Table 3 sets each against the part of E it acts on and the standard behind it, and the mapping is ours.

A scope limit says what the agent may state and do, and everything else becomes a hand-off. In practice that is a short list of facts the firm supplies and keeps current, such as services, hours and policies, a short list of actions, such as answering from the list, booking and taking a message, and a list of subjects that go to a person. We think that last list includes anything that is advice. OWASP defines Excessive Agency as the vulnerability that enables damaging actions in response to unexpected, ambiguous or manipulated outputs from an LLM, whatever causes the malfunction171. A scope limit also fixes which rules are engaged, because several attach to content and not to technology. California's health rule turns on clinical information and excludes scheduling and billing177, and Utah's high-risk definition turns on sensitive information and personalized advice183. OWASP also names Unbounded Consumption, excessive and uncontrolled inferences that lead to economic losses among other risks110, which bears on the usage-priced cost lines of Chapter 5.

Disclosure lowers the consequence of a failure to disclose by removing the failure, and several texts name it as the exit. Utah's safe harbor and California's no-liability provision both turn on disclosure173,177. The health-care texts add timing: Texas requires disclosure by the date the service is first provided181, and California's health rule calls for a verbal disclaimer at the start and the end of audio communications177. The AMA's policy, which is policy and not law, says AI use should be clearly disclosed at the beginning of an encounter where patients engage directly with AI237.

Disclosure may have a cost that we cannot size. If the stated preferences in the Gartner survey carried into behavior, some contacts might decline to continue67. We estimate that the share who decline is below the 53% who told Gartner they would consider switching67, because considering is not doing, so that figure is a ceiling and not a forecast. An independent survey of patients or clients about AI receptionists or voice agents in any of the four professions is NF; the Gartner survey is the nearest evidence. Chapter 7's designs can measure it for a firm that records outcomes.

The record on hand-off is thin and mixed: the CFPB's report collected complaints of being stuck in loops with chatbots165, and the telecommunications experiment found that replacing the IVR raised demand for human service temporarily while reducing complaints persistently50. California's health rule pairs its disclaimer with clear instructions for reaching a human177. A hand-off needs a named person who receives it, a time within which that person acts and a summary of what the agent heard. Chapter 8 treats hand-off as the design center.

Logging gives the firm a record of what the agent said and did, the raw material for the measurement in Chapter 7 and for finding errors before a customer or a regulator does. The log is also data. A recording raises the questions in the recording statutes above162,177. Its contents may be health, client or tax return information subject to the sector rules, and leakage and unauthorized use of sensitive data is a risk that NIST's profile lists under Data Privacy167. The control therefore comes with decisions about what is recorded, who may read it and how long it is kept. This paper gives no retention period.

Review is a person reading what the agent did and correcting it, and texts in several fields refer to it. California's health rule excepts communications that a licensed or certified provider has read and reviewed177. ABA Formal Opinion 512 says that submitting output without appropriate independent verification could violate the duty of competence205, and the IRS alert says practitioners cannot rely solely on AI222. Review has its own failure mode: NIST's profile lists automation bias and over-reliance among its human-AI configuration risks167, and a reviewer who approves what the system produced without checking supplies no control. At the volumes of a small firm, a person can read every transcript for a period, and whether the volume stays small enough is a question the firm's own counts answer. Chapter 12 puts reading the calls in the monthly cadence.

ControlPart of E it acts onStandard
Scope limitsThe number of contacts to which a kind of error can applyOWASP Excessive Agency171 and Unbounded Consumption110
DisclosureThe consequence of a failure to discloseGOVERN, the AI RMF's cross-cutting function, where we think a standing disclosure policy sits166
Hand-offThe chance an out-of-scope exchange ends in a wrong answer, and the harmNIST profile, human-AI configuration167
LoggingDetection, correction and measurement, at the price of creating dataMEASURE in the AI RMF166; NIST profile, data privacy167
ReviewThe chance an error reaches a person uncorrectedMANAGE in the AI RMF166; NIST profile, confabulation167

Table 3. The five controls, the part of E each acts on, and the standard behind each.

NIST's AI Risk Management Framework 1.0, of January 2023, describes four functions, GOVERN, MAP, MEASURE and MANAGE, broken down into categories and subcategories166. It says governance is infused throughout the other three166. NIST describes the framework as voluntary and flexible for organizations of all sizes166. Its Generative AI Profile, NIST AI 600-1 of July 2024, applies it to generative AI167. Assigning a control to a function, as Table 3 does, is our reading of the function's name. The OWASP Top 10 for LLM Applications 2025 lists Prompt Injection first as LLM01:2025168 and Unbounded Consumption as LLM10:2025110, and we treat the order as list position and not as measured frequency. ISO describes ISO/IEC 42001:2023 as an international standard that specifies requirements for an Artificial Intelligence Management System within organizations169. It is a standard and not a statute, and we read ISO's page in an archived capture.

These are frameworks and standards, not rules. The cited sources do not tie any statute in this chapter to them.

Chapter 10

What documented failures and reversals teach

AI failures in the public record: a selected sample

This chapter reads the public record of AI deployments that went wrong or were walked back, for what it shows about the paper's complements. Every case in our evidence became public: a company announced something, a tribunal decided something, a court sanctioned someone, or a journalist reported something. The set is a selected sample of the failures that surfaced, not a sample of deployments, and it supports no estimate of how often deployments fail.

The sources differ in kind, and each case is told from its source and no further. Three rest on primary documents: Klarna's February 2024 announcement; the decision in Moffatt v. Air Canada, 2024 BCCRT 149; and the sanction orders in Mata v. Avianca, Wadsworth v. Walmart and Johnson v. Dunn. Five rest on news or trade press, and every sentence about those five names its source and adds nothing to what the source reports.

Chapter 1 defined three complements: demand, a flow of contacts for the agent to act on; process, a defined path for each contact, including what the agent knows, what it may do and what it must pass on; and measurement, a record linking each contact to its source, its handling and its outcome. Two parts of process separate in these cases. Scope is what the agent is allowed to say and do; supervision is a person checking what it did. The classification is this chapter's reading: none of the companies or courts described their cases this way, and a different reader could classify some of them differently.

Selection runs in one direction. A failure becomes public when someone is hurt, someone sues, someone regulates or someone writes about it, and those events follow visible wrongs: a false statement, harmful advice, a fabricated citation. Failures of demand and of measurement are quiet. An agent installed where nobody calls produces no headline, and neither does a deployment whose owner could not tell what it earned. The report that 95% of organizations are getting zero return3 describes outcomes in which quiet failures may sit, with the limits Chapter 1 set out. The public cases here therefore bear mostly on scope and supervision, and only indirectly on the complements that carry the paper's thesis.

Klarna and AI customer service: an announcement, then a change of course

Klarna's announcement of February 27, 2024 is the fullest company-reported account of an AI customer service deployment in our evidence. Klarna said its OpenAI-powered assistant, live globally for one month, had held 2.3 million conversations, two-thirds of its customer service chats38. It said the assistant was doing the equivalent work of 700 full-time agents38. Customers, it said, now resolved their errands in less than 2 minutes compared with 11 minutes previously38. It also said the assistant was on par with human agents on customer satisfaction score and more accurate in errand resolution, leading to a 25% drop in repeat inquiries38. It estimated that the assistant would drive a $40 million profit improvement in 202438.

These are Klarna's statements about itself, with no independent check. The cited release does not state a method for what "equivalent" means, give a sample, period or definition for "resolve", or give a baseline or method for the repeat-inquiry figure38. The profit figure is an estimate, made in a release that describes the assistant as live for one month38.

For the complements, demand existed before the assistant did: two-thirds of a month's chats is a flow of millions of contacts38. The process was a customer service operation that already had measures, since a comparison with the 11 minutes "previously" implies a baseline. The reported figures cover volume, speed, repeat contact and satisfaction parity. The release says nothing on revenue per customer or retention, and we do not know what Klarna measured internally.

Entrepreneur reported that Klarna's chief executive told Bloomberg in May 2025 that Klarna was hiring human customer service workers again, and that he said it is critical to be clear to customers that there will always be a human if they want one111. We did not open Bloomberg, and the evidence for this paper holds no primary Klarna statement revising its 2024 claims. What we cite is one press account of one interview.

The record does not show that Klarna's assistant lost money, that the 2024 figures were wrong, or how many contacts the assistant handles now. A company can use an assistant for most contacts and still hire people to take the rest, so "Reversal" is the wrong word for what the sources show, and we use "change of course" for what a news report describes.

We think the reported change concerns scope and hand-off, not demand, since the chief executive's remark is about the route to a person, the control Chapter 9 calls hand-off. We do not know why Klarna changed its staffing, and his words as reported name a brand perspective and a company perspective, not a failed measure.

Moffatt v. Air Canada: a company is held to what its chatbot said

In Moffatt v. Air Canada, 2024 BCCRT 149, issued on February 14, 2024, the Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation231. It reached that finding because Air Canada's website chatbot told a customer he could apply for bereavement fares retroactively231. The customer later learned from Air Canada employees that Air Canada did not permit retroactive applications231. The decision is a tribunal's, not a court's, on one dispute between one customer and one airline, and we draw no rule from it for any other jurisdiction.

Three findings carry the case. Air Canada suggested, in effect, that the chatbot was a separate legal entity responsible for its own actions, which the tribunal called a remarkable submission. The tribunal wrote that Air Canada is responsible for all the information on its website231. It added that it makes no difference whether the information comes from a static page or a chatbot231. It found that Air Canada did not take reasonable care to ensure its chatbot was accurate, and that it was reasonable for the customer to rely on the chatbot231.

In money terms the case is small. The tribunal found the customer entitled to damages of $650.88, the difference between the $1,630.36 he paid and the $979.48 he should have paid231. It ordered Air Canada to pay a total of $812.02: $650.88 in damages, $36.14 in pre-judgment interest and $125 in tribunal fees231. We infer the amounts are Canadian dollars, since the decision text we read does not label the currency231. The evidence says nothing about the airline's own legal costs or what the matter cost it otherwise.

What the decision shows, whatever the sum, is where responsibility sat in that dispute: with the business whose website carried the chatbot. The cited decision does not say why the chatbot gave the answer it did, what material it drew on, or whether Air Canada read its transcripts. We classify the failure as one of scope and supervision: an agent spoke about a policy, the tribunal found no reasonable care to ensure accuracy, and the cited decision shows no check that its statement matched the policy.

A statute in a different jurisdiction points the same way. Utah Code 13-77-102 concerns violations of any statute administered and enforced by the Division of Consumer Protection184. It says it is not a defense to such a violation that generative AI made the violative statement, undertook the violative act or was used in furtherance of the violation184. The two come from different law, and one does not show what the other would do.

Four more cases from the press: NEDA, McDonald's, MyCity and Commonwealth Bank

CBS News reported in May 2023 that the National Eating Disorders Association said, after critics' complaints, that the current version of its Tessa chatbot, running the Body Positive program, may have given information that was harmful and unrelated to the program238. It also reported that NEDA took the program down pending an investigation238. NEDA's statement, as CBS News reported it, says the chatbot "may have given" such information, which is a possibility NEDA stated and not a determination of what the chatbot said238. The report gives the sequence as critics' complaints first, then the statement and the shutdown. The case is a scope failure in a high-stakes domain, since a chatbot running one program may have produced information unrelated to it. The report does not say what NEDA monitored, so we say nothing about its supervision beyond that sequence.

Nation's Restaurant News reported on June 17, 2024 that McDonald's confirmed it would end its Automated Order Taker test with IBM at about 100 locations after two and a half years239. The outlet quoted a company statement that, after review, McDonald's had decided to end the partnership on that technology beyond that year239. The same outlet reported that a 2022 report by BTIG showed voice ordering accuracy in the low 80% range, against McDonald's goal of 95%-plus order accuracy before a broader rollout239. We did not open the BTIG report, and its method is unknown. The company's statement, as reported, gives no reason, so we do not say the test ended because of accuracy. The reports show a deployment with a stated threshold before wider rollout, which is what a measured process supplies, and a decision not to go further. The case is as much an example of a threshold doing its job as of a technology failing, and the sources do not let us choose between the readings.

THE CITY and The Markup reported on January 30, 2026 that New York City's MyCity AI chatbot, released in fall 2023, gave answers that, if followed, would lead businesses to illegal behavior, such as taking a cut of employees' tips240. They also reported that a spokesperson for the new mayor confirmed the administration planned to take the chatbot down240. The report gives no reason for the removal beyond that confirmation. The case is a scope failure, since the reported answers concerned what businesses may lawfully do, a subject where wrong answers are costly, and the report describes no limit on what the chatbot would address. On supervision, the account has the errors reaching the public through reporting and does not say what checks the city ran.

Information Age reported in August 2025 that Commonwealth Bank reversed its decision to cut 45 customer service jobs due to the use of AI and admitted that it made an error by declaring the roles redundant241. The Finance Sector Union said the bank had announced in July that the 45 roles would be replaced by an AI-powered voice bot241.

This case concerns the capacity mechanism. The identity values released capacity as H times w, and H is released only if the work the agent absorbs goes away. We infer that a bank that admitted an error in declaring 45 roles redundant had counted on releasing the capacity of those roles; the report does not say why the roles were declared redundant.

Courts supply the best-documented family of AI failures in our evidence, because a sanction order is a public document that names the court, the date and the penalty. In Mata v. Avianca, on June 22, 2023, Judge Castel of the U.S. District Court for the Southern District of New York imposed a $5,000 penalty, jointly and severally, on two lawyers and their firm232. They had filed non-existent cases generated by ChatGPT232. The court said there is nothing inherently improper about using a reliable AI tool for assistance, but that existing rules give attorneys a gatekeeping role in ensuring the accuracy of their filings232. It found that the respondents abandoned their responsibilities232. They had submitted non-existent opinions with fake quotes and citations and then stood by them after judicial orders called their existence into question232.

The order of the U.S. District Court for the District of Wyoming in Wadsworth v. Walmart, dated February 24, 2025, recounts that plaintiffs' attorneys filed motions in limine citing nine cases, of which eight did not exist233. The court revoked one attorney's pro hac vice admission and fined him $3,000, and it fined two other attorneys $1,000 each233.

In Johnson v. Dunn, on July 23, 2025, the U.S. District Court for the Northern District of Alabama publicly reprimanded three attorneys for ChatGPT-fabricated citations and disqualified them from further participation in the case234. It also ordered publication of the order and referred the matter to the Alabama State Bar234. The attorneys confirmed that the citations in two motions were hallucinations of ChatGPT, and the court called fabricating legal authority serious misconduct that demands a serious sanction234. The cited claims name no monetary amount for this order.

Damien Charlotin's AI Hallucination Cases database tracks legal decisions in which a court found or implied that a party relied on AI-hallucinated content242. It listed 2,289 cases when read on November 10, 2026 (secondary)242. The page calls itself "a work in progress" that will expand as new examples emerge. A running count of decisions has no denominator of filings, so it shows scale and cannot say how often AI-assisted filings contain fabrications.

Courts have also begun to require disclosure. The Northern District of Texas's Local Civil Rule 7.2(f) requires a brief prepared using generative AI to disclose that fact on its first page under a stated heading243. If the presiding judge directs, the brief must also disclose the specific parts so prepared243. The rule was adopted by orders of June 10, 2024 and took effect September 3, 2024244. Judge Starr's page says he no longer requires a certification on the use of artificial intelligence in light of that adoption245. It is the rule of one district, and the cited claims do not show a nationwide requirement.

In the paper's terms these are supervision failures in professional work: citations that a tool generated were filed in a court before anyone caught that the cases did not exist. Test results show why a check is needed. A preprint by Dahl and co-authors reports legal hallucinations between 58% of the time with ChatGPT 4 and 88% with Llama 2 on specific, verifiable questions about random federal court cases26. Magesh and co-authors report between 17% and 33% for AI legal research tools from LexisNexis and Thomson Reuters27. Those are test results, not rates in practice.

The orders place responsibility on the lawyers who filed. ABA Formal Opinion 512 says that submitting output without an appropriate degree of independent verification or review could violate the duty of competent representation205. The monetary amounts in these orders are small. In Johnson v. Dunn the court disqualified the lawyers from the case, reprimanded them publicly and referred the matter to the state bar234. We think these cases share a structure with Moffatt: the party answerable for what the AI produced is the party that put it to use.

The pattern across the cases: demand, process, measurement, scope and supervision

Table 1 sets the cases beside each other; only its last column is this chapter's reading.

CaseSource and its kindWhat the source reportsOur reading of what failed
Klarna, February 2024The company's own announcement, with no audit among the cited sources382.3 million conversations in a month; work equal to 700 full-time agents38Not a failure in the record; abundant demand and measures of volume, speed and repeat contact
Klarna, May 2025Entrepreneur, reporting a Bloomberg interview (secondary)111The chief executive on hiring human staff again and always offering a humanScope and hand-off
Moffatt v. Air Canada, 2024Tribunal decision231Liability for negligent misrepresentation; $812.02 in total231Scope and supervision
NEDA, May 2023CBS News (secondary)238Information that may have been harmful and unrelated to the program; program taken downScope; supervision
McDonald's and IBM, June 2024Nation's Restaurant News (secondary)239Test at about 100 locations ended; accuracy reported in a 2022 report as in the low 80% range, against a 95%-plus goalA threshold was set; the reason for ending is not stated
MyCity, January 2026THE CITY and The Markup (secondary)240Answers that would lead businesses to illegal behavior; removal plannedScope; supervision
Commonwealth Bank, August 2025Information Age (secondary)241Reversal of 45 job cuts; the bank admitted an error in declaring the roles redundantThe capacity assumption, H
Mata, Wadsworth, JohnsonCourt sanction orders232,233,234$5,000; $3,000 and $1,000 each; reprimand, disqualification and bar referralSupervision

Table 1. The documented cases, the kind of source behind each, and this chapter's reading of what failed.

Scope and supervision appear in most rows and demand in none as the thing that failed. Measurement appears in two rows, as the reason the case is informative: Klarna had figures to report and McDonald's had a threshold to test against. That is not evidence that demand and measurement never fail; it is the selection effect.

The reading has two further limits. The sources are of uneven quality, with five of the eight rows resting on press, and the classification is judgment: the McDonald's case could be a failure of accuracy, of scope in an open setting, or a business decision, and the sources do not say which. The table organizes the record; it does not establish causes.

Not every deployment in the record fails, and the evidence on customer contact is mixed. Klarna's 2024 figures, taken at face value, describe gains in speed and repeat contact38. In a natural field experiment at a large telecommunications company, Wang and co-authors found that a voice-based AI system replacing the IVR temporarily increased machine-service duration and customers' demand for human service50. The system persistently reduced customer complaints, though50. A 2026 preprint reporting a randomized field experiment on Alibaba's Taobao found shorter chats and substantially lower ratings for AI-eligible chats51. And Brynjolfsson, Li and Raymond found that access to a generative AI conversational assistant raised the productivity of customer support agents, measured as issues resolved per hour, by 14% on average246. That assistant is a tool for human agents, not an AI agent replacing them246. These differ in task, in what they measure and in whether the AI replaced or assisted a person, and they cannot be pooled. What they share is an outcome measure, and the measure decides the verdict: speed and ratings moved in opposite directions in the same Alibaba deployment.

The cases support narrower statements. First, the business answers for what its agent says, as in Moffatt and in the sanction orders. Second, the failures that surfaced cluster where words carry consequences, in law, health advice and legal obligations. Third, a statement of results is not a result: Klarna's release was a statement, a later press report has its chief executive speaking of hiring human staff again, and the McDonald's reports show a stated threshold and an outcome the company did not explain.

Fourth, the problem surfaced through outsiders. In the cited accounts of Moffatt, NEDA, MyCity and Mata (press reports for NEDA and MyCity), it came through a customer, critics, reporters or a court231,232,238,240. A firm that logs and reads its own contacts can find problems before an outsider does. We think that is the measurement complement in a protective role; none of the sources states it.

What does not transfer from enterprise AI failures to a small firm

Most of the cases concern organizations with a scale that a small service business does not have: a company handling millions of chats, an airline, a fast-food chain, a city government and a bank. The differences cut in both directions.

Scale makes error rates measurable and makes small rates costly. A company handling millions of conversations, as Klarna reported in one month38, can estimate a small rate from its logs with little delay. A rate of a fraction of a percent still produces many wrong answers. A small firm handles far fewer contacts, so it cannot estimate a small error rate from its own logs quickly. An error that would be tolerable at scale may be a rare event it does not see until it matters. The same smallness lets a person read every transcript for a period, which an enterprise cannot. That is an argument from first principles, and Chapter 7 treats how long a small firm must wait before a difference means anything.

Several enterprise cases concern replacing staff: Klarna's equivalent of 700 agents, by its own statement38, and Commonwealth Bank's 45 roles, as Information Age reported241. The use this paper centers on is different: for a small firm the comparator for the agent is often a contact that no one would have answered, the call after hours or in a busy hour, not a person the agent replaced.

How often such calls go unanswered has been measured by vendors and by one secret-shopper study. Patient Prism† (call analytics) reports that of every 100 patient calls in 2025 to 8,280 dental locations, 31 hung up before reaching an agent34. Invoca† (call tracking) reports that a person answered 54% of inbound calls to health care organizations, across six health care fields64. Clio†'s 2024 secret-shopper study of 500 law firms found that 40% answered the phone call and that, by Clio†'s reading, 48% were essentially unreachable by phone63. These measure different things (hanging up, being answered by a person, being reachable) in different professions, so we do not average them or set them side by side as one rate.

A study of missed calls at accounting firms is NF; the two nearest neighbors put the share of calls not reached by a person at 46% among health care organizations in Invoca†'s data (100 - 54) and at 60% among the law firms in Clio†'s test (100 - 40). We estimate that an accounting firm without a dedicated receptionist misses calls at a rate of the same order, roughly four to six in ten (from the two neighbors, not their mean, since the measures differ). A firm with a full-time receptionist will sit below it. The share of calls that arrive after hours in any of the four professions is NF in a source we could open; the nearest evidence is the figures above. We think, then, that the comparator is often an unanswered contact; the figures above fit that and none proves it.

If it holds, a replacement failure like the bank's does not map onto capture, while a quality failure such as a wrong answer maps closely. A wrong answer given to someone who would otherwise have reached voicemail can still cost the firm that customer and can carry the legal consequences set out in Chapter 9.

The domain differs too. NEDA's chatbot ran a program of the National Eating Disorders Association, and New York City's chatbot gave answers that, as reported, would lead businesses to illegal behavior if followed. We think both are settings where users bring high-stakes questions and a wrong answer is costly. A booking agent confined to a list of facts is a narrower tool, and the scope failures in those cases transfer to it only if it is allowed to drift. A professional practice is likewise a setting where callers can be expected to ask clinical, legal or tax questions of the front desk, and the sector rules in Chapter 9 attach to such content. What transfers is the discipline of saying in advance what the agent will not answer.

Resources differ as well. A large organization is likelier to have legal staff, and a small firm often has no legal or compliance department, as Chapter 1 argued. The cited Moffatt decision places responsibility on the business for the information on its website and says nothing of its size, but the buffer around that responsibility depends on size. A firm of 10 people may have one person to supervise the agent, and the supervision then competes with that person's other work.

None of the cases documents a failure of an AI agent answering a small service business's calls, and a set selected through visible wrongs cannot show that none has happened. Only two touch on voice, McDonald's voice ordering test and the voice bot that the union said Commonwealth Bank had announced, and the first is an ordering test and the second a staffing decision. The others concern chatbots or court filings. Chapter 9 estimates how often a bounded script states something false, from one hospital study of outbound calls, and says the estimate likely errs low. A randomized or field study of an AI agent answering inbound calls in dental, medical, legal or accounting settings is NF; the nearest evidence is that hospital study of outbound calls, so there is no measured rate to set beside the cases.

We think what does transfer is a set of questions a small firm can answer before launch and check after it:

  • Which facts may the agent state, and who keeps each one current?
  • Which subjects does it hand to a person, who receives the hand-off, and how fast?
  • What does the firm record for each contact, and who reads it?
  • What result, set before launch, would lead the firm to narrow the agent or switch it off?
  • Whose name does the agent speak in, and does the person on the other end know it is an AI?

Chapter 13 turns these questions into signs that a deployment is not paying, and into what to switch off.

Chapter 11

AI profit in professional practices: one identity, four sets of constraints

A dental practice, a medical practice, a law firm and an accounting firm can earn from an AI agent in the same four ways. In each, a different constraint makes one of the four worth more than the others. This chapter applies the profit identity of Chapter 4 to the four professions and compares them.

The identity has four mechanisms for the agent, and the system ledger adds one line above them. Demand is the gross margin from the contacts the system adds, dN x c x V, where dN is the contacts a month the system adds.

Capture is the gross margin from contacts that would otherwise have been lost, N x m x a x c x V. N is inbound contacts a month, m the share that go unanswered or late without the agent, a the share of those that the agent answers and holds, c the share of answered contacts that become booked customers, and V the first-year gross margin of one new customer. Conversion is the lift dc in c on contacts already being answered, N x (1 - m) x dc x V. Capacity is H hours of staff or professional time released a month, each worth w in its next best use, so H x w. Retention is B customers on the books, a change dr in the monthly share kept and a lifetime value L, so B x dr x L. Against these stand the run cost k x n, the fixed monthly cost F and the expected monthly cost of error E.

The mechanism each section calls plausibly largest is our argument from the profession's structure. A randomized or field study of an AI agent answering inbound calls or scheduling in dental, medical, legal or accounting settings is NF; the nearest is one hospital's study, in which an AI voice assistant making outbound calls before a procedure completed 86.4% of them17.

The search figures are US monthly searches in Google Ads keyword data collected in November 2026. The phrases are ones a practice owner or a vendor would type, not ones a patient or client would. Volume shows interest, not purchase, and cost per click is what advertisers pay, not what a client is worth. Nothing here is legal, medical, tax or financial advice. Where the chapter describes a rule, it says what the rule says. The companion paper, Technical marketing for professional practices, sets out each profession's advertising and privacy rules in depth and reports our study of AI recommendations in full.

How the model ledgers are built

Each profession's case is a ledger with cited figures on both sides and one unknown solved for. Every input is a published price, a cited payment, rate or count, or our arithmetic on them, shown in brackets or marked calc. A quantity nobody has measured, such as the share of unanswered callers an agent turns into a conversation, gets no value of its own. The case shows the count or share needed to cover each purchase and sets it beside a cited base rate, such as the new patients a practice sees in a month or the share of visits missed. Where a neighboring measure exists, the text estimates from it and says which way the estimate likely errs.

One rule fixes what is set against what: a fee counted as a cost has its product counted as a gain. The plan produces demand, so it is set against new clients. The agent answers contacts, so it is set against answered calls and released hours. For a dental or medical practice the HIPAA-ready program is the compliance layer for tools that touch patient information. Every tool that runs on it shares its fee, so no single gain is asked to carry it, and the medical case sets it against kept visits together with the reminders add-on.

Our prices set the cost side. Plans are $999, $1,999 and $2,999 a month per location18. The two lower plans carry a one-time onboarding fee of $499 and $999 and the third a $3,000 setup18. A 12-month term paid in full makes two months free, at $9,990, $19,990 and $29,990 a year, and a 12-month term waives the two onboarding fees and cuts the setup to $1,400; the cases use monthly billing and leave these plan fees out. The voice agent is $299 a month with 250 call minutes included, and $0.39 a minute after that18. HIPAA-ready Essentials is $699 a month after $1,500 setup, per location and after a compliance consultation56. A plan carries no patient information, so we sell it to a health care practice with no program56, and the least a dental or medical practice can buy is the $999 plan18.

The program is a layer added to a plan and does not include the agent56. The agent on patient calls, the site assistant, appointment forms, texts to patients and call tracking each need one56, so its fee is shared by everything that runs on it. Four front-desk add-ons run inside it: reminders and rebooking at $149 a month, recall and reactivation at $199, waitlist backfill at $99, and intake with insurance capture at $149 plus $250 setup, or $449 a month for all four, per location, plus usage56. The cases charge only the reminders add-on, against kept visits in the medical case.

The agent's 250 included minutes are about 62 four-minute calls (250 / 4), and every call chain below stays inside them, so the agent costs $299 in the cases18. Past them a four-minute call costs $1.56 (4 x $0.39), and where a case uses more minutes the fee is $299 + (minutes - 250) x $0.39. With the Essentials setup spread over 12 months, the program costs $824 a month in the first year ($699 + $1,500 / 12) and $699 after, and Table 1 gives what each purchase costs.

What is boughtMonthly cost
A plan alone: the marketing. It carries no patient information, so a dental or medical practice can buy it with no HIPAA program$999
Agent added to a plan the firm already has, with 250 minutes included$299
Lowest plan plus agent (law, accounting)$1,298
HIPAA-ready Essentials, first year (setup over 12 months): the compliance layer that every tool handling patient information runs inside. It is not part of the agent's price$824
Essentials after the first year$699
Agent plus Essentials added to a plan, first year: a practice whose only patient-facing tool is the agent$1,123
Lowest plan plus agent plus Essentials, first year (dental, medical)$2,122
The same after the first year$1,997

Table 1. What each purchase costs a month, by scope (calc. from our prices18,56; the first year's $824 is $699 + $1,500 / 12).

The agent costs $299 wherever it is bought, inside its included minutes (Table 1). Essentials is a layer around it and not part of it. The $1,123 row of Table 1 is the narrow case: a practice with no program that buys one for the agent and runs nothing else on it. A practice that runs the site assistant, patient texts or the add-ons on the same program shares its fee among them.

A client is valued from cited payments and cited frequency. These are payments before the practice's costs, not the gross margin V of the identity, and the dental case shows how a count moves when the practice keeps only part of a payment.

Errors are not priced: Chapter 9 takes their consequences from cited penalties and leaves their probabilities as unknowns, and each case here shows the margin an error would have to consume before a purchase stopped paying. Conversion lift and retention are not needed to reach break-even in any case, so they stand as gains beyond the counts shown.

Dental AI, dental marketing and HIPAA

Which mechanism is plausibly largest in a dental practice

For an established dental practice we argue that retention is plausibly the largest mechanism, with capture second. Dental care recurs: in the early release of the 2025 National Health Interview Survey, 66.7% of US adults had a dental exam or cleaning in the past year247. A practice's patients are a base kept in a recall cycle, and a small rise in the monthly share kept, dr, is multiplied by a lifetime value, L, across a base, B, that may number in the thousands. Demand sits above both for many practices: in the ADA Health Policy Institute's Q3 2026 poll of 704 private-practice dentists, nearly two in five said they were not busy enough and could have treated more patients74.

The nearest evidence on keeping patients concerns booked visits, and it is thin. A 2013 Cochrane review pooled trials of reminders for healthcare appointments in which attendance was 67.8% with no reminders, 78.6% with text reminders and 80.3% with phone call reminders77. The authors called the evidence insufficient to inform policy77, and the trials concerned healthcare appointments in general. Text reminders did about as well as phone calls, so a plain automated text system would be expected to capture most of such a gain. The part attributable to an AI agent is the increment over that system, ~1.7 points (the gap between call and text attendance in the trials, 80.3% - 78.6%; likely high). Three trials that compared the two directly found a similar effect, with a risk ratio of 0.99248.

The strongest case against this ranking is capture, which would be larger in a practice with a thin base or unanswered calls. Patient Prism† (call analysis) reports that of every 100 calls to its 8,280 customer locations in 2025, 69 connected with the front desk and 31 hung up before reaching anyone34. We take ~31% as m for a dental office (from that hang-up share; likely high as a share of callers lost, since some ring back). A secret-shopper phone survey of 185 cataloged Arizona Medicaid pediatric dental practices reached 86% (159) on behalf of a Medicaid-enrolled child and 81% (149) on behalf of a commercially insured child, "reached" being the paper's own term62. That leaves 26 and 36 practices not reached (185 - 159 and 185 - 149), in one state and one kind of practice.

The binding constraint: privacy law

HIPAA binds a dental agent most tightly, because a caller who books a visit gives it protected health information. HHS describes a business associate as a party that performs functions on behalf of a covered entity that involve creating, receiving, maintaining or transmitting such information163. The Privacy Rule permits disclosure to a business associate if the covered entity obtains satisfactory assurance that it will safeguard the information195, documented in a written contract195. We think a dental agent runs only under a business associate agreement, and that is our practice.

In our check of 28 software tools that small practices commonly use, the vendors' own pages showed that 9 will sign without a special tier, 15 sign only on certain plans and 4 offer none20. The four are Zapier, Calendly, Google Analytics and Meta Pixel20. Telephony, voice and call-tracking products are among the 15 plan-gated tools20, and 6 of the 15 do not publish the price of the qualifying plan20, so the fixed cost of such a stack may not be readable from public price pages.

HHS's tracking bulletin gives the example of a patient who makes an appointment through a clinic's website that uses third-party tracking technology to send the appointment and IP address to a vendor: the vendor is a business associate and an agreement is required159. In June 2024 a court vacated a different element, concerning an IP address and a visit to an unauthenticated public page about health conditions or providers159. The court said the vacatur does not limit the legal operability of the other guidance160. Separately, the Privacy Rule defines marketing, in general, as a communication about a product or service that encourages recipients to buy or use it, with listed exceptions249. It requires an authorization for marketing uses of protected health information, with narrow exceptions250. Whether a recall message from an agent is marketing is for the practice's counsel.

What search demand shows dental owners want

"Dental marketing" and "dentist marketing" are each searched 1,300 times a month and "dental seo" 1,000 times5. The AI phrases are smaller: "dental ai" 880, "ai dental receptionist" 170 and "ai for dentists" 1405. The largest marketing phrase is searched about 1.5 times as often as the largest AI phrase (1,300 / 880 = 1.48). Dentistry is therefore the exception to law and accounting, where the most-searched AI phrase draws several times the searches of the most-searched marketing phrase.

In our study of AI recommendations, which put 199 local questions to three assistants in 16 US cities19, the assistants named 266 dental businesses; 3 (1.1%) were named by all three and 88% by only one19. Among the ten kinds of business asked about in every city19, only accountants had a lower share, 0%19; the next lowest were real estate agents at 1.5% and financial advisors at 1.8%19. Each question was asked once, names were matched by spelling, up to eight names were read per answer, and answers came through an API, which is a baseline and not what every signed-in user sees19. The study shows how rarely the assistants agreed on dentists, not that any of them is wrong.

What a dental agent may do and what stays with a person

A dental agent may do administrative work: answer calls and forms, book and move appointments, give hours, take insurance details for the front desk, and hand off. It tells every caller at the start that it is an AI assistant, which is our practice. California's section 1339.75 names health facilities, clinics and physician and group practice offices, not dental offices177. It excludes administrative matters such as scheduling from the clinical information to which its disclaimer duty attaches177. Dentists already use AI at the front desk. In the ADA Health Policy Institute's 2026 panel, 13.6% of responding dentists used it for insurance verification and 10.1% for check-in, and 43.3% used it for at least one type of task251.

Clinical content, public replies and promises stay with a person. HHS announced that Elite Dental Associates agreed to pay $10,000200. Its resolution agreement recites a patient's complaint that the practice disclosed her information in a reply to her online review201. New Vision Dental paid $23,000 over disclosures of patient information in response to online reviews252. Both were settlements with no admission of liability. A North Carolina dental practice that did not answer OCR's data request had a $50,000 civil money penalty imposed, which was a penalty and not a settlement202. Promises about results stay with a person because the agent's words are advertising. The ADA's Principles of Ethics say no dentist shall advertise in a manner that is false or misleading in any material respect253. The code is an ethics code and not law.

Dental market facts

The ADA Health Policy Institute counts over 200,000 US dentists, 60.0 per 100,000 people in 2025254, and one in three works in a single location with no other dentist254. Separately, the Quarterly Census of Employment and Wages shows 136,040 private offices of dentists in 2024 with 1,039,896 employees on annual average255, or about 7.6 employees per office (1,039,896 / 136,040). The two sources count people and establishments by different rules, and we do not combine them. The Institute reports that expenses per dentist are rising more than revenue254. What a patient pays is measured from the household side: average dental spending from all payers per person with a visit was $887 in 2021, as households reported it55.

The dental case: what a patient is worth and what each purchase needs

Take a general practice of one dentist or two, sizes chosen for the example. An average general dentist in private practice billed $965,660 in 2025 and earned $215,32053, which is $80,472 of billings a month ($965,660 / 12), with net income at 22.3% of billings ($215,320 / $965,660). Statista, which does not name its source on the page, gives about 3,429.5 patient visits a year per dentist in 202554. That makes about $282 of billings a visit ($965,660 / 3,429.5) and about 286 visits a month (3,429.5 / 12). One patient is worth $887 a year in spending55. Henry Schein One† (practice software) reports that its customer practices see an average of 30 new patients a month13.

At $999 a month, the plan alone is covered by 1.1 new patients a month on their first year of spending ($999 / $887), 3.8% more than the 30 of the average practice in Henry Schein One†'s data ($999 / $887 / 30).

Everything together, the plan, the agent and Essentials, costs $2,122 a month in the first year, 2.6% of one dentist's billings ($2,122 / $80,472) and 1.3% of two dentists' (calc.). The count it needs depends on how much of a patient's spending the practice keeps. If it keeps all of it, 2.4 new patients a month cover the system in their first year ($2,122 / $887); at half, 4.8; at a quarter, 9.6; and at 22.3%, the average dentist's net income share, 10.7; in visits, it is 7.5 more kept visits a month at average billings per visit ($2,122 / $282), out of about 286 a month per dentist. That line is rough: the visit count may include hygiene visits, which would make billings per visit look lower than a dentist's own visits bring. The net income share is a floor, because rent and salaries are already paid when one more patient is seen.

The agent is the part that answers calls, and it is a service purchase. Added to a plan, it costs $299 a month, its published price with 250 call minutes included18. A dental office also needs the HIPAA-ready program around it56, $824 a month in the first year and $699 after (Table 1). The program is a layer and not a tool, and its fee is shared by everything that runs on it, so it is not charged to answered calls alone.

The agent is set against front-desk time and against calls. Front-desk time costs $34.59 an hour, the employer's full cost of office and administrative work in health care in June 202673, so the agent's $299 is 8.6 hours a month ($299 / $34.59), about 25 minutes a working day (8.6 x 60 / 21 working days). Set against calls, the agent rests on a chain of cited links with one link open: the share of callers who hang up that an agent turns into a conversation. Table 2 gives the links, and Table 3 solves for the open one. Patient Prism† counts new and existing patients together, so a booking is valued at two ends: one visit at the low end and a new patient's first year at the high end.

LinkFigureSource
Calls a month per location116Patient Prism†: 11,552,668 calls at 8,280 locations in 2025, divided by 12 (calc.), so a mean of totals and not a measured average per office34
Callers who hang up before reaching anyone31%, so 36 a month; 69 in 100 get throughPatient Prism†34
Of the 69 who get through, calls that are a real chance to book34, or 49%; the other 35 are not bookablePatient Prism†'s own classification34
Chances that end in a booking21 of 34, or 62%Same source
Bookings a month among the hang-ups, if they resemble callers who get through and every one were answered11.0 (36 x 21 / 69)calc.
A booking, low end: one visit$282ADA billings per dentist divided by a visit count whose source is not named, so rough53,54
A booking, high end: a new patient's first year$887AHRQ MEPS55
Value that hangs up each month$3,089 to $9,730calc.

Table 2. The dental call chain: cited links (Patient Prism†) and calc. on them.

Share of hang-ups turned into conversationsCalls a monthBookingsGain, one visit eachGain, a new patient eachLess the agent ($299)Less the agent and a program bought for it alone ($1,123)
10%3.61.1$309$973$10 to $674-$814 to -$150
25%9.02.7$772$2,433$473 to $2,134-$351 to $1,310
50%18.05.5$1,544$4,865$1,245 to $4,566$421 to $3,742
75%27.08.2$2,317$7,298$2,018 to $6,999$1,194 to $6,175

Table 3. The open link at four values, with each booking valued at both ends (calc. from Table 2 and Table 1). The range is not a forecast, and the four shares are values we chose. Answered calls are the agent's gain, so the chain is set against the agent's own fee; the last column is the narrow case of Table 1.

The agent is covered when it turns 3% of hang-ups into conversations if each booking is a new patient (0.3 bookings a month, $299 / $887), or 10% if each is one visit (1.1 bookings, $299 / $282; calc., Table 2). A practice's own phone log shows where between the two ends it falls. In the narrow case, the last column of Table 3, the shares are 12% (1.3 bookings) and 36% (4.0 bookings). The plan is not charged against this chain, because its product is new demand, which the counts above already set against it; charging it here too would count a fee with nothing credited for it.

The chain assumes callers who hang up resemble callers who get through, which the source does not show. It counts none of the callers who ring back, which lowers every figure, most of all at the low end, where the callers are existing patients. The one peer-reviewed figure for an AI voice agent finishing its calls is the 86.4% of the hospital study above, which is a guide and not a measure of a front desk.

At half of hang-ups turned, the agent clears its $299 by $1,245 to $4,566 a month (Table 3), or by $421 to $3,742 in the narrow case. That is the margin an error would have to consume before the purchase stopped paying.

Medical practice AI, the business associate agreement and the schedule

Which mechanism is plausibly largest in a medical practice

For a busy physician practice we argue that capacity and kept visits are plausibly the largest mechanisms, and that demand and capture change form rather than vanish. In AMN Healthcare's 2025 survey, research specialists called 1,391 physician offices in 15 large US metro areas to schedule a new-patient appointment59. The average wait across six specialties was 31 days, up 19% since 2022 and 48% since 200459. For family medicine it averaged 23.5 days and ranged from one day to 20759.

The survey is AMN Healthcare's own and covers only those metros, so it does not describe every practice. Where waits run to weeks, a further inquiry does not add a visit this week, but it fills the slots that missed and late-canceled visits leave, which run at 18.8% to 23% of appointments in the studies below. It also lets the practice choose better-paying work: for most specialties in 2016, payments per visit were highest for privately insured patients and lowest for Medicaid80. And a queue that stays long is the case for adding a provider.

Retention keeps its value for the same reason. In a study of patients aged 65 and over of 23 primary care physicians in Oklahoma, with data from 1999 and 2000, the median relationship with the physician had lasted 8 years81. A patient who leaves takes years of visits with them, not one slot, though the study says nothing about working-age patients.

Capacity comes in two kinds: staff hours spent on routine calls, and clinician slots lost to missed visits. A 2018 review of 105 studies found an average no-show rate in healthcare scheduling of the order of 23%, highest in Africa (43.0%) and lowest in Oceania (13.2%)77. In a randomized initiative at one safety-net clinic, adding live reminder calls for appointments with a predicted no-show risk of 15% or higher lowered the no-show rate to 33% from 36%86. The calls were model-driven only in that a predicted risk selected the appointments. The nearest study of a voice agent placing calls is the hospital study above, outbound calls before a procedure with no control group, so the effect of an agent placing reminder calls is unmeasured.

Cost pressure raises the stakes: in a June 2026 poll with 251 applicable responses, 84% of medical groups reported higher year-to-date operating costs, by about 11% on average256. The poll reports only on the groups that answered.

The binding constraint: the schedule, then HIPAA

The schedule sets the form a gain takes: in a busy practice a recovered call fills a gap or joins a waitlist rather than adding a visit the week it arrives. The HIPAA business associate rule applies to a physician's office as it does to a dental one, so the agent runs only under a business associate agreement and the findings on tools in the dental section carry over20,195.

Medicine adds disclosure rules aimed at clinical communication. California's section 1339.75 requires a physician's office or clinic that uses generative AI for patient communications about clinical information to include a disclaimer that the communication was AI-generated and instructions for reaching a human177. In an audio communication the disclaimer is given verbally at the start and end177. A licensed provider's reading and review of the communication exempts it177. Texas H.B. 149 requires the provider of a health care service in which an AI system is used to give a disclosure by the date the service is first provided181. The AMA's House of Delegates adopted a policy in November 2024 that direct patient interaction with AI be disclosed at its start, which is policy and not law237. HHS has proposed that a regulated entity's risk analysis consider the type and amount of ePHI an AI tool accesses; that Security Rule change is not final203.

What search demand shows medical owners want

The phone phrases outdraw the AI phrases. "Medical answering service" is searched 5,400 times a month, against 140 for "ai medical receptionist"5; the first is about 39 times the second (5,400 / 140 = 38.6). The agency phrases are also large: "healthcare marketing agency" 5,400 and "medical marketing agency" 4,4005. The one large AI phrase near medicine, "hipaa compliant ai", draws 6,600 searches a month but names no profession, so dental and other health owners may be typing it too5. Our study of AI recommendations has no row for physicians, so we draw nothing from it about medical practices.

What a medical agent may do and what stays with a person

A medical agent may do the administrative work the schedule creates: answer calls, give hours and directions, take scheduling requests, handle the reschedules a reminder produces, and hand off. Text reminders do not need an AI agent, since the Cochrane trials found them about as effective as calls77. An outbound call in an AI-generated voice falls within the FCC's 2024 ruling that such calls require the called party's prior express consent to initiate174. In a 2024 proposed rule, the FCC said the TCPA's requirements do not extend to technologies used to answer inbound calls175. That is a statement in a proposal, not a ruling; Chapter 9 treats the law.

Symptoms, medications, test results and triage stay with a person. A capability study explains why a rule is needed and does not show it can be relaxed. In a JAMA Internal Medicine study, licensed health professionals preferred a chatbot's reply to a verified physician's in 78.6% of 585 evaluations77. The questions came from a public forum and not a practice's patients, and the evaluators were professionals and not patients. Public replies to reviews stay with a person too. Manasa Health Center, a New Jersey psychiatric provider, paid $30,000 and agreed to a corrective action plan, according to HHS257. The resolution agreement says Manasa impermissibly disclosed the protected health information of four patients in response to their negative reviews on Google Reviews258. The agreement is not an admission of liability.

Medical market facts

The Bureau of Labor Statistics counts 862,800 physician and surgeon jobs in 2025 and reports a median wage of $275,930 in May 2025 for wage-and-salary workers259. The Quarterly Census of Employment and Wages shows 239,840 private offices of physicians in 2024 with 2,954,911 employees260, or about 12.3 per office (2,954,911 / 239,840). In the AMA's 2024 survey, 42.2% of physicians were in private practice, 18 percentage points fewer than in 2012, and 47.4% worked in practices of 10 or fewer physicians261. The AMA changed its survey vendor in 2024, so the comparison with earlier years carries that caveat261.

The medical case: one physician and four, and the tight case

Take a primary care practice of one physician or of four, sizes chosen for the example. Table 4 gives the inputs and their sources.

InputFigureSource
Payment per primary care office visit$186 mean, 2016AHRQ MEPS80
Payment per office visit to any medical doctor$347 mean, 2023AHRQ MEPS82
Office visits per person a year, all physicians3.21, 2019CDC NAMCS79
Share of visits made to primary care50.3%CDC NAMCS79
New patients as a share of primary care visits12.5%CDC NAMCS108
Appointments per primary care physician51.1 a week, 2021Rotenstein et al., Epic data107
No-show rate18.8% to 23%Ten main clinics of one VA medical center76; review of 105 studies77
Effect of reminder calls on no-shows in two studies0.5 to 3 pointsPhone reminders at the same VA center, a before and after comparison76; live calls, one clinic, randomized86
Years a patient stays with a physician8 median (patients 65 and over, one state)Mold et al.81
Front-office staff cost with benefits, health care$34.59 an hour, June 2026BLS73

Table 4. Inputs to the medical case, with their sources.

A person makes 1.61 primary care visits a year (3.21 x 50.3%), worth $300 a year at $186 a visit (1.61 x $186) and $2,400 over 8 years ($300 x 8). A physician's weekly load of 51.1 appointments comes to 221 a month (51.1 x 52 / 12), including the visits that are missed. The $300 a year, from Table 4's inputs, is an average over everyone, including people who made no visit, so a person who seeks a practice out is likely to bring more.

Some inputs in Table 4 carry limits: the $186 is a 2016 mean. NAMCS defines primary care more broadly than MEPS does. The appointment count comes from health-system data and is not a small-practice figure. The no-show rates, from one VA medical center and a worldwide review, are joined here with appointment volumes from a third source, so the missed-visit counts are a scale and not a forecast.

Each fee is set against the gain it buys. The plan, $999 (Table 1), is set against new patients and needs no program. The agent, $299, is set against answered calls, which Table 6 counts, and front-desk hours. The program, $824 a month in the first year and $699 after, is shared by every tool that runs on it (Table 1), so no single gain is asked to carry it. Table 5 sets the program and the reminders add-on, $973 a month together ($824 + $149)56, against kept visits, because reminders are what the two no-show studies measured.

A few costly visits pull the mean above the typical visit, which the median shows. At the 2016 median of $107 a primary care visit80, a patient pays $173 a year (calc.). The plan then needs 5.8 new patients a month ($999 / $173), the program and reminders 9.1 kept visits ($973 / $107) and everything together 19.8 kept visits ($2,122 / $107). At the 2023 mean of $347 for a visit to any medical doctor82, a patient pays $560 a year and the same counts are 1.8, 2.8 and 6.1 (calc.). Table 5 uses the $186 between them. These are payments, not what the practice keeps: a kept visit in a slot already staffed adds nearly all of its payment, and a new patient who needs added capacity adds less.

Table 5 sets each purchase against what it produces.

One physicianFour physicians
Scheduled appointments a month221886
Kept visits a month, once 18.8% to 23% are missed171 to 180682 to 719
New-patient visits a month now, at 12.5% of kept visits21 to 2285 to 90
Plan alone, $999: new patients a month needed, first year only3.3 (15.6% more)3.3 (3.9% more)
The same, counting three years, and 8 years1.1 and 0.41.1 and 0.4
Agent alone, $299: the fee in front-desk hours, at $34.598.6 a month (about 25 minutes a working day)8.6 a month (about 25 minutes a working day)
Program and reminders, $973: missed visits a month at 18.8% to 23%42 to 51167 to 204
What one point off the no-show rate is worth a month$412$1,647
Kept visits needed to cover the program and reminders5.2 (2.4 points off the no-show rate)5.2 (0.6 points off the no-show rate)
Kept visits needed to cover reminders alone, $149, where the program is already in place0.80.8
Kept visits if no-shows fall by 0.5 to 3 points, as in the two studies1.1 to 6.6 ($206 to $1,236)4.4 to 26.6 ($824 to $4,942)
Everything together, $2,122: share of payments at $186 a kept visit6.3% to 6.7%1.6% to 1.7%
Kept visits needed, as a share of missed visits22.4% to 27.4%5.6% to 6.9%
New patients a month needed, first year only7.1 (33.2% more)7.1 (8.3% more)
The same, counting three years, and 8 years2.4 and 0.92.4 and 0.9

Table 5. The medical case by scope, for one primary care physician and for four (calc. from Table 4 and Table 1). The agent's answered calls are in Table 6.

The agent's own fee is $299 a month, the cost of 8.6 front-desk hours ($299 / $34.59) or of the answered calls in Table 6, where covering it takes 2% to 33% of missed calls turned into conversations. The program is the larger fee, and every tool on it shares it. The program and reminders need 5.2 kept visits a month ($973 / $186), which is 2.4 points off the no-show rate for a solo physician, toward the top of the 0.5 to 3 points found in the two reminder studies, and 0.6 points for four physicians. A solo office therefore covers the program by running more than one tool on it: the agent's answered calls, reminders, recall, waitlist backfill and intake all sit on the same layer. A practice whose only patient-facing tool is the agent is the narrow case of Table 1, shown in the last column of Table 6.

Staff hours released are worth $692 a month at 20 hours, $1,384 at 40 and $2,767 at 80, at the loaded rate ($34.59 x hours). The hours at stake are the call time itself: ~20 to 34 hours a month, from the 299 to 505 calls a month per office location that Call Box† (call tracking and answering) counted33 and four minutes a call (errs both ways). It errs low as a measure of phone work, since it leaves out holds, callbacks and the tasks after a call. It errs high as a measure of hours an agent releases, since the agent hands some calls on.

The agent's $299 is 8.6 of those hours ($299 / $34.59), but it covers 250 minutes, 4.2 hours of talk (250 / 60), and each minute after costs $0.39 against $0.58 of front-desk time ($34.59 / 60). Hours alone therefore cover the fee when the agent takes 1,080 minutes of calls a month, 18.0 hours: $299 + (1,080 - 250) x $0.39 = $622.70, and 18.0 x $34.59 = $622.62. That is ~90% of the ~20 hours at 299 calls (18.0 / 19.9, from 299 x 4 / 60) and ~53% of the ~34 hours at 505 calls (18.0 / 33.7, from 505 x 4 / 60). So hours alone cover the agent only if it takes over that share of an office's phone time, and they add to the answered calls in Table 6 and do not replace them.

Call Box† counted those 299 to 505 calls a month across more than 1,700 health care practices on its platform in 2020, a pandemic year33. Two measures of calls that reach no one are cited, not averaged. Across VA medical centers, 12.0% of calls were abandoned before a staff member answered in October 201465 and 8.3% in September 201665; Table 6 uses the later figure. Among the health care clients of Invoca† (call tracking), 46% of calls were not answered by a person (calc., from the 54% answered in its 2026 report)64. Invoca† also reports that 43% of answered calls were leads and 45% of those leads converted on the call64, across six health care fields and not physician offices alone. The callers are new and existing patients together, so Table 6 values each lost booking at two ends: one primary care visit at $186 and a new patient's first year at $300.

Share of calls reaching no oneCalls a monthCalls missedBookings lost if they convert like answered callsValue, one visit each to a new patient eachShare of missed calls the agent must turn into conversations to cover its own fee ($299)The same, if a program is bought for the agent alone ($1,123)
8.3% abandoned (VA call centers, 2016)299254.8$893 to $1,44121% to 33%78% to more than all of them
8.3% abandoned (VA call centers, 2016)505428.1$1,509 to $2,43312% to 20%46% to 74%
46% not answered by a person (Invoca†'s health care clients, 2026)29913826.6$4,950 to $7,9854% to 6%14% to 23%
46% not answered by a person (Invoca†'s health care clients, 2026)50523245.0$8,361 to $13,4872% to 4%8% to 13%

Table 6. The medical call chain at two cited miss rates and two cited volumes, with each booking valued at both ends (calc. from the sources in the text and Table 4). Answered calls are the agent's gain, so the fee they are set against is the agent's own; the last column is the narrow case of Table 1.

Neither miss rate comes from private physician offices, and the two are different measures from different settings. Against the agent's own fee, the share of missed calls it must turn into conversations runs from 2% to 33% across Table 6: 12% to 33% at the VA rate and 2% to 6% at 46% not answered by a person. At each of those shares the agent's calls fit inside its 250 included minutes, so its fee is $299. Every row is below the 86.4% of calls that the hospital study's agent completed17, and the highest, 33%, is below the 36.6% of calls the same agent finished with no person17. That study placed outbound calls before a procedure and is not a front desk answering calls, so the comparison gives a scale and not a forecast.

For an office with 299 calls a month and one visit per booking, all 25 missed calls turned into conversations are worth $893 against the agent's $299 (Table 6); in the narrow case $893 does not cover the agent and a program bought for it alone. The lead and booking shares are Invoca†'s figures for answered calls, applied here to missed callers on the assumption that the two are alike, which the source does not show. Kept visits and front-desk hours, in Table 5, are separate gains on top of this one. In the four-physician practice, each point off the no-show rate beyond the 0.6 that covers the program and reminders leaves $1,647 a month (Table 5) for an error to consume before the purchase stopped paying.

Which mechanism is plausibly largest in a law firm

For a firm that takes new matters we argue that capture is plausibly the largest mechanism. An inquiry arrives when the person chose to make it, and an unanswered one is not queued, so the share unanswered or late, m, turns directly into matters not opened.

The value of a matter is large beside the cost of handling a contact. A Washington survey of private defense attorneys put the average attorney's fee for a DUI at $3,500 in 2020102, against at most $1.56 for a four-minute call (4 x $0.39). A firm short of time gains from demand in another form, by choosing the better-paying matters. Few firms are short of time: in the data of Clio† (legal software), the average lawyer bills 38% of an eight-hour day75. Retention is small where matters do not recur, as in a DUI defense. Capacity is the weaker term for a reason in the rules. ABA Formal Opinion 512 says a lawyer who bills by the hour must bill actual time205, so a released hour is worth its next best use, w, not the hourly rate.

The capture evidence comes from one source: a legal-software vendor's 2024 secret-shopper study, in which 500 law firms were emailed and phoned with a client inquiry63. It reports that 40% of the firms answered the phone call, down from 56% in 2019, and that 52% answered or called back63. On the vendor's reading, 48% were essentially unreachable by phone63. It also reports that 33% responded to an emailed inquiry, against 40% in 201963. The page we read does not say how the 500 firms were drawn, and the 2019 figures are the vendor's own restatement.

A share of firms that did not answer one call is not a share of calls that go unanswered. We estimate m for a law firm at six in ten, the firms that did not answer Clio†'s test call (likely high, since a firm that misses one call may answer the next). In Clio†'s 2019 consumer survey, 57% of those who had ever shopped for a lawyer said they contacted more than one firm262. And 42% said that if they like the first lawyer they speak with they will not need to speak with any others262. Those are stated intentions, seven years old.

The same study reports that, among firms that replied to email, 84% did so within eight hours but 18% gave clear next steps or cost information66. That bears on conversion more than capture, since a fast reply that says little is a different failure from no reply.

The binding constraint: conduct rules

Every rule named here is a model rule or an advisory opinion, and each state adopts its own version. A person who consults a lawyer about forming a client-lawyer relationship is a prospective client, and ABA Model Rule 1.18 bars the lawyer from using or revealing what was learned even when no relationship follows208. An intake agent collects exactly that information. Florida Ethics Opinion 24-1, an advisory opinion, says that website chatbots used for intake carry the added risk that a prospective-client or even a lawyer-client relationship is created without the lawyer's knowledge210. It also says that chatbots communicating with clients or third parties must comply with the restrictions on lawyer advertising and carry a disclaimer that they are an AI program and not a lawyer or employee of the firm210. We read the opinion in a copy hosted by a law school.

Rule 1.6(c) requires reasonable efforts to prevent inadvertent or unauthorized disclosure of information relating to a client's representation206. Opinion 512 says a lawyer who relies on others outside the firm to employ generative AI tools has additional supervisory obligations205. It also says that because many self-learning tools could lead to disclosure of client information, informed consent is required before such information is input, with engagement-letter boilerplate not enough205. Whether a prospective client's first call yields information of that kind is for the firm's counsel. Rule 7.1 bars false or misleading communications about the lawyer's services263. Rule 7.3(b) bars soliciting employment by live person-to-person contact when a significant motive is pecuniary gain, with listed exceptions264. The design here answers inbound contacts. An agent that calls prospects raises questions under that rule and under the telephone law of Chapter 9 that this chapter does not settle.

What search demand shows law firm owners want

"Legal ai" is searched 18,100 times a month, "ai for lawyers" 2,400 and "ai for law firms" 1,9005. The marketing phrases are smaller: "law firm seo" 3,600, and "attorney seo" and "lawyer seo" 2,900 each5. "Legal ai" is searched about five times as often as "law firm seo" (18,100 / 3,600 = 5.0), the reverse of dentistry. The phrases closest to intake are small: "legal answering service" 720, "legal intake software" 140 and "law firm intake" 705.

Advertisers pay the most for the phone phrase: a click on "legal answering service" costs $214.51, the highest cost per click in our whole keyword table5. "Legal ai" may be typed by people who are not practice owners. In our study of AI recommendations, the assistants named 240 family law attorneys across the 16 cities; 13 (5.4%) were named by all three and 75.8% by only one19.

In the design we recommend, a legal intake agent gives no legal advice. It answers inbound calls and forms, says at the outset that it is an AI assistant and not a lawyer, collects contact details and the type of matter within a script the firm approves, offers consultation times, and hands the contact to a named person. The Federal Trade Commission's action against a company that advertised the world's first robot lawyer ended in $193,000 in monetary relief under a final order approved in January 2025186,187. The complaint alleged that the company had not tested whether its chatbot's output equaled a human lawyer's and had not hired attorneys186. Those were allegations that a settlement resolved.

The decision to accept a matter, the conflicts check, the advice and anything filed with a tribunal stay with a person. In Mata v. Avianca the court imposed a $5,000 penalty on two lawyers and their firm for non-existent cases generated by ChatGPT232. The court said that existing rules impose a gatekeeping role on attorneys to ensure the accuracy of filings232. Opinion 512 says that relying on AI output without appropriate independent verification could violate the duty of competent representation205. These matters concern drafting and not intake, and they are the nearest evidence on what a wrong output costs a lawyer.

The ABA's 2025 Profile reports 1.37 million lawyers in the US in 2025, up from 1.35 million265. The Bureau of Labor Statistics separately counts 863,700 jobs for lawyers and reports a median wage of $159,670 in May 2025266. The ABA counts licensed active lawyers and the Bureau counts jobs, so the two do not compare. The Quarterly Census of Employment and Wages shows 188,520 private offices of lawyers in 2024 with 1,098,918 employees on annual average267, about 5.8 per office (1,098,918 / 188,520). In the first quarter of 2024, 141,518 of 187,335 offices, or 75.5% (calc.), had fewer than 5 employees268,269. These are establishments and not firms: a firm with five offices counts five times, and a lawyer with no employees is not counted. We estimate the share of lawyers in small firms from the offices' size, which overstates it, since a large office holds many lawyers.

The ABA's 2024 Legal Technology Survey, with 512 respondents on this topic, found that 30.2% of attorneys said their offices were using AI-based tools270. The share was 17.7% for solo practitioners and 47.8% at firms of 500 or more lawyers270. A legal-software vendor's survey of 1,028 US legal professionals, lawyers and support staff, found that 79% used AI in some capacity63. The two instruments count different things and are not a trend.

The law case: what a matter is worth and what each purchase needs

Take a small firm. A law firm buys no HIPAA program, so its two purchases in Table 1 are the agent added to a plan it already has, at $299 a month, and the lowest plan with the agent, at $1,298. Table 7 gives the inputs.

InputFigureSource
Hourly billing rate$300 median, 2023State Bar of Michigan survey of 2,191 attorneys109
Hourly rate, family law$343 average, 2025Clio†75
Share of the working day billed (utilization)38%; 25% for solo lawyersClio†75,88
Share of billed work invoiced, and of invoices collected88% and 93%Clio†75
Attorney's fee for a DUI$3,500 average, Washington, 2020Washington State Office of Public Defense102
Average auto bodily-injury claim paymentmore than $20,000, 2022Insurance Research Council103
Usual personal injury contingency feeone-thirdState Bar of California advisory104
Firms that answered a test call40%; 48% could not be reached by phone at allClio† secret-shopper study of 500 firms, 202463

Table 7. Inputs to the law case, with their sources.

An hour billed at $300 brings in $245.52 after realization and collection ($300 x 88% x 93%); a lawyer who bills 38% of an eight-hour day bills 3.0 hours (38% x 8) and leaves 5.0 unbilled (8 - 3.0); and one-third of a $20,000 injury payment is $6,667 ($20,000 / 3). The injury payment is what insurers paid, not a lawyer's revenue. In Clio†'s data, 59% of firms billed flat fees alone or beside hourly rates in 202475, so a rate times hours is not what every client pays.

What is boughtMonthlyA yearCollected hours a month to cover itDUI matters a yearInjury cases a year
Agent added to a plan the firm already has$299$3,5881.21.00.5
Lowest plan plus agent$1,298$15,5765.34.52.3

Table 8. The law case by scope (calc. from Table 7 and Table 1). The agent alone is set against answered calls; the plan and agent together against new matters.

The base rates sit far from the counts in Table 8. A law firm is not short of hours: the average lawyer in Clio†'s data bills 38% of the day, and Clio†'s solo lawyers 25% (Table 7). Six firms in ten did not answer Clio†'s 2024 test call, and most shoppers contacted more than one firm.

Calls a month at a law firm are NF; the nearest evidence is Clio†'s one-call test. New matters can be bounded instead. In Clio†'s data, a lawyer collects on 2.4 hours of billable work a day75, which at the $300 Michigan median is $720 a working day. We estimate that a lawyer who handled only DUI matters would need a new one about every five working days to keep that pace (3,500 / 720 = 4.9; likely high for practices with larger matters or repeat clients). Against that pace, the 4.5 DUI matters a year that cover the lowest plan and the agent are about four and a half weeks of one lawyer's new matters (4.5 x 5 working days = 22.5 days). Each DUI matter a year beyond the 4.5 in Table 8 leaves $3,500 for an error to consume.

AI in accounting: the talent pipeline, tax return information and capacity

Which mechanism is plausibly largest in an accounting firm

For an accounting firm we argue that capacity is plausibly the largest mechanism, from the supply of people. Graduates earning a bachelor's or master's degree in accounting fell to 55,152 in the 2023-24 academic year, down 6.6% from the year before, according to the AICPA's 2025 Trends report271. The bachelor's count was 40,817, down 3.3%, and master's degrees in accounting or taxation fell about 15% to 14,335271. These figures cover one academic year, so they do not show whether the fall continues. New candidates entering the CPA Exam pipeline numbered 42,626 in 2023, the highest since 2016, and 28,082 in 2024271. The AICPA says such a surge typically comes the year before a major exam change, which suggests that part of the fall is a pull-forward.

Against a smaller supply stands demand: the Bureau of Labor Statistics counts 1,595,200 jobs for accountants and auditors in 2025 and projects 5% growth from 2025 to 2035, over accountants in any industry and not only in accounting firms272. If the supply of new accountants falls while projected jobs grow, an hour of staff time is scarce and an agent that releases hours has a buyer for them. This paper does not measure hiring difficulty in any firm. Busy-season hours point the same way: in a longitudinal study of public accountants at a single national firm in 2002, weekly workload rose from 49 hours before the busy season to about 63 during it273.

A released hour must still be put to use. The IRS Office of Professional Responsibility's June 2026 alert says that billing clients for time not actually spent, or double billing for AI-assisted tasks, may violate Circular 230 section 10.27, depending on the facts222. The strongest objection is that capture matters more, because new clients are a firm's growth. That may hold in a firm with spare hours. A firm short of hours gains from demand in another form: it chooses better-paying work and raises its minimum fee. The median firm in the AICPA's 2025 MAP Survey grew its net client fees 6.7% over the prior year274.

Under Treasury Regulation 301.7216-3(a)(1), unless section 7216 or a regulation specifically authorizes it, a tax return preparer may not disclose or use a taxpayer's tax return information before obtaining the taxpayer's written consent216. The regulation defines the preparer to include a person who provides auxiliary services in connection with preparing returns, including one who develops software used to prepare or file a return217. Whether an agent vendor falls within that definition is for the firm's counsel.

The FTC lists tax preparation firms among the financial institutions its Safeguards Rule covers275. IRS Publication 5708 says tax and accounting professionals are financial institutions regardless of size and must keep a written information security plan218. A financial institution must notify the FTC no later than 30 days after discovering a notification event involving the information of at least 500 consumers219. The AICPA Code says a member shall not disclose confidential client information without the client's specific consent223. Before disclosing it to a third-party service provider, a member should either enter a contractual confidentiality agreement with the provider or obtain the client's consent223. Nothing in our claim for the interpretation mentions AI. We think a vendor that handles client information belongs in the firm's security plan.

What search demand shows accounting owners want

"Ai accounting", "ai for accountants" and "ai in accounting" are each searched 6,600 times a month at $38.95 a click5. The identical volume and cost suggest variants of one query, so we count them as one reading. The marketing phrases "accountant marketing", "accounting marketing" and "marketing for accountants" are each searched 260 times a month5. The AI phrase is about 25 times the marketing phrase (6,600 / 260 = 25.4), a larger gap than in law. We leave out "cpa marketing", at 1,300 searches a month5, because the data cannot show that searchers mean the profession. Searchers for "ai accounting" may do accounting inside a company or be students.

In our study of AI recommendations, the assistants named 272 accountants across the 16 cities, and none was named by all three; 90.4% were named by only one19. Accountants are the only one of the ten kinds of business asked about in every city with no business named by all three.

What an accounting agent may do and what stays with a person

An accounting agent may answer inbound inquiries about services and the engagement process, book consultations, tell a prospective client which documents the firm usually asks for, and take a message. The line it must not cross is the return itself. We think a first-line agent should discuss no client's return or account, and any consent position that status questions would need is for the firm to set.

The IRS Office of Professional Responsibility's alert says technology is a powerful tool and not a substitute for professional judgment222. Due diligence requires verifying the accuracy of facts, citations and calculations produced by AI, and practitioners cannot rely solely on it222. Firm procedures should cover staff training, secure data handling and AI accuracy monitoring, and third-party AI tools should be vetted222. The alert applies existing Circular 230 duties and is not a new rule.

Advertising rules reach the agent's words: the AICPA Code treats promotional efforts that create false or unjustified expectations of favorable results as false, misleading or deceptive223. An agent that told a caller what refund to expect would be making such a promise. Capability tests are a weak guide: ChatGPT 4 with few-shot training and the ability to reason and act, such as by using a calculator, passed the CPA, CMA, CIA and EA exams tested at an average of 85.1%226. We think an exam score does not by itself show that AI can do the work of accountants.

Accounting market facts

The Bureau of Labor Statistics reports a 2025 median pay of $83,680 a year, or $40.23 an hour, for accountants and auditors272. The Quarterly Census of Employment and Wages counts 63,177 private offices of certified public accountants in 2024 with 539,890 employees on annual average276. It also counts 37,615 tax preparation establishments and 71,408 in other accounting services277,278. In the first quarter of 2024, 44,800 of 62,828 CPA offices, 71.3% (calc.), had fewer than 5 employees268,279. As with law offices, these count establishments, not firms.

The one adoption figure here is a vendor's: a survey of 700 US accounting professionals that an accounting-software vendor commissioned in April 2025 reports that 81% said AI had positively affected their productivity280. The answers are perceptions and not measured productivity, and we do not set the figure beside the legal or dental ones.

The accounting case: what a client is worth and what each purchase needs

Take a small CPA firm. Like a law firm it buys no HIPAA program, so its purchases in Table 1 are the agent on a plan it already has, $299 a month, and the lowest plan with the agent, $1,298. The client that matters most to the count is an advisory client at a CPA firm that sells advisory services, from the CPA.com and AICPA benchmark of 206 self-selected practices, which is the high end for a small tax firm. The low end is a bookkeeping band from Ignition† (software for accounting firms). Table 9 gives the inputs.

InputFigureSource
Revenue per advisory client at a CPA firm that sells advisory services$17,867 a year, median, 2024CPA.com and AICPA benchmark of 206 self-selected practices105
Typical monthly fee for an advisory client$3,000, medianSame survey105
Advisory clients lost a year10%, medianSame survey105
Advisory clients served, and new ones67 and 14, mediansSame survey105
Hourly billing rate, all positions$159 median, fiscal 2022AICPA MAP survey281
Monthly bookkeeping and accounting fee, most common band$250 to $499 a month, 2025Ignition† benchmark of 219 customers, solo bookkeepers included106
Business tax return feemost often $1,000 to $1,499Ignition†106
Individual tax return feemost often $400 to $599Ignition†106
Base Form 1040 fee$236 average, 2025NATP fee study282
Net client fees per full-time professional$189,695 median, fiscal 2022AICPA MAP survey281

Table 9. Inputs to the accounting case, with their sources.

An advisory client who brings $17,867 a year and leaves at 10% a year stays about 10 years (1 / 0.10) and brings about $178,670 over that time ($17,867 x 10). We infer this from two medians; the survey does not print it. A bookkeeping client in the most common Ignition† band pays $3,000 to $5,988 a year ($250 x 12 to $499 x 12), and the lowest plan and the agent cost what 8.2 billed hours a month bring in at $159 an hour ($1,298 / $159). The advisory figures are medians for client advisory services only, reported by 206 firms with a median firm revenue of $7,050,000105, so they describe firms that sell that work well. Solo bookkeepers in the Ignition† sample pull its bands down.

Kind of clientFee a yearNew clients a year to cover the agent on a plan ($299 a month, 1.9% of one professional's fees)New clients a year to cover the lowest plan plus agent ($1,298 a month, 8.2%)
Advisory client at a CPA firm that sells advisory services (AICPA benchmark of 206 self-selected practices)$17,8670.20.9
Monthly bookkeeping client, most common band (Ignition†, solo bookkeepers included)$3,000 to $5,9880.6 to 1.22.6 to 5.2
Business tax return$1,000 to $1,4992.4 to 3.610.4 to 15.6
Individual return (Ignition†)$400 to $5996.0 to 9.026.0 to 38.9
Individual Form 1040, NATP average$23615.266.0

Table 10. The accounting case by scope (calc. from Table 9 and Table 1). The percentages are of one professional's median net client fees.

At a CPA firm that sells advisory services, 0.9 new advisory clients a year cover the lowest plan and the agent; one such client pays for more than a year of it and stays about ten years. At the low end, 2.6 to 5.2 bookkeeping clients a year cover it, and a firm that sells only base 1040s needs about 66. The count depends on which clients the firm markets for, which is the case for marketing for fit. The median advisory practice in the benchmark reports 14 new clients, with no period stated.

A measure of missed calls at accounting firms is NF, so the nearest neighbors stand in, and they measure different things. Six law firms in ten did not answer Clio†'s one test call63, and only 54% of calls to Invoca†'s health care clients were answered by a person64. We estimate from these that a small accounting firm without a dedicated receptionist misses calls on that scale; a firm with a full-time receptionist will sit below it. One advisory client a year leaves $2,291 beyond the $15,576 cost for an error to consume ($17,867 - $15,576).

What the four professions share

Three features are candidates for what the four share: high value per client, small teams and regulated speech. They hold to different degrees in our evidence. For value, divide each Bureau of Labor Statistics median annual wage by 2,080 hours, a 40-hour week over 52 weeks. That gives about $84.67 for dentists, $132.66 for physicians, $76.76 for lawyers and $40.23 for accountants and auditors259,266,272,283. The last reproduces the Bureau's own hourly figure, which is why we use 2,080. These are wage-and-salary medians before benefits, which were 30.0% of employer compensation costs across private industry in June 2026100, and owners' earnings differ. A four-minute call costs at most $1.56 (calc., above). An hour of professional time is expensive beside it, and we think that gap is why these practices can profit from an agent, and why an intake agent seldom releases the hour that matters, the professional's.

Small teams show in all four. In the Quarterly Census of Employment and Wages, dental offices average about 7.6 employees and physician offices about 12.3255,260, law offices about 5.8 and CPA offices about 8.5267,276 (calc., above; CPA offices 539,890 / 63,177). In the first quarter of 2024 the share of offices with fewer than 10 employees was 74.2% for dentists, 75.0% for physicians, 88.3% for lawyers and 86.2% for CPAs (calc. from the size counts)268,269,279,284,285,286. These count offices, not firms. Among physicians, 47.4% work in practices of 10 or fewer261. Chapter 1's point follows: a small practice has no analyst and no slack, and its owner is the measurement system.

Speech is bound by a professional code or rule in all four, some of them ethics codes and not law, and the agent speaks for the practice. The ADA's code, ABA Model Rule 7.1 and the AICPA's advertising rule each bar false or misleading communication223,253,263. The AMA's Opinion 9.6.1 allows no restriction on physician advertising except those justified to protect the public from deceptive practices287. Utah's statute says it is no defense to a violation of the statutes its consumer protection division administers that generative AI made the statement184. It lists health data, financial data and legal, medical and financial advice among high-risk interactions183.

Trust is a fourth feature. Darby and Karni (secondary) distinguished credence qualities, which are expensive to judge even after purchase72, and Day and Barksdale (secondary) argue that professional services are intangible and customized, so that evaluation before purchase is difficult71. Neither is about AI. We think the four professions sell services high in credence qualities, so a first contact is part of the evidence the client uses, and an agent that errs spends trust the practice built.

For the cost of error, E, HIPAA civil money penalties run from $145 per violation at the lowest tier to a calendar-year cap of $2,190,294 at the highest, in the 2025-adjusted amounts HHS published on January 28, 2026229. HHS's 2019 notice sets lower annual limits for the three lower tiers, and the current figures should be confirmed230.

A tax return preparer who knowingly or recklessly discloses or uses return information other than to prepare the return commits a misdemeanor punishable by a fine of up to $1,000, or $100,000 in some cases, or by up to one year in prison214. A civil penalty of $250 per disclosure also applies, capped at $10,000 a year215. For lawyers, the sanctions in our evidence concern fabricated citations and not intake. Insurers' statements on AI errors in medical, dental and accounting liability, and bars' statements on malpractice coverage of AI errors, are NF; the nearest prices for an error are these penalties and settlements. The probabilities are unknown in all four, which is why each case shows the margin an error would have to consume.

Comparing the four professions

The system loses money in one corner only: a practice that is not growing and whose intake and scheduling are already sufficiently optimized. The cited base rates put a typical practice some way from it. Nearly two in five dentists said in Q3 2026 that they could have treated more patients74. Of every 100 calls to Patient Prism†'s dental customers in 2025, 31 hung up before reaching anyone34. Missed visits ran at 18.8% across the ten main clinics of one VA medical center and about 23% across 105 studies76,77. Six law firms in ten did not answer Clio†'s test call63, and the median CPA firm's fees grew 6.7% in a year274.

Table 11 sets the four side by side. The search rows are monthly US searches.

DentalMedicalLegalAccounting
Mechanism plausibly largest (our argument)Retention, then capture; demand where the schedule has roomCapacity and kept visits; demand as filled gaps and better-paying workCaptureCapacity; demand as better-paying work
Binding constraintPrivacy law: HIPAA business associate rules195The schedule, which sets the form of the gain, then HIPAA59Conduct rules, including duties to prospective clients208Written consent for tax return information216
Evidence bearing on the mechanismPooled trials of reminders for healthcare appointments: attendance 67.8% without reminders, 78.6% with text reminders77Average new-patient wait of 31 days in 15 large metros59Vendor secret shop: 40% of 500 firms answered the phone63Accounting graduates 55,152 in 2023-24, down 6.6%271
Most-searched AI phrase naming the profession"dental ai" 8805"ai medical receptionist" 1405"legal ai" 18,1005"ai accounting" 6,6005
Most-searched phrase not about AI"dental marketing" 1,3005"medical answering service" 5,4005"law firm seo" 3,6005"accountant marketing" 2605
Size: private establishments and their employees, 2024136,040 offices, 1,039,896 employees255239,840 offices, 2,954,911 employees260188,520 offices, 1,098,918 employees26763,177 CPA offices, 539,890 employees276
Median annual wage, wage-and-salary workers$176,110, May 2025283$275,930, May 2025259$159,670, May 2025266$83,680, 2025272
Stays with a personClinical content, public replies, promises about resultsClinical content and triage, public repliesLegal advice, accepting a matter, filingsReturn information, tax positions, review of AI output
Cited consequence of error (probabilities unknown)OCR: $10,000 and $23,000 settlements, a $50,000 penalty200,202,252OCR: $30,000 settlement with a psychiatric provider257$5,000 sanction for fabricated citations in a filing232Civil penalty of $250 per disclosure, capped at $10,000 a year215
What one client brings in a year$887, spending per person with a visit55$300, 1.61 primary care visits at the $186 mean (Table 4)$3,500 a DUI matter102; $6,667 an injury case (Table 7)$17,867 an advisory client at a firm that sells advisory services, the high end105; $3,000 a bookkeeping client at the low end (Table 9)
Least a practice can buyThe $999 plan, no program neededThe $999 plan, no program neededThe $999 planThe $999 plan
Full system, a month in the first year$2,122 with the agent and Essentials$2,122 with the agent and Essentials$1,298$1,298
New clients needed to cover the full system, on first-year value2.4 new patients a month7.1 new patients a month4.5 DUI matters or 2.3 injury cases a year0.9 advisory clients a year, or 2.6 to 5.2 bookkeeping clients
Base rate beside it30 new patients a month, average practice (Henry Schein One†)21 to 22 new-patient visits a month for one physician, 85 to 90 for fourLawyers bill 38% of the day in Clio†'s dataThe median advisory practice reports 14 new clients, period not stated
The agent's own caseOn a plan: 3% to 10% of hang-ups turned. In front-desk hours the fee is 8.6 a monthOn a plan: 2% to 33% of missed calls turned. In front-desk hours the fee is 8.6 a month. The shared program and reminders: 5.2 kept visits a month, 2.4 points off no-shows for one physician and 0.6 for fourOn a plan: 1.0 DUI matters a yearOn a plan: 0.2 advisory clients a year

Table 11. The four professions compared. The mechanism and the items that stay with a person are our argument, the last six rows are calc. from Tables 1 to 10 with sources given there, and the other cells state cited facts.

Chapter 12

Operating the system: people, vendors and cadence

Technical marketing operations: why returns come from running the system

An agent that is installed and left alone drifts. The facts it states about hours, prices and services go out of date, the calendar it books into changes, a vendor changes a model, a voice or a price, and callers ask things nobody planned for. Each of these moves a term in the profit identity: the share of contacts held, the share booked, the run cost per contact or the expected cost of error. None shows itself unless someone looks.

Brynjolfsson, Rock and Syverson argue that technologies such as AI "enable and require significant complementary investments", including co-invention of new processes, products, business models and human capital43. Brynjolfsson and Hitt used data on 527 large US firms over 1987 to 199444. They found that the contributions of computerization were up to 5 times greater over 5- to 7-year differences than over one-year differences, which they link to large, time-consuming investments in complementary inputs such as organizational capital44. That finding concerns large firms and earlier computing and says nothing about small businesses or AI. We infer that in a small firm the complementary investment is mostly the attention of the owner and staff each month, a cost that belongs in the fixed monthly cost of the system (F).

Two legal facts make an owner for the system unavoidable. In Moffatt v. Air Canada, the tribunal wrote that Air Canada is "responsible for all the information on its website", whether it comes from a static page or a chatbot231. Utah's Code says it is not a defense to the violation of a statute administered and enforced by its Division of Consumer Protection that generative AI made the violative statement184. Chapter 9 sets out the rules and Chapter 10 reads the case. A firm cannot hand responsibility to the tool, so someone inside the firm must hold it.

We think operation is five jobs: keep the facts true, read what happened, reconcile it with the ledger, change one thing at a time, and handle the exception.

Roles in an AI receptionist deployment: owner, front desk and outside team

Four roles cover the work in a small practice, and one person may hold more than one. The system owner is a named person inside the practice with authority to approve what the agent says and does. The hand-off recipient is the named person whose phone rings when the agent passes a contact on. The reviewer for regulated content is the practice's own counsel or compliance lead, who approves wording that touches a regulated topic. The outside team keeps the technical system running.

RoleDecidesDoes each month
System owner, inside the practiceWording, hours, hand-off rules and any change to what the agent may doReads the monthly record, approves or declines the proposed change, signs the change log
Hand-off recipient and a named backupWhat to do with an urgent call or a messageAnswers within the agreed time and reports contacts that arrived wrongly
Reviewer for regulated contentWhether wording or a data flow fits the rules that bind the practiceReviews changes that touch a regulated topic and decides on any notification after an incident
Outside team, named peopleTechnical configuration within what the owner approvedPrepares the monthly record, proposes one change, applies approved changes and keeps the vendor register

Table 1. Roles, decisions and monthly duties in a small practice. The assignment is ours, not a standard.

The system owner matters most. Without one, nobody decides, and the agent's wording is whatever was typed at set-up. We think the owner of a ten-person service business is the proprietor or the office manager, because the role needs authority over wording, hours and hand-offs more than technical skill. The role's time is a cost, and the build, buy or hire section prices it.

An agent is only as good as the person who answers when it passes a contact on, and an urgent call that rings an unanswered phone is a failure of the whole system. Chapter 8 argues that hand-off is the design center; operationally it means a named person, a named backup and a written definition of urgent, set by the practice and reviewed monthly.

Regulated practices add the reviewer, who decides how a legal or professional rule applies to the practice. In our managed marketing, anything that names a customer, makes a claim or touches a regulated topic waits for the client's sign-off. NIST describes governance as "a cross-cutting function to inform and be infused throughout the other three functions" of its AI Risk Management Framework166. Table 1 is a small-firm reading of that idea: the same named people appear in every routine below. The framework is voluntary166, and we do not claim that it prescribes these roles.

The outside team in our practice is small and named. Our HIPAA-ready programs are architected by Keegan Justis and built only by Mike Golliet and Kyle Grenier. The program's accounts stay with those three people, with no subcontractors or rotating support staff, and nothing changes in a program without the same named team. Managed marketing is led by Sezen Güner. A small named team gives the practice someone to call and a short list of people with access. It also concentrates knowledge in few people, which is a continuity risk, and a practice choosing any outside team should ask how its work is documented and who covers when a named person is away.

AI vendor agreements: who signs a BAA and why the plan tier matters

A voice agent is a chain of services, and each link is a contract. In our check of 28 software tools that small practices commonly use, 9 will sign a business associate agreement (BAA) without requiring a special tier20. Another 15 sign one only on certain plans and 4 offer none, according to the vendors' own pages20. Table 2 names the 4 and the 1520. Vendors' positions change, so a register has to carry the date each was checked.

Position in our checkToolsNames listed
Signs without a special tier209not named in our published count
Signs only on certain plans2015, of which 6 do not publish the qualifying plan's priceChatGPT, Gmail, Claude, Jotform, HubSpot, Supabase, Salesforce, Twilio, Vercel, Typeform, GoHighLevel, Acuity Scheduling, ElevenLabs, CallRail, Vapi
Offers none204Zapier, Calendly, Google Analytics, Meta Pixel

Table 2. Business associate agreement positions of 28 tools in our check, from the vendors' own pages.

A BAA is the written arrangement that HIPAA's rules contemplate when protected health information moves to a vendor. HHS describes a business associate as a person or organization that performs functions or activities on behalf of a covered entity that involve creating, receiving, maintaining or transmitting protected health information163. A subcontractor that does the same on behalf of a business associate is itself one163. The rules permit a covered entity to disclose protected health information to a business associate if it obtains satisfactory assurances, in a contract or other written arrangement, that the information will be safeguarded163. Which tools in a given practice's chain need an agreement is for counsel.

In our HIPAA-ready programs, the practice signs a BAA with us, and every service in the chain signs one with us. No signed BAA means no patient data. These are our rules, not general requirements: compliance depends on a signed BAA and on how the tool is set up, and we apply the labels "compliant" and "certified" to no tool and not to ourselves, saying "HIPAA-ready" instead.

The plan tier matters for three reasons. The first is availability: the plan that signs may not be the plan whose price is published. Chapter 5 priced a voice stack at the rates vendors publish, and our check found Twilio, ElevenLabs and Vapi among the tools that sign only on certain plans20. A regulated practice cannot assume Chapter 5's k and F apply to it. The second is price opacity: for 6 of the 15 the qualifying price is not published, so the practice learns the cost only by asking20. The third is that the agreement is tied to a plan, so a change of plan, by the practice or the vendor, raises a question the register must answer and the reviewer decides: does the agreement still apply?

FieldWhy it is kept
Service and its role in the chainShows which links touch caller information
Whether an agreement is needed, and who decidedRecords the reviewer's decision
Agreement status, date signed and plan tier it coversTies the contract to the plan actually in use
Date the vendor's position was last checkedPositions change, and a stale entry misleads
Renewal and price-change dates, as the vendor prints themLets a cost step be planned for
Data held, where and for how longSupports the incident routine
Owner of the relationship inside the practiceGives each vendor one named contact

Table 3. Fields of a vendor register, and why each is kept. The list is ours.

Vendors print dates on which their prices change, and the register should carry them. Google lists Gemini 4 Carbon at $2.00 per million input tokens through November 30, 2026 and $4.00 from December 1, 202692, and lists Gemini 3.8 Flash at $0.75 through December 31, 2026 and $1.50 from January 1, 202792. These are prices as published on 2026-11-10. A cost step with a printed date is a scheduled event and belongs on the calendar before it arrives.

The monthly cadence: check the answers, read the calls, reconcile the ledger, change one thing

The cadence is four jobs that run once a month in a fixed order, with at most one change (Table 4). The first job looks outward at what customers will be told elsewhere, the second and third look at what the agent did, and the fourth acts only after the other three have been read.

JobInputOutputDone by
Check the answersA fixed list of questions per location, asked of each assistant the same way each monthA record of which answers name the firm and what was changedOutside team, read by the owner
Read the callsCall summaries for the month, every handed-off or abandoned call, and a sample of full transcriptsA list of wrong facts, missed hand-offs, unanswered questions and complaintsOwner and outside team together
Reconcile the ledgerThe carrier's call log, the summaries, the calendar and the customer recordContacts handled, contacts held, bookings, revenue and cost per booked customerOutside team, checked by the owner
Change one thingThe three outputs aboveOne approved change with its reason and the way it will be read, or no changeOwner approves, outside team applies

Table 4. The monthly cadence: four jobs, their inputs and outputs, and who does each. The cadence is ours, not an industry standard.

The first job is to check what assistants say about the business, because assistants are now one of the doors through which customers reach a business, as Chapter 6 sets out. In our practice a monthly AI answer check comes with every plan. A single month's answer is a sample. In our study each question was asked once, and answers obtained through an API are a baseline, not what every signed-in user sees19. About 4% of the businesses named in our study were named by all three assistants19, so we think one assistant's answer says little about another's. The record is useful for the direction of change, not for any one answer.

Reading comes second: what the agent actually said. Every call produces a summary. A reader should open every call that was handed off, ended early or drew a complaint, and a sample of the rest in full, looking for wrong facts, missing hand-offs, bookings made against the rules, a missing disclosure line and questions the agent could not answer, which become additions to the facts it holds. In our practice we read the call summaries with the client each month. Reading presupposes a record, and whether a call may be recorded is a legal question. Federal law says it is not unlawful for a private person to intercept a call when that person is a party to it or one of the parties has given prior consent, unless the purpose is a criminal or tortious act161. Washington's statute makes it unlawful to intercept or record a private communication transmitted by telephone or other device without first obtaining the consent of all the participants162. Which rule covers the practice's callers is for counsel.

A second reason to read comes from NIST, which lists human-AI configuration among the risks of generative AI, including automation bias and over-reliance167. An agent that has been right for many months invites its owner to stop reading, and the monthly job exists to prevent that.

The third job reconciles three records of the same month, which should agree. Where they do not, such as a call the log shows and no summary, a booking in a summary and none in the calendar, or a booked customer with no source, the gap is a finding and often a plumbing fault. The reconciled counts feed the identity: contacts handled (n), contacts held, bookings and revenue. With the month's invoices they give the cost per booked customer, which Chapter 5 argued is the figure that matters. Chapter 7 gives the data model.

Changing one thing, or nothing, is the fourth job. Reading a change is hard even for large advertisers. In 25 field experiments with major US retailers and brokerages, the median confidence interval on the return on digital advertising was over 100 percentage points wide48. The authors say informative experiments can easily require more than 10 million person-weeks, which makes them costly and potentially infeasible for many firms48. In Facebook experiments, observational methods often failed to produce the effects found in randomized experiments, even after conditioning on extensive demographic and behavioral variables49. Those findings concern digital advertising, not a small firm's phone line. We think they still warn that comparing one month with the next can mislead. Our rule is to change one thing, record why, judge it against leading indicators first (Chapter 7) and against revenue only after several months.

A short monthly meeting between the system owner and the outside team follows the four jobs. Its agenda is fixed: what the checks showed, what the reconciliation found, the one change proposed, and any open incident. In our HIPAA-ready programs this is the monthly call, alongside access reviews and log checks. The NIST AI Risk Management Framework names four functions, GOVERN, MAP, MEASURE and MANAGE166. The system map of Chapter 8 is MAP, reading the calls and reconciling the ledger are MEASURE, the change and incident routines below are MANAGE, and the roles above are GOVERN. The mapping is ours, as a memory aid, and takes from the framework only the names of its four functions.

Change control for prompts, scripts and rules

Everything that can change what the agent says or does is under change control. That includes its instructions, the facts it draws on, its hand-off rules, the tools and permissions it has, the disclosure line, the voice and model, and the telephone routing. A change that adds a tool or a permission is larger than a change of wording, because it widens what the agent can do. OWASP defines Excessive Agency as the vulnerability that enables damaging actions in response to unexpected, ambiguous or manipulated outputs from a language model, whatever is causing the malfunction171.

Six rules, ours, make a workable procedure for a small firm.

  1. Every change has a written request that says what changes, why, the expected effect and how the effect will be read.
  2. One change is made per cadence month, except a correction made under the incident routine.
  3. The system owner approves every change, and the reviewer for regulated content approves any that touches a regulated topic.
  4. A fixed set of scripted test calls runs before and after every change. It includes the hardest scenarios and attempts to make the agent leave its rules.
  5. The previous version is kept so that a change can be undone the same day.
  6. The change log records the date, the person, the reason and the result read at the next cadence.

The test calls are the core of the procedure. In our practice we call the agent with a client's most common and most difficult calls before it goes live, and for health settings we test with synthetic records and never real patients. Two risks justify the adversarial calls. NIST defines confabulation as the production of confidently stated but erroneous or false content that can mislead or deceive users167. OWASP defines prompt injection as user prompts altering a model's behavior or output in unintended ways170. It says such inputs can affect the model even if they are imperceptible to humans170. Scripted calls are therefore a floor, not a proof: some attempts will be unscripted.

Vendor-initiated changes go through the same procedure. A vendor can change a model, a voice or a price without the practice asking, and we think a new model can answer the same instructions differently, so the test calls are rerun after any such change. This paper does not measure how often this happens. The nearest evidence is the price calendar: Google prints price steps for two of its models within two months of the date we read its pages92. Where a change would alter what a caller is told about being served by an AI, it goes to the reviewer first. Callers in our practice are always told that they are speaking with an AI assistant, and the agent hands off to a person when asked. Chapter 9 covers the law on disclosure.

Incident handling: what happens when the AI agent is wrong

An incident, in this chapter, is any event in which the agent's behavior harmed or could have harmed a caller, the practice or a third party, or ran up an unplanned cost. Table 5 gives a first action for each kind. Given the two legal facts in the first section, an incident cannot be passed to the vendor or the tool: the practice owns the consequence, and the routine exists so that it is handled quickly and in order.

Kind of incidentFirst actionReference
A wrong statement of fact, such as hours, a price, a policy or a promiseCorrect the fact where the agent reads it, then the wording; the owner contacts the affected callerNIST confabulation167; the firm's responsibility231
A hand-off that failed, such as an urgent call not transferred or a message lostReach the caller by the named person; repair the route; rerun the test callsThe design center of Chapter 8
A disclosure failure, such as a caller not told that it is an AI assistantRestore the disclosure line at once; the reviewer advises on the rules that applyChapter 9
An exposure of personal information, such as a summary sent to the wrong person or data held in a tool without an agreementStop the flow; preserve the logs; the reviewer decides on notificationNIST data privacy risk167; Table 2
Manipulated behavior, such as a caller talking the agent into an action outside its rulesRemove the capability concerned; add the case to the test callsOWASP prompt injection and excessive agency170,171
Runaway cost, such as calls held open or a loopApply or lower the cap; check the invoiceOWASP unbounded consumption110

Table 5. Six kinds of incident, a first action for each and the reference used. The actions are ours.

The routine has seven steps.

  1. Detect. Incidents surface through a caller's complaint, the hand-off recipient, the monthly reading, the carrier log or a cost alert.
  2. Contain. Stop the harm before finding its cause, by routing calls to the named person or to voicemail, switching off the capability or lowering the cap. The owner and the outside team both know how, and the switch is tested before it is needed.
  3. Preserve. Keep the call, the transcript, the configuration version and the change log as they stood, and do not edit them.
  4. Assess. Establish what was said or done, how many callers were affected, and which rule or contract applies.
  5. Notify. Whether anyone must be told, and by when, is a legal question for the reviewer. The owner contacts affected callers as the reviewer directs.
  6. Repair. Correct the fact, rule or permission through the change procedure and rerun the test calls.
  7. Review. At the next cadence, record the cause and ask whether the same cause exists elsewhere in the system.

In the identity, E is the sum, over kinds of error, of probability times consequence. Chapter 9 takes the consequences from cited penalties and cases and leaves the probabilities as unknowns to be managed. The routine gives E a record, since an incident log is how a practice learns its own probabilities over time. A firm that logs nothing records a probability of zero, which is not a probability of zero in fact.

Cost control for an AI voice agent

Cost control is a set of limits, set before launch, on the ways the per-minute meter can run. Chapter 5 priced the agent and showed where its cost can move; this section turns that into controls. OWASP names the general risk Unbounded Consumption: an application that allows excessive and uncontrolled inferences, leading to risks that include economic losses110. For a voice agent the exposure is minutes. Chapter 5 showed that one call left open for 60 minutes costs $7.11 in its Configuration A and that 100 such calls would cost $711.

ControlWhat it limitsBasis
Cap on the length of a callCalls left open, loops and pranksChapter 5; OWASP unbounded consumption110
Daily spend limit with an alertA busy day running up a billThe same; in our chat assistant, rate limits and a daily cap are part of the service
Concurrency setting reviewed against volumeBurst pricing that doubles the per-minute rate above a limitElevenLabs prints a burst price double the standard rate95
Calendar of printed price stepsSurprises from vendor price changesTable 3; Gemini price steps92
Monthly check of the invoice against contacts handled times kDrift in k, unbilled or mis-billed usageChapter 5
Cost per booked customer against VWhether capture pays at allChapter 5

Table 6. Six cost controls, what each limits and its basis. The controls are ours.

Anthropic prints a 50% discount on input and output tokens for batch processing90, and Google lists a Batch API with a 50% cost reduction92. A live call cannot wait for a batch. The discount could apply to work done after the call, such as the summaries and the monthly reading of transcripts; we have checked no specific call flow, and a vendor's terms decide.

The controls come with a stop rule. Cost per booked customer, the monthly cost divided by the customers the agent added, can be compared with V, the value of one new customer over the first year as gross margin. If the first exceeds the second, the capture mechanism alone is losing money. A rule, ours, is to pause and review when that happens in two successive months, after allowing for the small numbers that make a single month noisy (Chapter 7). The review asks whether the cause is cost, volume or customers added. It also counts what else the fee buys: the staff hours the agent released and, where the plan is bought too, the new clients the plan produced.

Build, buy or hire: AI automation compared with staffing

Three routes are open to a small firm covering its inbound contacts: hire a person, buy an agent as a service, or build one from vendors' parts. The comparison follows the scope rule of Chapter 11: a fee counted as a cost has its product counted as a gain. All three routes answer calls and release or consume hours, so each is set against answered calls and hours. The plan, which is set against new clients, is left out of every route, on both sides of the ledger.

The comparison uses only cited inputs, the Bureau of Labor Statistics wages and employer costs, our published prices and the vendor prices of Chapter 5, plus labeled assumptions. This chapter does not price a developer's, a marketer's or an operations specialist's time, or a live answering service, so those routes are left out.

Suppose the agent handles 300 contacts a month of three minutes each, as in Chapter 5, which is 900 minutes (300 x 3) or 15 hours of call time (900 / 60). Suppose a hire is one half-time seat of 20 hours a week for 52 weeks. That is 20 times 52 divided by 12, or 86.67 hours a month (the table uses the unrounded 86.6667). The Bureau reports median hourly wages of $18.27 for receptionists98 and $21.53 for customer service representatives99, for May 2025. Each is a wage, not the full cost of an employee.

For private industry workers in June 2026, wages and salaries were 70.0% of employer compensation costs and benefits the remaining 30.0%100. Dividing each wage by 0.70 gives $26.10 an hour for a receptionist ($18.27 / 0.70) and $30.76 for a customer service representative ($21.53 / 0.70); these are our derivation, not Bureau figures. The shares are an average across private industry, per hour worked, so the derivation is approximate. For health care the Bureau prints the full cost directly. Office and administrative support occupations in health care and social assistance cost employers $34.59 an hour worked in June 2026, of which wages were $23.50 and benefits $11.0873. The figure covers the whole sector, not physician or dental offices alone. Equipment, space, recruiting, training and supervision remain outside every figure.

Suppose the buy route is our voice agent at its published price, $299 a month with 250 call minutes included and $0.39 a minute after that18. A dental or medical practice that uses the agent on patient calls also needs a HIPAA-ready program around it56. The program is a compliance layer and does not include the agent56. The site assistant, appointment forms, texts to patients and call tracking need it as well, and the front-desk add-ons run inside it at their own prices56, so its fee is shared by everything on it. It costs $824 a month in the first year ($699 + $125, the $1,500 setup over 12 months)56.

The agent's buy route is $552.50 a month for those 900 minutes ($299 + (900 - 250) x $0.39). Inside the 250 included minutes, about 83 three-minute calls (250 / 3), it is $299. Only a practice that buys the program for the agent and runs nothing else on it pays for both: $1,376.50 a month in the first year ($552.50 + $824), or $1,123 inside the included minutes ($299 + $824). Chapter 11's worked cases use that figure, because they stay within about 62 four-minute calls (250 / 4). That is the narrow case. Suppose the build route pays vendors' charges only. Table 7 sets the routes side by side.

RouteCited inputsMonthly cost, wage only (calc.)Monthly cost with benefits (calc.)Left out
Hire a half-time receptionist86.67 hours x $18.2798; benefits at 30.0% of compensation100$1,583.40$2,262.00Equipment, recruiting and supervision; covers 20 hours a week
Hire a half-time customer service representative86.67 hours x $21.5399; the same benefit share$1,865.93$2,665.62The same
Hire a half-time front-office worker in health care86.67 hours x $23.50 in wages and $34.59 in full cost73$2,036.67$2,997.80The same
Buy our voice agent$299 with 250 minutes included, then $0.39 a minute18: $299 + (900 - 250) x $0.39$552.50$552.50The practice's own time
Add HIPAA-ready Essentials (dental, medical): the compliance layer that every patient-facing tool shares$699 + $125 (the $1,500 setup over 12 months)56$824.00$824.00The agent, which it does not include; the other tools that run inside it
Buy the agent and a program bought for it alone (dental, medical; the narrow case)$299 + $699 + $125 (the $1,500 setup over 12 months)18,56 + (900 - 250) x $0.39$1,376.50$1,376.50The same
Build from vendors' parts$1.1597 + 300 x $0.4155$125.80$125.80All time to build, test and maintain; platform plan fees, which this chapter does not price

Table 7. Monthly cost of the routes from cited inputs, with 300 contacts of three minutes a month. The costs are calculated on assumptions and are not a forecast. The benefit columns for the first two rows are our derivation from the private-industry shares. A practice that runs other tools on the program spreads its $824 across them56, and the narrow case is the one in which it runs nothing else.

Reading Table 7 needs the value of the practice's own time. Let w stand, as in the notation, for the value of one hour in its next best use, here an hour of the practice's own time. Hours spent on the cadence, on supervising a build or on managing a hire all carry that value, and three cited rates give it a range depending on who spends the hour. The first is $34.59, the employer's full cost of an hour of office and administrative work in health care73. The second is $159, the median hourly billing rate across all positions at CPA firms in fiscal 2022281. The third is $245.52, what a lawyer collects for an hour billed at the $300 Michigan median109. It applies the realization of 88% and collection of 93% in Clio†'s data75 ($300 x 88% x 93%).

Billing rates are ceilings on w. In Clio†'s data the average lawyer bills 38% of an eight-hour day75, so an hour on the cadence may come from unbilled time and cost the firm less than its rate. This paper does not measure the hours the cadence takes, so Table 8 shows a range of 2, 6 or 10 hours a month at each rate (calc.).

Hours a month of the practice's timeAt $34.59 (office staff, health care, with benefits)At $159 (CPA firm billing rate)At $245.52 (a lawyer's collected hour)
2$69.18$318$491.04
6$207.54$954$1,473.12
10$345.90$1,590$2,455.20

Table 8. Cost of the practice's own time on the monthly cadence, for a range of hours at three cited hourly values (calc.).

At 6 hours of office time a month, the practice's time costs 38% of the $552.50 of the bought agent in Table 7 ($207.54 / $552.50). At a CPA's billing rate and a lawyer's collected rate it costs 1.7 and 2.7 times the agent (calc., Tables 7 and 8). That changes the break-even of Chapter 5, whose Table 6 counts the new clients a year that cover our published prices, because the practice's time belongs in that cost too. Counted in clients, even the dearest column stays small. That table puts the agent on a plan at 1.0 DUI matters a year, at the $3,500 average attorney's fee in Washington's 2020 survey102. With a lawyer's six hours added to the $552.50 of Table 7, the monthly cost is $2,025.62 ($552.50 + $1,473.12), or 6.9 DUI matters a year ($2,025.62 x 12 / $3,500). With a CPA's six hours added, $1,506.50 a month ($552.50 + $954) is $18,078 a year ($1,506.50 x 12), a little more than one advisory client's median revenue of $17,867 a year at a firm that sells advisory services105.

Returns come from operation and not from installation in this sense: the software can cost less than the attention it needs, and who gives the attention decides what it costs. Our own monthly reading of call summaries with the client moves part of that time from the practice to us; the practice's remaining share is a range that a reader should replace with a measured figure.

The first reading compares building with buying. The gap between the buy and build lines in Table 7 is $426.70 a month ($552.50 - $125.80). Building is the cheaper route only if every hour spent building, testing and maintaining the agent, including a share of the original build, averages less than that gap divided by w, which Table 9 gives at each rate. The build line uses the dearest configuration's k. The lean Configuration B would put it at $54.93 (1.15 + 300 x 0.17925), and the gap would be larger.

The second reading compares buying with hiring. Table 9 divides the gap between each half-time hire and the bought agent, on wage alone and with benefits, by w. The result is the hours a month the practice could spend on the bought agent before the hire became cheaper.

ComparisonGap a monthHours at $34.59Hours at $159Hours at $245.52
Build against buy$426.7012.342.681.74
Buy against a half-time receptionist, wage only$1,030.9029.806.484.20
Buy against a half-time receptionist, with benefits$1,709.5049.4210.756.96
Buy against a half-time customer service representative, wage only$1,313.4337.978.265.35
Buy against a half-time customer service representative, with benefits$2,113.1261.0913.298.61

Table 9. Hours a month of the practice's own time at which two routes cost the same (calc., Tables 7 and 8). Even with benefits the hire rows leave out equipment, recruiting and supervision, so they are floors.

For a dental or medical practice the agent is set against answered calls and front-desk hours. The program is shared by everything that runs on it, so no single gain is asked to carry it. The matching hire is a front-office seat. Against the agent itself, the half-time front-office seat of Table 7 costs $1,484.17 a month more on wages and $2,445.30 more with benefits, or 42.91 and 70.69 hours of office time at $34.59 (calc., Table 7: $2,036.67 - $552.50 and $2,997.80 - $552.50, each divided by $34.59). Against the agent with a program bought for it alone, the narrow case, the gaps are $660.17 and $1,621.30, or 19.09 and 46.87 hours (calc., Table 7). The program's own add-ons for reminders, rebooking, recall and waitlist backfill run on the same layer56, so the work that runs on it is wider than the call minutes.

These comparisons are not like for like. A person does other work between calls and handles the unusual call, and an agent hands off what it cannot do. The agent covers every hour, while a half-time seat covers 20 of the 168 hours in a week. On the supposed volume, 15 hours of call time falls in the 86.67 paid hours, an occupancy of 17.3% (15 / 86.67). To cover the 128 hours outside a 40-hour week with one receptionist at a time would cost $2,338.56 a week (128 x $18.27). At 52 weeks that is $10,133.76 a month on wage alone ($2,338.56 x 52 / 12) and $14,476.80 at the derived $26.10 with benefits (128 x $26.10 x 52 / 12). Nobody staffs the night for a few hours of calls, so we think the real alternative to the agent after hours is voicemail or nothing, and then the relevant comparison is the capture term of the identity, not a wage.

The capacity term gives a ceiling for the case in which the agent does replace staff time. At 300 contacts the call time is at most 15 hours. On wage alone that is worth $274.05 at the receptionist median (15 x $18.27) and $322.95 at the customer service median (15 x $21.53), both below the $552.50 of the bought agent. With benefits it is worth $391.50 and $461.36, both also below it (calc.).

Capacity alone, assuming every call minute is released as paid time, has to cover a fee that rises by $1.17 for each three-minute call past the included minutes (3 x $0.39). On wage alone a three-minute call releases $0.91 at the receptionist median ($18.27 x 3 / 60) and $1.08 at the customer service median ($21.53 x 3 / 60), both less than $1.17, so capacity alone does not pay at any volume. With benefits a call releases $1.305 and $1.538 ($26.10 and $30.76 x 3 / 60), and capacity alone pays above about 1,493 and 548 contacts a month. Each is $201.50 ($299 - 250 x $0.39) divided by the difference between that value and $1.17 (calc.).

At the same ceiling in a health care office, 15 hours at $34.59 is $518.85 (15 x $34.59). That is $33.65 short of the $552.50 of the agent ($552.50 - $518.85), which equals 16.0 hours of office time ($552.50 / $34.59), and well short of the $1,376.50 of the agent with a program bought for it alone, the narrow case. The program is a layer shared by the tools that run on it, and call minutes are not asked to pay for it.

The ceiling is generous, because staff time is released in fragments between calls, not as whole hours. At 300 contacts a month, then, capacity does not cover the bought agent even at the ceiling, and the larger and surer gain is capture, the finding of Chapter 5 seen from the other side.

Building multiplies the vendors that a practice must contract with and track, and for a regulated practice that means a BAA chain and the plan tiers of Table 2 held by the practice itself. Buying places part of that chain with the seller, as it does in our HIPAA-ready programs. Hiring avoids the chain and adds the burdens of employment that even the full cost leaves out. This chapter does not price what an outside team costs to run the cadence for the system as a whole. Our managed marketing add-on is priced at $1,800 a month on every term18, and it covers campaigns, content and media, so we do not set it against the lines above.

What this chapter cannot say

The chapter rests on a mix of cited prices and wages, descriptions of our own practice, and judgment, and the reader should keep them apart. This paper does not measure the hours that operation takes, the incidents a practice should expect or the time a build requires, so every such number above is a range or a supposition.

How often an agent needs a person can be estimated only from a neighbor. At one hospital an AI voice assistant making outbound calls before a procedure completed 86.4% of them17. We estimate that roughly one call in seven to an inbound agent will need a person or a second attempt (100 - 86.4 = 13.6; likely low, since inbound calls are more varied than a scripted call before a procedure). A randomized or field study of an AI agent answering inbound calls in dental, medical, legal or accounting settings is NF; the nearest evidence is the outbound hospital study above17. This paper does not measure how small firms divide the work between owner and outside team, and it compares no cadence with another. A field experiment could test whether the cadence, the change-control rules and the incident routine improve outcomes, which Chapter 14 includes in the research agenda.

The standards cited here are voluntary guidance on risk management in general, and neither sets a cadence for a ten-person firm. OWASP's list puts Prompt Injection first168, though this paper does not read that order as a measure of frequency. Prices move on the dates vendors print and may move on others, and law changes. Chapter 13 turns the procedures of this chapter into a 90-day and a 12-month plan, with the order in which to build them.

Chapter 13

A maturity model and a roadmap

AI strategy for a small business: a maturity model

This chapter gives you a way to place your business and an order in which to work. The model has five stages, from a brochure site and an unanswered phone to a measured system with agents at several points of contact. The stages are our framework for ordering work and make no claim about how businesses progress, and this paper cites no study that followed businesses through them. A business may skip a stage, sit between two, or buy an agent before it has the records, so use the model as a checklist.

The identity is Chapter 4's. The agent has four mechanisms: capture (N x m x a x c x V), conversion (N x (1 - m) x dc x V), capacity (H x w) and retention (B x dr x L). The ledger for the whole system adds a line above them, demand (dN x c x V), where dN is the contacts a month that the plan adds. Against these stand the run cost (k x n), the fixed cost F and the expected cost of error E, which this paper does not price (Chapter 9). Each fee is counted against what it produces: the plan against demand, the agent against answered contacts and hours. We call a term observed when your own records give its value and estimated when it needs a design that compares what happened with what would have happened without the agent (Chapter 7). Any other term is assumed.

We sell the kind of system this paper describes, and our prices appear in this chapter as one reference point. You can do every step yourself or with another provider.

The five stages of AI adoption in a service business

Stage 1: a brochure site and an unanswered phone

The site shows the business and does little else. Calls ring out or reach a voicemail that nobody tracks, forms land in an inbox, and you cannot say where your last ten customers came from. Chapter 3 argued that this missing record is the cheapest missing complement. At this stage you can observe V and F from your books, and N only roughly, from the phone carrier's call log and your inbox. The terms dN, m, a and c are unobserved, so every claim about what a plan or an agent would add is an assumption.

The next step is to count. Record every call, form and message with its time, and ask each new customer how they found you. Measure new customers a month by the source each one names, and the calls received and missed. Nothing in this stage needs an agent, and nothing in it is wasted if you never buy one.

Stage 2: counted

Every inbound contact is logged with its door, source and time, and staff still answer by phone. There is no agent, and the path after a contact is improvised. You can now observe N by door and by source, m from the log of missed and late contacts, the time to first response, V and F. The contacts a plan adds show up as a rise in N by source once the plan starts, though telling the plan from the season needs a design.

Marketers call response time speed to lead. Oldroyd, McElheran and Elkington audited 2,241 US companies with a web-generated test lead58. As a 2011 news report quotes their Harvard Business Review article, the average response time among companies that responded within 30 days was 42 hours288. In the same report, 23% of the companies never responded at all58. These are companies and web leads, not service practices and their phone calls, so the figures show that slow response is common in one setting and do not give your m. The next step is to define a path for each contact. Measure m by hour and day, the time to first response and the share of contacts that end with a recorded outcome.

Stage 3: routed

Each door has a written path: who answers, what they say, what they may promise, how the booking is made, who receives the hand-off and how fast. The customer record shows each booking and its revenue, and you have checked the rules of your profession and state with counsel. This is the process complement of Chapter 1. You can now observe c by source, bookings by source, B from the customer record, and V and L, the customer lifetime value, from billing history over time. H is observable if staff log how they spend time, though w remains an assumption. There is still no agent.

The next step is to choose one door for an agent in a bounded role. Measure c by source, revenue by source and the staff time spent on contact handling over a few weeks. That last record is the baseline against which any capacity claim will be judged.

Stage 4: assisted

One agent answers one door, such as after-hours calls or website chat. It tells every contact that it is an AI assistant, which is our practice. It follows the path of Stage 3, hands off to a named person, and logs each conversation. You read a sample every week, keep an incident log, and sign any agreement your sector requires before the agent handles protected information. A design lets you compare with and without the agent, such as alternating weeks or held-out hours.

You can observe n, k, F, the bookings among agent-handled contacts, and the incidents in the log. The next step is a monthly reconciliation that ends in one of three decisions: extend, change one thing, or switch off. Measure contacts handled, held and booked, hand-offs completed, requests for a person, complaints, and cost per booked customer, which Chapter 5 argued is the headline in place of cost per contact.

Stage 5: measured

Agents sit at several doors, for example phone, chat, reminders and intake, inside one system with one monthly ledger. Prompts, scripts and rules change under control, vendor agreements are filed, and the cadence of Chapter 12 runs every month. You can observe every cost term and every count, and estimate demand and each mechanism against a counterfactual wherever your designs allow it. What stays assumed is what you choose to assume and label, and the probabilities in E, since the measured error rates in Chapter 9 belong to other tools and other tasks.

This stage has no fixed next step except to keep the ledger accurate. Re-run the designs when volume, prices, the law or the vendor change. Measure the full ledger, with the leading and lagging indicators for each mechanism that Chapter 7 sets out. Table 1 shows, for each stage, what you can observe, what needs a design and what stays assumed.

StageObserved from your recordsNeeds a design to estimateStill assumed
1. Brochure site, unanswered phoneV and F; N roughlyNone possible yetdN, m, a, c, dc, H, w, B, dr, L, k, n, E
2. CountedN by door and source, m, response time, V, FNone possible yetdN, a, c, dc, H, w, B, dr, L, k, n, E
3. RoutedThe above, plus c by source, B, L, H if time is loggedNone possible yetdN, a, dc, w, dr, k, n, E
4. AssistedCost terms n, k and F; bookings among agent-handled contacts; incidentsDemand (dN), capture (through a), dc, whether released hours were useddr, the probabilities in E
5. MeasuredEvery cost term and countDemand and all four mechanismsOnly what you choose to assume and label

Table 1. What you can observe at each stage of the maturity model (our framework).

AI implementation order: demand and measurement before automation

The order is to count demand, define the process, record outcomes, automate one door and only then extend. An agent's payoff depends on a flow of demand to act on, a process to sit inside and a ledger to be judged by (Chapter 1), and each earlier stage builds one of those. An agent installed before them has no work to do, no defined way to do it, or no way of being judged.

On failure and the shape of gains (Chapters 1 and 2), a report from MIT Project NANDA found that 95% of organizations were getting zero return3. Its stated method was a multi-method design that combined a review of over 300 initiatives, interviews with 52 organizations and 153 survey responses3. Gartner, in a 2024 prediction (secondary), named poor data quality, inadequate risk controls, escalating costs and unclear business value as reasons for abandonment12. We read unclear value as a failure of measurement and inadequate controls as a failure of process, and none of the four reasons is that the model could not do the task. Chapter 2 points the same way. The controlled studies recorded an outcome for each worker or task, and the one result in which AI-assisted work was worse concerned correctness, which shows only when answers are checked24.

On complements (Chapter 3), organizational investments have a large influence on the value of IT investments, and the benefits are often intangible and hard to measure14. IT use is complementary to new workplace organization15. A model of the Productivity J-Curve can explain the productivity slowdowns often accompanying the advent of general-purpose technologies, as well as the increase in productivity later30. Among 179 large publicly traded firms, adopters of data-driven decision making had output and productivity 5% to 6% above what their other investments and IT use predicted45. These studies concern earlier technology and large firms, and applying them to your agent is an inference. They bear on order in two ways: the complement comes first, and early returns may be low.

On demand and measurement (Chapters 6 and 7), contacts now arrive through several doors, and no single assistant's answer can be assumed to be the answer. Across the 160 questions our study asked about ten kinds of business in all 16 cities, the assistants named 2,414 businesses, of which 84 (3.5%) were named by all three19. Each question was asked once, and answers came through an API, which is a baseline, not what every signed-in user sees19. So count where your contacts come from. Measuring advertising is hard even at scale: in 25 large field experiments the median confidence interval on the return of digital advertising was over 100 percentage points wide48. Observational methods often failed to reproduce the experimental effects49. The baseline you record in Stages 2 and 3 is what lets you run a before-and-after or held-out design at Stage 4.

On process and the cost of error (Chapters 8 to 10), NIST's AI Risk Management Framework describes four functions, govern, map, measure and manage, with governance running through the others166. A tribunal held an airline to what its chatbot told a customer, writing that it makes no difference whether information comes from a static page or a chatbot231. Klarna said in February 2024 that its assistant had held 2.3 million conversations in a month38. An Entrepreneur report (secondary) of May 2025 on its chief executive's remarks to Bloomberg says the company was hiring human customer service staff again111. This paper cites no primary Klarna statement revising the 2024 figures.

The strongest objection is that an unanswered phone loses money now, so waiting to count and to define a process delays the agent. The order does not ask you to leave the phone unanswered. It asks you not to buy a claim of return before you have a record to test it against. Counting missed calls for several weeks needs no agent. If missed calls turn out to be rare, m is small, the capture mechanism is small, and you have saved F. If they are common, you hold the number that justifies the agent. We concede one case: where you have few doors and an obvious path, Stages 2 and 3 can run alongside a first pilot, provided the log starts first.

A 90-day plan

The plan below promises no result, and no study cited in this paper would support such a promise. It produces a record, a baseline, a bounded deployment and a decision. For the proposal step, our AI voice agent is $299 a month with 250 call minutes included and $0.39 a minute after that18, and our plans are $999, $1,999 and $2,999 a month per location18. The two lower plans carry a one-time onboarding fee of $499 and $999, and the top plan a $3,000 setup fee. On a 12-month term paid in full two months are free, which makes the plans $9,990, $19,990 and $29,990 a year, and a 12-month term waives the two onboarding fees and cuts the setup fee to $1,40018. Collect other proposals and put the questions in the last section of this chapter to every provider.

  1. Days 1 to 14. Name one owner of the system inside the business (Chapter 12 describes the role). List every door: phone numbers, forms, chat, email and directory profiles. For your last ten customers, write down how each found you, by asking them. Start a contact log with date, time, door, source, whether it was answered, the response time, the outcome and the revenue. A spreadsheet is enough.
  2. Days 15 to 30. Log every contact and tag every missed call, filling gaps from the carrier's call detail. Fix the cheapest leaks: the voicemail greeting, the routing of forms to a named person, and the hours shown online.
  3. Days 31 to 45. Write the path for each door, with the hand-off person and the time allowed. Check with counsel the rules that apply to you. If you will record calls, they include the consent rules. Washington requires the consent of all participants in a private communication162. California requires the consent of all parties to a confidential communication177. The federal floor is one-party consent161. Note which states your callers are in.
  4. Days 46 to 60. Choose the first door for an agent: the one where the log shows the largest m beside the largest V, and where a failure leaves a record. Write the measurement design and the review date before the agent starts. Obtain written proposals, or decide to build.
  5. Days 61 to 90. Run the agent at that one door in a bounded role, disclosed, with a named person for hand-off. Read a sample of conversations every week. Compare with the baseline using the design. At day 90 hold a review and decide: extend, change one thing, or switch off.

A 12-month AI strategy for a small business

Months 1 to 3 are the 90-day plan. In months 4 to 6, reconcile the ledger each month as Chapter 12 describes, and keep the design running until its answer is clear enough to decide on, using the small-numbers calculation of Chapter 7. Add a second door only after the first door's ledger reconciles. The second door is often the one that serves a different mechanism, such as reminders for retention or intake for capture.

In months 7 to 9, extend or deepen. Review the incident log as the evidence on E, because each wrong answer is a data point on a probability that only your own log can supply. Revisit the vendor agreements and prices. Test the exit: can you export every call, transcript, source and outcome, and can you switch the agent off without losing the record? If the answer to either is no, fix that before you extend.

In months 10 to 12, hold an annual review. Place the business on Table 1 again. Re-run the build, buy or hire comparison of Chapter 12 with current prices. Re-check the law and the prices, since both move. Colorado repealed and reenacted its AI statute in 2026, with most of the new enactment effective January 1, 2027182. One model vendor's price page, read on 2026-11-10, prints a doubling of a model's per-token price on a stated date92. Neither is a forecast of what will change for you.

Signs that an AI deployment is not paying, and what to switch off

We give no universal thresholds, because the line depends on what you pay and what a client is worth to you. You can solve for it: divide the monthly cost of a purchase by what one client, visit or hour brings, and compare the resulting count with your own log.

A dental practice pays $299 a month for the agent, which includes 250 call minutes, about 62 four-minute calls (250 / 4)18. It needs 0.3 bookings a month to cover that if each is a new patient's first year of spending, $887 per person with a dental visit ($299 / $887)55. If each booking is one visit, it needs 1.1 at about $282 ($299 / $282), a rough figure from ADA billings per dentist and a visit count with no named source53,54. The average location in the data of Patient Prism† (dental call analysis) has 36 calls a month that hang up, and of every 100 of its calls 69 connect and 21 book34. Answering every hang-up would then book 11.0 a month if hang-ups resemble the callers who get through (36 x 21 / 69), so the agent must turn 3% to 10% of those hang-ups into conversations to break even ($299 / $887 / 11.0 and $299 / $282 / 11.0).

The practice also needs the HIPAA-ready Essentials program around the agent56, $824 a month in its first year ($699 + $125 of its $1,500 setup)56. The program is a compliance layer that the site assistant, texts to patients and the front-desk add-ons also run inside56, so its fee is shared by everything on it and is not charged to answered calls alone. A practice that buys it for the agent alone and runs nothing else on it pays $1,123 a month ($299 + $824) and needs 12% to 36% of those hang-ups (1.3 / 11.0 and 4.0 / 11.0), which is the narrow case.

A law firm pays $1,298 a month for the lowest plan plus the agent, inside the agent's included minutes18. A $300 median hourly rate109 brings $245.52 an hour after the 88% of work invoiced and the 93% of invoices collected in the data of Clio† (legal practice software)75. The firm needs 5.3 collected hours a month to cover the $1,298 ($1,298 / $245.52).

If your log shows fewer than the count your own prices call for, you are below break-even on first-year value; above it, you have only covered cost. Set your own line before you start, from your baseline, and use the small-numbers calculation to know how long to wait.

  • Demand. The log shows no rise in contacts by source after the plan starts, so dN x c x V is close to zero in your own record. Read the log for the other forms a gain takes before you conclude that. A practice with a full schedule cannot book more visits, but more demand lets it fill canceled slots, choose better-paying work, raise its minimum fee or justify another provider. Each of those leaves a trace in the records.
  • Capture. The log shows m near zero, because staff already answer almost every contact, so there is little for the agent to capture. Or a is low, because most callers the agent handles hang up, ask for a person or do not book. Or contacts the agent captured never book over a period long enough that your design says some should have. In each case N x m x a x c x V is close to zero in your own record.
  • Conversion. The booking rate on contacts that staff already answered falls after launch, or hand-offs reach staff without a summary, so dc is negative. A 2026 arXiv preprint, posted before peer review, reports a randomized field experiment on text chats at a large e-commerce platform51. An agentic system reduced chat duration and had limited effects on retrial rates but substantially lowered ratings for the chats it handled51. Track satisfaction beside speed.
  • Capacity. Hours are reported as released, but payroll, overtime, bookings and billable work are the same as before, so w is zero. Time saved is profit only if it is sold, spent on something that earns, or removes a cost.
  • Retention. The share of customers kept is no different from the held-out group's, so dr is zero.
  • Run cost and fixed cost. The cost of k x n rises faster than the contacts the agent books, through long calls, loops or spikes. OWASP lists unbounded consumption among the risks of language-model applications110. A 2023 report from the Consumer Financial Protection Bureau records complaints from customers stuck in "doom loops" with chatbots, a count of complaints and not a rate165. A rising share of calls in which the caller asks for a person is the same sign from the customer's side. F can also rise when a tier you need is more expensive than the one you started on. In our check, 6 of the 15 plan-gated tools did not publish the price of the qualifying plan20.
  • Error. The incident log shows wrong statements about price, hours, eligibility or scope. The business answers for what its channels say231. Utah's statute says it is no defense to a violation of the consumer-protection statutes its division enforces that generative AI made the statement184. A cluster of wrong answers on one topic is a reason to narrow the scope, not to write a longer script.
  • The ledger itself. You cannot compute it. If you cannot say what the agent earned, you are at Stage 3, not Stage 4, and you are paying F for a claim. Gartner's predicted reasons for abandonment include unclear business value12.

A field experiment at a large telecommunications company found that a voice AI system replacing a menu system temporarily increased machine-service duration and customers' demand for human service, and persistently reduced complaints50. A dip right after launch is therefore not a sign by itself, which is why the review date is set in advance; a flat line at the review date is.

To switch things off, go in this order. First, whatever the agent can do beyond answering and booking, which is the excessive agency that question 7 below defines. Second, the door with the weakest ledger. Third, the agent. Keep the log, the path and the baseline throughout: they are the complements, and they keep their value if the agent goes.

What to ask an AI automation agency or vendor

A good answer is specific, in writing and checkable. Some of these questions apply only to health-care, legal or accounting businesses, and they say so.

  1. What is your measured result, on what kind of business, with what method and with what interval? The FTC's February 2023 business post (archive capture) said AI performance claims would be deceptive if they lack scientific support or hold only for certain users or conditions194. A figure with no design behind it is a claim and not a measured result.
  2. Do you say your product replaces a person, and on what evidence? The FTC alleged in a 2025 complaint that one company advertised that its product could replace human customer service representatives189. That is an allegation. The controlled customer-support study in Chapter 2 measured an assistant that suggests replies to human agents, not an agent that takes the call246.
  3. What does the agent say at the start of a call, and what does it say when asked whether it is a person? Utah requires a supplier to disclose when asked clearly172 and requires a person in a regulated occupation to disclose in a high-risk interaction without being asked172. California's health statute requires a disclaimer on AI-generated clinical communications from physician offices177. Florida's advisory opinion says lawyers' chatbots must carry a disclaimer that they are an AI program and not a lawyer210. Ask which of these the vendor builds for.
  4. Who is responsible for what the agent says? The tribunal in the airline case rejected the argument that the chatbot was a separate legal entity231.
  5. For a health-care business: will you sign a business associate agreement, on which plan and at what price? HIPAA requires the assurances to be documented in a written contract or other written arrangement that meets the rule's terms195. In our check of 28 tools, 9 sign without a special tier, 15 only on certain plans and 4 offer none20.
  6. What do you mean by "certified"? HHS says no standard requires a covered entity to certify its compliance with the Security Rule196. It does not endorse private certifications, which do not absolve covered entities of their obligations under it196.
  7. What can the agent do besides answer, with which permissions? OWASP lists prompt injection first in its 2025 list of risks for language-model applications168 and defines it as user prompts that alter the model's behavior or output in unintended ways170. It defines excessive agency as the vulnerability that enables damaging actions in response to unexpected, ambiguous or manipulated outputs171.
  8. How is usage billed, what limits a spike, and what happens when the underlying model's price changes? Ask for the rate in writing with its date. A calculator's default is not a bill93.
  9. Where are calls recorded and stored, and how is consent captured in the states where my callers are? See the recording statutes cited in the 90-day plan162,177.
  10. Can the agent be switched off for chosen hours or lines so that I can measure the difference, and can I export every call, transcript, source and outcome?
  11. How do you log, review and fix errors, and which framework do you follow? NIST's generative AI profile defines confabulation as confidently stated but erroneous or false content167. If the vendor claims certification to ISO/IEC 42001, ask what the certificate covers: the standard specifies requirements for an AI management system, and it is not a law169.
  12. What do you promise about rankings or AI answers? Google says a page must be indexed and eligible to be shown with a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional technical requirements137. The paper that introduced generative engine optimization reports visibility gains of up to 40% that vary by domain138. In our own study, only 3.5% of the businesses named for the ten kinds asked in every city were named by all three assistants19. A guarantee of placement has no basis in any of this.
  13. Where do your testimonials and reviews come from? The FTC's rule bars writing or selling fake reviews, including AI-generated ones191,289, and bars compensation conditioned on reviews of a particular sentiment290.
  14. For a law or accounting firm: who supervises the tool under our professional rules? ABA Formal Opinion 512 describes additional supervisory obligations when others outside the firm employ AI tools205. The AICPA Code asks for a confidentiality agreement or the client's consent before client information goes to a service provider223. The IRS Office of Professional Responsibility says outsourced AI tools should be vetted222.
  15. What are the exit and refund terms, in writing? The FTC alleged that consumers, many of them small business owners, lose as much as $250,000 and are often left in debt after relying on one company's alleged false promises189. That too is an allegation, and it is a reason to read the contract before the proposal.

Chapter 14

Objections, limits and a research agenda

The thesis, the objections and how this chapter answers them

This chapter states the six strongest objections we know to the paper's thesis and answers each from the evidence, conceding what should be conceded. It then sets out the limits of the evidence and proposes studies that would settle what the evidence leaves open.

The thesis has three parts, and the objections attack different ones. The first is that AI is a general-purpose input whose payoff depends on complements. Its support is the economics of earlier information technology. The cited evidence suggests that organizational investments had a large influence on the value of IT investments and that IT use was complementary to new workplace organization14,15. The second is our hypothesis that, for a small service business, three complements matter most: a flow of demand for the AI to act on, a process for it to sit inside, and measurement that can tell what it earned. Chapters 2, 3, 6, 7 and 8 argue it, and the evidence is consistent with it without testing it. The third is that a technical marketing system supplies the three, and that an agent installed anywhere else in such a business is a cost whose return the firm cannot see. We have a commercial interest in the third part, as Chapter 1 said.

AI cost and capability: the objection that systems will not matter

The first objection says that the thesis describes a passing condition. Models are becoming cheaper and more capable, and when an agent costs almost nothing and handles contacts as well as the best staff, nothing needs to be built around it, justified or measured. The three complements then stop mattering. Part of this is right.

Cost is already low beside value in the published figures. Chapter 5 priced a three-minute call at between $0.18 and $0.42 in the three configurations it assembled from published prices. Usage is not charged separately until the included minutes run out: the agent is $299 a month with 250 call minutes included, about 62 four-minute calls (250 / 4), and $0.39 a minute after that18. A call past them costs $1.56 at four minutes (4 x $0.39).

The larger lines are fees. The whole health care stack, with the lowest plan at $999 a month18, costs $2,122 a month in its first year. That is $999 plus the agent's $299 plus $824 for the HIPAA-ready Essentials program, which is $699 a month and $125 a month of its $1,500 setup56. The program is the compliance layer that every tool handling patient information runs inside, including the agent on patient calls, the site assistant, texts to patients and the front-desk add-ons, and it does not include the agent56. The $2,122 is 6.3% to 6.7% of the monthly payments for the kept visits of one primary care physician and 1.6% to 1.7% of four, at $186 a visit80, 51.1 scheduled appointments a week107 and 18.8% to 23% of them missed76,77 (calc.).

A $300 median hourly rate109 brings a lawyer $245.52 after the 88% of work invoiced and the 93% of invoices collected in Clio†'s data75. The $1,298 a month that the lowest plan plus the agent costs is then 5.3 hours of collected work ($1,298 / $245.52). A further fall in k changes little for a firm with a decent client value. It matters for a firm with a low client value and many contacts, and for such firms the objection widens the set for which an agent pays.

Capability also moves quickly. In one study ChatGPT 3.5 passed none of the CPA, CMA, CIA or EA exams, with an average score of 53.1%226, and a later configuration with a newer model, few-shot training and reasoning and acting passed all those tested, averaging 85.1%226. Error rates reported in 2024, such as the 17% to 33% hallucination rate of commercial legal research tools from two providers27, are dated, and we cannot say what they are now.

The first thing that survives the objection is a ceiling. A better agent raises a, the share of otherwise-missed contacts that it answers and holds, toward 1 and no further, and a better closer raises c toward 1 and no further. The capture term then cannot exceed N x m x V, the value of the contacts that would have been missed. If N is small, or m is small because the firm already answers its calls, the ceiling is low however capable the model is. The contacts that the plan adds, the demand line dN x c x V of Chapter 4, do not depend on the model at all. Chapter 6 made the same point from the structure of the identity.

Responsibility is the second. The tribunal in Moffatt v. Air Canada wrote that Air Canada is responsible for all the information on its website, whether it comes from a static page or a chatbot231. Utah's code says it is not a defense to a violation of a statute administered by its Division of Consumer Protection that generative AI made the violative statement184. A better agent lowers the chance of error and, on these two sources, leaves the firm answerable for what remains, so E shrinks and does not reach zero.

The third is the history of the argument. Brynjolfsson and Hitt used data on 527 large US firms over 1987 to 1994. They found that the contributions of computerization were up to 5 times greater over 5- to 7-year differences than over one-year differences, and they link this to large, time-consuming investments in complementary inputs44. The finding concerns large firms and earlier computing, and it does not say what happened as computing cheapened further.

We concede one more point. A capable agent may write much of its own process: read the website, propose the hand-off rules and draft the summaries. That would lower F, and it would take over part of what our own practice does for a client. We expect it, and the evidence neither supports nor contradicts the expectation. If both k and F fell toward zero, the claim that an agent placed outside a system is a cost would weaken to a narrower one, that it earns nothing the firm can see. The narrower claim holds because the owner's attention and the cost of error do not fall with the price of a token (Chapters 9 and 12).

AI agents for business: the objection that the platforms will bundle the agent and the demand

The second objection says the platforms will do all of it. Search engines and assistants already stand between customers and firms. If they add answering and booking, perhaps with a paid lead attached, a firm that buys the bundle needs no system of its own, and the website as a working rep belongs to a period in which the platforms left the last step to the firm.

Half of this is true already. In Pew Research Center's analysis of US adults' browsing data, users who encountered a Google AI summary clicked a traditional search result link in 8% of visits, against 15% of visits without a summary112. In our study, which asked each question once19, the three assistants mostly named different businesses: 79.7% of the businesses were named by only one of them19. The 25 most-cited websites, directories, review platforms, list sites, publishers and a forum drew 29% of all citations19. The GEO authors describe content creators as having little to no control over when and how their content is displayed138. A firm's access to customers already passes through parties it does not control, and if one of them sells answering and booking beside placement, the offer may be attractive.

First, the statements we cite describe the present. Google says that a page must be indexed and eligible to be shown with a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional technical requirements137. Its Business Profile help page says there is no way to request or pay for a better local ranking291. These are Google's descriptions as read, and they say nothing about what Google or anyone else will sell.

Second, a bundle does not remove the need for the third complement. If one party delivers the contact, answers it and reports the result, the firm cannot check the report. The large experiments show why that matters. For eBay's paid search, typical OLS methods gave a return of over 1,400% even with time and geographic controls153, where the experiment found -63%153. In that case a number read straight from observational data lay far from the effect, and the same can happen whoever supplies the data. We think a party that sells the contact and marks its own work has an interest in the mark; we name no platform that does so.

A platform's own count can also be a different measure from the firm's. OpenAI's Business Hub help page says that for a local business a mention counts when ChatGPT shows a result for that location, even if the customer does not open the listing. The insights it describes show mentions and website clicks129. That is a definition, not a misreport.

Third, the firm's own record is what lets it act as a buyer. It logs contacts by source whoever delivers them, keeps the customer record in a system it controls, and asks whether logs can be exported. Chapter 13 gives questions to put to a vendor. We concede the rest. In a world where one party supplies demand, process and measurement, the firm's own system has less to do, and a business like ours would shrink. This paper does not forecast whether that world will arrive, and we have a commercial interest in saying that it will not. The thesis survives in a conditional form: whoever supplies the three complements, the firm must be able to see the ledger, because a buyer needs a record to audit, to compare and to leave.

Marketing attribution for a small firm: the objection that measurement costs more than it returns

The third objection is the one a practical owner raises. Measurement is a fixed cost, and a small firm has little to spread it over. Lewis and Rao found that informative advertising experiments can easily require more than 10 million person-weeks, which makes them costly and potentially infeasible for many firms48. Large experiments with major retailers and brokerages, most reaching millions of customers, found the returns to advertising difficult to measure48. A firm with a few dozen bookings a month should not expect a precise return, and the owner's hours spent trying could earn more elsewhere.

We concede most of this. Chapter 7's worked example hid a supposed effect of two bookings a month inside ordinary swings, and none of its designs gives a precise return to a firm of that size. The evidence on scale points the same way. Brynjolfsson and McElheran report that the key correlates of adoption of data-driven decision making are plant size, the presence of complements such as information technology and educated workers, and firm learning46. Their data are Census Bureau data on manufacturing plants46. That is consistent with measurement costing more, relative to size, for small organizations, though it does not show it. The cited evidence on data-driven decision making and output comes from 179 large publicly traded firms45. Adopters had output and productivity 5% to 6% higher than expected given their other investments and information technology usage45. It is survey data, and we do not read it as showing that measurement caused the difference.

The agent added to a plan the firm already has costs $299 a month, or $3,588 a year (12 x $299), inside its included minutes18. The hours that the cadence of Chapter 12 takes are not measured, so we solve for them. At $299 a month the cadence costs more than the agent once it takes more than 8.6 hours a month of health care front-office time at a loaded $34.59 an hour73 ($299 / $34.59). For a lawyer the threshold is 1.2 hours a month, since an hour brings $245.52 at a $300 median rate after realization and collection75,109 ($299 / $245.52).

Priced in full, the owner's time can cost more than the agent, and whose hours are spent decides whether it does. We think most of those hours are operation, not measurement. Reading calls to find wrong statements and reconciling the carrier's log are jobs a firm needs to do to run an agent safely, whether or not it ever asks what the agent earned. The extra cost of measurement is the source tags, the arm label and a holdout, most of it set once.

The test for whether measurement pays is whether it can change a decision worth more than it costs. Take the decision to keep or drop the agent, which costs $299 a month, or $3,588 a year (calc., above). It is such a decision if the agent earns nothing, because dropping it saves the whole annual cost. It is a larger one if the agent earns something, because a firm that drops a paying agent for lack of evidence forgoes the whole benefit. The program is a separate decision, because the other tools run inside it too, and dropping the agent saves its fee only if nothing else uses the program.

In a dental call chain built from published figures, turning half of a month's hang-ups into conversations brings 5.5 bookings (calc.). At one visit each, about $282 and rough, that is $1,544 a month; at a new patient's first year each, $887, it is $4,865 (calc.). After the agent's cost that leaves $1,245 to $4,566 a month ($1,544 - $299 and $4,865 - $299), or $14,940 to $54,792 a year (12 x $1,245 and 12 x $4,566).

The inputs come from Patient Prism† (dental call analytics): 116 calls a month (11,552,668 calls / 8,280 locations / 12) and a 31% hang-up share34. In its own funnel, 69 of every 100 calls connect and 21 book34. Spending per person with a dental visit was $887 in 202155. About $282 a visit comes from ADA billings per dentist and a visit count whose source is not named53,54. The share turned into conversations is the link nobody has measured, and half is one of the values shown, not a forecast. It shows an asymmetry: the loss from keeping a non-paying agent is bounded by its price, and the loss from dropping a paying one is not.

Proportionate measurement is the answer. Keep the cheap parts always: the record of Table 1 in Chapter 7, the source tags and a monthly reading. Run a holdout once, not continuously. State at the start what result would end the trial and on what date, and when the result is unclear, decide on cost, as the stop rule of Chapter 12 does. "Not known yet" is a legitimate output of measurement, and it is better than a belief. The expensive part is a poor fit for one firm. A vendor or agency with many clients can pool results, with a conflict of interest that applies to us, and open reporting by many firms could do it without one.

AI profitability under competition: the objection that the gains will be competed away

The fourth objection is an old one in economics. If a technology is open to everyone and its payoff is a gain in output or service, competition passes the gain to customers, and the firms that adopted first keep little. On this view the paper's identity describes a transitional rent that will vanish as agents spread.

Some of our evidence fits the objection. In the controlled studies that report who gained, the gains went to the less experienced. Novice and low-skilled support workers improved their productivity by 34%, with minimal impact on experienced and highly skilled workers2. In the writing study, ChatGPT compressed the productivity distribution by benefiting low-ability workers more23. We infer that a tool that narrows the gap between weaker and stronger performers will narrow the gap between firms that use it, if their rivals use it too. The studies are of tasks, not firms. Acemoglu's estimate for the whole economy is modest, no more than a 0.66% increase in total factor productivity over 10 years, and it is a calculation from task-level estimates29.

The capture mechanism is the most exposed. If the customers in a local market are a fixed pool, a contact that one firm captures is a contact that another misses, and this paper does not test that. When every firm answers every contact, m falls toward zero for all of them and the capture term disappears as an advantage. We think expectations also move. A 2013 literature review of the Gaps Model of service quality describes its customer gap as the discrepancy between customers' expectations and their perception of the service70. The same review defines responsiveness, the willingness to help customers and perform the service promptly, as one of the SERVQUAL dimensions70. If a prompt answer becomes the norm, it stops distinguishing a firm, and its absence starts to count against one.

The first thing that resists the objection is that the same tool has paid differently in different organizations. Bloom, Sadun and Van Reenen, combining pan-European firm-level IT data with a management practices survey, found that the US IT-related productivity advantage is primarily due to tougher "people management" practices16. Bloom and Van Reenen found significant differences in management quality within countries, with a long tail of extremely badly managed firms47. Both concern multinational or manufacturing firms, and neither concerns AI. If a tool is open to all and its payoff depends on a complement, competition equalizes the tool and leaves the complement. That is what the thesis predicts: buying the same tool does not copy the complement. The second is that capacity and retention do not depend on beating rivals to a contact. A freed hour is worth what it earns, and a customer who stays is worth the margin whether or not rivals also have agents. This second point rests on reasoning, not data.

We concede that the tool's own contribution may well be competed away, and that if a vendor makes the complements cheap to buy, those could be competed away too. An agent may also become a condition of staying even. If rivals answer every call, a firm without an agent loses contacts it used to keep, and the profit measured against a no-agent counterfactual can stay positive while the firm's position against its rivals does not improve. The identity's counterfactual would then have to be a market in which rivals answer. And our own data are a snapshot taken in 2026. The adoption figures of Chapter 1 cannot be read as a trend, and this paper does not follow cohorts of firms through time.

AI for small business: the objection that the evidence is enterprise evidence

The fifth objection says that the paper argues about small service businesses from studies of large organizations. We think it is largely true. Table 1 lists the settings of the main bodies of evidence, as the cited claims state them.

EvidenceSetting as the claims state itClaim
Customer support with an AI assistant (Brynjolfsson, Li and Raymond)5,179 customer support agents246
Professional writing (Noy and Zhang)Mid-level professional writing tasks23
Consulting tasks (Dell'Acqua and co-authors)Consultants, with tasks chosen inside and outside the frontier24
Software (Peng and co-authors)Developers, one HTTP server task25
Voice AI (Wang and co-authors)A large telecommunications company50
Agentic AI (Alibaba preprint)Taobao platform, customer chats; preprint, not peer reviewed51
Klarna's announcementOne company, 2.3 million conversations in a month, in its own words38
Advertising experiments (Lewis and Rao)Major US retailers and brokerages48
Paid search (Blake, Nosko and Tadelis)eBay153
Advertising at scale (Gordon and co-authors)Facebook49
Data-driven decisions (Brynjolfsson, Hitt and Kim)179 large publicly traded firms45
Computerization (Brynjolfsson and Hitt)527 large US firms, 1987 to 199444
Management practices (Bloom and Van Reenen)732 medium-sized manufacturing firms47
Census Bureau BTOSBusinesses in the survey, December 2025 to May 2026; it counts use and not return1

Table 1. Settings of the main evidence in the paper, as the claims state them.

As far as the cited claims show, no row is a small service business. The nearest are the 732 medium-sized manufacturing firms and the Census counts, and neither measures an agent's return. The first four controlled studies measure an AI used by a person, the two on customer-facing AI ran at a large telecommunications company and on one platform, and the one company announcement reports a flow of contacts on a scale that a ten-person practice does not have.

First, the sizes in the studies are possibilities, and the paper uses none as an expectation for a small firm. Every quantity in the identity that concerns a small firm is taken from a cited source about firms of that kind, solved for as the count or share needed to break even, or marked as an estimate. Second, the mechanism transfers and the magnitude does not. That unanswered contacts are lost, that errors have a price and that a firm which records nothing cannot see what it earned are statements of logic that need no enterprise sample, although their sizes do.

Third, the foundation of the thesis, the literature on complements, is itself enterprise evidence from 179, 527 and 732 firms (Table 1), and applying it to a small business is an extension by argument, as the step from earlier information technology to AI is, which Chapter 1 noted. Fourth, each small firm is its own study, which is why Chapter 7 gives designs a firm can run.

The nearest cited item about small businesses and AI products is an enforcement allegation. In an August 25, 2025 complaint the FTC alleged that consumers, many of them small business owners, lose as much as $250,000 and are often left in debt after relying on Air AI's alleged false promises189. It is an allegation, we did not check its outcome, and it concerns one seller. From this one item, what reaches small firms in the cited evidence appears to be sellers' claims and complaints against them, while the outcome studies are about other firms.

AI profit model: the objection that it can never show a loss

The sixth objection targets the model. A model that counts a fee against the demand it produces, values a client from cited payments and solves for what it does not know can look as if it cannot lose. A model that cannot lose proves nothing. This one can show a loss, in one corner and in the tight cases named below.

The corner is a practice that is not growing and whose intake and scheduling are already sufficiently optimized. There the plan adds few clients, few inquiries are lost at first contact, few visits are missed or canceled late, little staff time is released and few people stay longer, so all the gains are small and the fees remain.

The corner is the exception, not the base case: the cited base rates put a typical practice a long way from it. In Patient Prism†'s 2025 data on dental practices, 31 of every 100 calls hung up before reaching an agent34. In Invoca†'s data (call tracking) on health care, 54% of calls were answered by a person64. In Clio†'s test, 40% of 500 law firms answered the call63. The mean no-show rate across ten main clinics of one Veterans Health Administration medical center was 18.8%76. Nearly two in five dentists told the American Dental Association's Health Policy Institute that they were not busy enough74. A practice that answers nearly every call, loses few visits and wants no more clients would sit in the corner.

The model shows losses short of the corner, and the dental call chain shows them only where the share of hang-ups turned into conversations is low. The chain is set against the agent only, $299 a month18. The plan's product is new demand, which is counted separately, and charging the plan to answered calls would charge a fee with nothing credited. The program is not charged to answered calls either, because its fee is shared by everything that runs on it. If the agent turns 10% of the 36 monthly hang-ups into conversations, it books 1.1 and the gain is $309 to $973 a month, valuing each booking at one visit or at a new patient's first year with the inputs above (calc.). The agent is then ahead at both ends, by $10 to $674 ($309 - $299 and $973 - $299), and at 25% it is ahead by $473 to $2,134 (calc.). To cover its cost it needs 3% to 10% of the hang-ups turned into conversations, which is 0.3 to 1.1 of the 11.0 bookings they would yield (calc.); below those shares it loses.

A practice that buys the program for the agent alone and runs nothing else on it pays $1,123 a month with the agent ($299 + $824), the narrow case. It is then behind by $150 to $814 a month at 10% ($299 + $824 - $973 and $299 + $824 - $309) and from $351 behind to $1,310 ahead at 25%, and it needs 12% to 36% of the hang-ups turned into conversations, which is 1.3 to 4.0 bookings (calc.).

Kept visits, hours and the plan's own demand come on top, and the chain does not subtract the callers who ring back and book, so every figure in it is too high, most of all at the low end, where the callers are existing patients.

Solo primary care is the tightest case, and we do not present it as typical of medicine. The whole health care stack at $2,122 a month needs 7.1 new patients a month on a first-year count (calc. from76,77,80,81,107,108). That is 33.2% more than the 21 to 22 new-patient visits a month now, at 12.5% of kept visits. If patients stay three years, it needs 2.4 a month. The plan alone needs 3.3 new patients a month, which is 15.6% more for one physician and 3.9% more for four.

The agent's own fee is $299 a month: 8.6 hours of front-desk time at a loaded $34.59 an hour73 ($299 / $34.59), about 25 minutes on each of 21 working days (8.6 x 60 / 21), set against the hours it frees and the answered calls of the call chain above. The larger fee is the program. With the reminders add-on running on it at $149 a month56, the two cost $973 a month in the first year ($824 + $149), and covering that takes 5.2 kept visits a month at $186 a visit ($973 / $186). That is 2.4 points off the no-show rate for one physician, toward the top of the 0.5 to 3 points that two studies of reminders found76,86, and 0.6 points for four. A solo office therefore covers the program by running more than one tool on it, since the agent's answered calls, reminders, recall, waitlist backfill and intake share the same layer.

The $186 mean is pulled up by a few costly visits. At the 2016 median primary care visit of $10780, a patient's payments in a year are $173, the plan alone needs 5.8 new patients a month, the program and reminders need 9.1 kept visits a month and the whole stack needs 19.8, against 3.3, 5.2 and 11.4 at the mean. At the 2023 mean for a visit to any doctor, $34782, the counts are 1.8, 2.8 and 6.1 (calc.). Payments are what insurers and patients paid and not what the practice keeps.

A firm that sells only base Form 1040 returns at the $236 average fee282 needs 66 new clients a year to cover the lowest plan plus the agent. A dentist who counts only the average net income share of 22.3% needs 10.7 new patients a month for the whole stack, over a third more than the 30 a month that Henry Schein One† (dental practice software) reports for its average practice13 (both counts calc. from53).

A full schedule is not the corner. More demand still lets a full practice fill canceled slots, choose better-paying work, raise its minimum fee or justify another provider, and the median accounting firm's fees grew 6.7% in a year274. Whether a given practice sits near the corner is a question its own records answer.

Limits of this paper

This paper reports no outcome data from agents we have deployed and no result from any client. It describes how our service works and what it costs. We sell the kind of system the thesis favors, and a reader should weigh the argument with that in mind.

What a second round of research found

A second round looked for figures the first round lacked. On unanswered calls it found the Patient Prism†, Invoca† and Clio† figures given earlier in this chapter, and a Veterans Health Administration study that put mean abandonment at 12.0% in October 201465 and 8.3% by September 201665. These measure different things, and we do not set them side by side. On volume, Call Box† reports 299 calls per office location in February 2020 across more than 1,700 health care practices33, and Henry Schein One† reports 30 new patients a month for its average dental practice13.

On what a client brings, it added the average attorney fee for a DUI in Washington, $3,500 in 2020102. It also added median revenue per advisory client at a CPA firm that sells advisory services, $17,867 in 2024, in a survey of 206 self-selected practices105, beside the primary care and dental payments used above. On missed visits, it added a review of 105 studies with a no-show rate of the order of 23%77. On agents, the one peer-reviewed measurement is outbound: at Mount Sinai an AI voice assistant completed 86.4% of calls in Phase I17. Relatient† (patient scheduling) says its voice product answered 62% of inbound calls, of which 72% were handled or escalated, which is not containment39. Each figure carries the limits its source states, and Chapters 5 and 9 use them.

What was estimated, and what was searched for and not found

Where nothing measures a quantity directly, we estimate it from cited neighbors and say which way the estimate likely errs. For AI referrals to a practice's website, the nearest figures are Conductor†'s (search marketing platform): AI referral traffic was 0.64% of website traffic for health care and 0.48% for financials124,126. The share for a practice is ~1% of website traffic or less (from those two figures; likely low). A recommendation that produces a call and no click is not in a referral count. For AI Overviews on accounting and tax searches, the nearest figure is Conductor†'s 25.8% of all analyzed queries for the Financials industry, which covers banks, insurance and financial services126. We estimate a quarter of such queries; the estimate may err either way, since Google changes when Overviews appear. For unanswered calls at accounting firms, Chapter 10 gives the estimate.

A Google figure for the share of local searches that end in a call, visit or booking is NF; the nearest evidence is a 2015 Google figure on retail store visits292. A repeat-asking study of AI recommendations for local professionals is NF; the nearest evidence is a vendor study of volunteers on unrelated prompts140. Call volume and inquiry-to-client conversion at law firms are NF; the nearest evidence is Clio†'s secret-shopper test. A peer-reviewed overall no-show rate for US dental practices is NF; the nearest evidence is Henry Schein One†'s 4% for its practices of one to seven locations and 3% for its dental groups, with no sample or definition given13,293. An independent survey about AI receptionists in the four professions is NF; the nearest evidence is a Pew result on AI in diagnosis294.

An insurer statement on AI errors in medical, dental or accounting liability cover is NF; the nearest evidence is ALPS on law, as Chapter 9 says. On the market side, firm-level size distributions, an ABA count of solo and small-firm lawyers, an AICPA count of CPA firms and the paid-preparer share of all tax returns are NF; the nearest evidence is the Bureau of Labor Statistics counts of establishments, such as 188,520 offices of lawyers in 2024267.

What only a firm's own records can settle

Several quantities cannot come from a published source for a given firm. The first is the one link in the call chains: the share of unanswered callers that an agent turns into a conversation, which no study has measured for a front desk and which this paper solves for. The second is the firm's own m and N by door and hour, and dN, the contacts a month that the plan adds. The third is the mix of clients it markets for, which sets V, since the count needed to break even depends on whether the firm sells advisory work or base returns. The fourth is whether released hours are resold or sit idle. The fifth is the firm's own error rate by kind of error, which Chapter 9 solves for. The sixth is whether its disclosure changes what callers do. Each is a number in the firm's log, not in a study.

The profit identity is an accounting model: it restates profit as a sum of parts, so it cannot be wrong, though it can be unhelpful. V and L come from cited figures, m from vendor base rates, and the links nobody has measured are solved for. The terms are not independent: an agent that holds more callers (a) may hold callers who fit the firm less well, and so lower c. The identity has no term for referrals, as Chapter 6 noted, or for the effect of the agent on reviews and so on later N. It counts every contact in m as lost, which Chapter 7 illustrated to be an upper bound. Its counterfactual, what would have happened without the agent, is never observed.

Prices and laws change, and each in this paper carries its date. Google prints that Gemini 4 Carbon doubles from $2.00 and $10.00 to $4.00 and $20.00 per million input and output tokens on 2026-12-0192, 21 days after we read the page (calc. from the two dates). Colorado's SB 26-189, a 2026 act, repeals and reenacts, with amendments, part 17 of article 1 of title 6 of the Colorado Revised Statutes, on automated decision-making technology in consequential decisions182. Except for provisions that take effect on passage, the act takes effect January 1, 2027182. The FTC set aside its Rytr order on December 22, 2025188. HHS's proposal to amend the HIPAA Security Rule, published in January 2025, was still a proposal on 2026-11-10, with final action listed for July 2027204. The State Bar of California replaced its November 2023 guidance with a 2026 version211. We read the price pages and the Unified Agenda entry on November 10, 2026. A reader should check each against today's date, and nothing in the paper is legal advice.

The documented failures are a selected sample. The public record of deployments that went wrong, which Chapter 10 reads, is made of cases that became public: a company's own announcements, a tribunal decision, press reports. Successes that stay private and failures nobody wrote up are not in it, so it can illustrate how deployments fail and cannot give a rate. The selection runs both ways. Klarna announced results after its assistant had been live for one month38. Entrepreneur reported that its chief executive later told Bloomberg the company was hiring human customer service workers again111. We can describe that sequence because the company is large and spoke publicly. This paper cites no comparable announcement from a smaller firm.

Our study of assistants' answers matched names by spelling, read up to eight names per answer and used API answers, which are a baseline and not what every signed-in user sees19. The search data come from phrases an owner or a vendor would type, volume shows interest and not purchase, and cost per click is what advertisers pay and not what a customer is worth. Several claims rest on secondary sources, and each is attributed where it is used. The reading in Chapter 2, that gains appear where the task is frequent, the output can be checked and the work is measured, is an interpretation that none of the studies tested, and a literature made of measured tasks will always seem to show that gains need measurement.

A research agenda

The studies below would settle what this paper cannot. As far as our evidence shows, none has been done in the form described. Table 2 lists each with its design and outcome. We state no sample sizes, because a statistician should set the size of each study from a pilot.

StudyQuestionDesignMain outcome
Field experiment with small service firmsDoes an agent's effect depend on demand, process and measurement being present?Many firms; within each, a random slice of hours or lines held out; across firms, support for the system assigned at random in a two-by-two with the agentBookings, net capture, first-year gross margin, cost per booked customer
Open reporting of containment and booking ratesWhat are the real rates, and how do they vary with contact volume?A common record (Table 1 of Chapter 7) and common definitions, reported by firms, vendors and agencies, including those that stoppedContainment, booking per answered contact, booking per missed contact
Audits of AI answers in regulated adviceHow often do agents and assistants give wrong answers within a profession's scope?Paired expert review of answers to realistic questions, repeated over time, scope-limited agents against open-ended onesError rate by kind of error
Access audit of small service firmsHow often do contacts go unanswered, and how fast are they answered?Simulated contacts at randomly drawn hours through each doorm, by trade and hour
Disclosure experimentDoes the wording or timing of the AI disclosure change what callers do?Random assignment of two disclosures; none omittedContinuation, booking, complaints
Calibration of cheap designsHow far do before-and-after and matched-log estimates stray from a randomized holdout?In firms that run a holdout, compute all three estimatesThe gap between estimates
Long-run follow-upDo returns dip before they rise, and are released hours resold?Cohorts of firms followed for yearsMargin, retention, hours redeployed

Table 2. Seven proposed studies. None has been done by us.

First and most important is a field experiment with small service firms. A single firm cannot run it, for the reasons of Chapter 7. With a few dozen bookings a month the noise hides a modest effect. Lewis and Rao's finding that informative advertising experiments can need more than 10 million person-weeks48 describes the same difficulty at a larger scale. Pooling many firms is the way through, with the two-by-two design in Table 2, in which support for the system means source tags, a defined process and the monthly record. The cell that tests the thesis is the interaction: whether the agent's effect is larger where the complements are present. Chapter 2's reading predicts that it is, and Chapter 6 predicts that agents in firms with a thin measured flow of contacts show less return than those in firms with a steady one.

Second comes open reporting of containment and booking rates. Containment is the share of contacts finished without a person, a measure that Chapter 8 declined to treat as a goal. Without a common record and common definitions, every reported rate has its own denominator, and no firm can compare itself with another. Reports would be aggregate and carry no personal data. Selective reporting is the main threat, which is why reports from firms that stopped using an agent matter most. This study would supply the benchmarks that this paper lacks.

Third is an audit of AI answers in regulated advice. In a preprint, Dahl and co-authors report legal hallucinations occurring between 58% of the time with ChatGPT 4 and 88% with Llama 2 when the models were asked specific, verifiable questions about random federal court cases26. Magesh and co-authors found 17% to 33% for the legal research tools of two providers27. Neither, as the cited claims describe them, tested a front-desk agent confined to facts the firm supplies. The nearest test of a bounded agent is the Mount Sinai study of outbound calls, in which a hallucination ended 0.7% of Phase I calls17. An audit would classify the errors by the kinds in Chapter 9 and repeat monthly, since average scores in the accounting exams moved from 53.1% to 85.1% with the model and its configuration226. It would test our expectation in Chapter 9 that scope limits change the problem.

Fourth is an access audit. The vendors and the one secret-shopper study cited above each measured how often calls go unanswered in their own setting and with their own definition34,63,64. None draws small firms at random across the four professions, none covers accounting firms, and the capture term depends on m for the firm in question.

The design has precedent. Oldroyd, McElheran and Elkington, as quoted from their Harvard Business Review article in a 2011 news report, audited 2,241 US companies on the time each took to respond to a web-generated test lead58. A simulated-patient telephone study of primary care practices in 10 states measured whether callers with different insurance received an appointment, which shows that audits with simulated callers can be run by telephone60. Applied to small service firms, such an audit would give m a random-sample basis by trade and hour.

A disclosure experiment would address a gap that Chapter 9 left open. In a December 2023 Gartner survey of 5,728 customers, 64% said they would prefer that companies did not use AI in customer service and 53% said they would consider switching to a competitor if they found out a company was going to use it67. Those are stated preferences, and a study of behavior would compare two compliant disclosures and never test a call with none. A calibration study would run the cheap designs of Chapter 7 beside a randomized holdout, in the manner of Gordon and co-authors' comparison of observational and experimental results49. A long-run follow-up would test the J-curve argument that unmeasured complementary intangible investments lead to initial slowdowns followed by later surges30.

These studies could contradict the thesis. It would be in doubt if agents in firms with thin or unrecorded contact flows earned as much as agents in firms with steady, recorded ones, if firms without a record could report their return as accurately as firms with one, if the net capture found by holdouts were close to zero in firms with steady after-hours demand, or if the cost of error proved too small to matter for the controls of Chapter 9.

Conclusion

Chapter 1 put the puzzle in the form the evidence supports: AI use is spreading fast, documented returns are scarce, and the instruments that count use count different things. The evidence points to one cause more than to any other we weighed: AI is a general-purpose input, and what it earns depends on what surrounds it. In a small service business three things matter most. Demand is a flow of contacts for the agent to act on. Process is a defined path for each contact, ending with a named person. Measurement is a record that links each contact to its source, its handling and its outcome.

The evidence fits that account and does not prove it. We think the controlled studies show that gains appeared where tasks were frequent, outputs could be checked and work was measured, and that in one study harm appeared on a task chosen to lie outside the frontier.

On the cited client values, the cost arithmetic shows that the system's cost is covered by a small count of new clients, kept visits or hours a month. The lowest plan plus the agent, $1,298 a month or $15,576 a year, equals 0.9 new advisory clients a year at a CPA firm that sells advisory services ($15,576 / $17,867105), which is the most favorable case for a firm that does mostly tax returns, and 5.3 collected hours a month at a law firm (calc., above). The whole health care stack, $2,122 a month, equals 2.4 new dental patients a month on first-year spending ($2,122 / $88755), and solo primary care is the tightest case at 7.1 new patients a month. The number of clients the system adds therefore decides profit, and a firm cannot look that number up. It has to measure it, with designs such as those of Chapter 7.

What follows for an owner can be done without us, and it is conditional. Count the contacts and record where they come from. Define what happens to each one, ending with a named person. Record what each became. Then place an agent where there are contacts for it to act on, hold out a slice to see what it adds, and set a date on which to keep it or switch it off. Whatever the answer for a given firm, it will come from the firm's own record, not from a vendor's claim or from studies of firms unlike it.

References

Before the paper cited a source, the page was fetched and the passage we rely on was matched against it, except where a reference is marked otherwise. The small numbers in the text point here.

  1. Khatiwoda. Large Firms With at Least 20 Employees Biggest AI Users. U.S. Census Bureau, 2026. census.gov/library/stories/2026/05/ai-use-businesses.html
  2. Brynjolfsson, Li, and Raymond. Generative AI at Work. National Bureau of Economic Research, 2023. doi.org/10.3386/w31161
  3. Challapally, Pease, Raskar, and Chari. The GenAI Divide: State of AI in Business 2025. MIT Project NANDA, 2025. mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.
  4. Thomson Reuters. The AI Adoption Reality Check: Firms with AI Strategies are Twice as Likely to see AI-driven Revenue Growth; Those Without Risk Falling Behind. Thomson Reuters (press release, June 26, 2025), 2025. thomsonreuters.com/en/press-releases/2025/june/the-ai-adoption-reality Secondary source.
  5. Technical Marketing. What service businesses search for: 2026 demand data. Technical Marketing, 2026-11. technicalmarketing.ai/reports/service-business-search-demand/ Our own data.
  6. Breaux and Dinlersoz. Only 3.8% of Businesses Use AI to Produce Goods and Services, Highest Use in Information Sector. U.S. Census Bureau, 2023. census.gov/library/stories/2023/11/businesses-use-ai.html
  7. Dinlersoz et al.. Tracking Firm Use of AI in Real Time. U.S. Census Bureau Center for Economic Studies, 2024. census.gov/hfp/btos/downloads/CES-WP-24-16.pdf
  8. McKinsey & Company. The state of AI: How organizations are rewiring to capture value. McKinsey & Company, 2024. Secondary source.
  9. Chatterji et al.. How People Use ChatGPT. National Bureau of Economic Research, 2025. doi.org/10.3386/w34255
  10. Anthropic Economics Team. Anthropic Economic Index report: Learning curves. Anthropic, 2026. anthropic.com/research/economic-index-march-2026-report
  11. Ryseff et al.. The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: Avoiding the Anti-Patterns of AI. RAND Corporation, 2024. rand.org/pubs/presentations/PTA2680-1.html Held through its record or abstract; passage not matched.
  12. Sallam. Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025. Gartner, 2024. gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30- Secondary source.
  13. Henry Schein One. 2026 Catalyst Index: Dental Practice Benchmarking Report. Henry Schein One, 2026. henryscheinone.com/insights/ebook/2026-catalyst-index/ Secondary source.
  14. Brynjolfsson and Hitt. Beyond Computation: Information Technology, Organizational Transformation and Business Performance. Journal of Economic Perspectives, 2000. doi.org/10.1257/jep.14.4.23
  15. Bresnahan, Brynjolfsson, and Hitt. Information Technology, Workplace Organization and the Demand for Skilled Labor: Firm-Level Evidence. National Bureau of Economic Research, 2002. doi.org/10.3386/w7136
  16. Bloom, Sadun, and Van Reenen. Americans Do IT Better: US Multinationals and the Productivity Miracle. American Economic Review, 2012. doi.org/10.1257/aer.102.1.167
  17. Annapoorna Kini et al.. Conversational artificial intelligence for pre-procedural patient preparation: implementation, validation and patient satisfaction. npj Digital Medicine, 2025. pmc.ncbi.nlm.nih.gov/articles/PMC13710839
  18. Technical Marketing. Pricing. Technical Marketing, 2026-11. technicalmarketing.ai/pricing/ Our own data.
  19. Technical Marketing. Who AI recommends locally: a study of local answers. Technical Marketing, 2026-11. technicalmarketing.ai/reports/ai-local-recommendations/ Our own data.
  20. Technical Marketing. Which software will sign a HIPAA BAA? 28 common tools, checked. Technical Marketing, 2026-11. technicalmarketing.ai/reports/hipaa-baa-availability/ Our own data.
  21. Kung et al.. Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models. PLOS Digital Health, 2023. doi.org/10.1371/journal.pdig.0000198
  22. Ayers et al.. Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum. JAMA Internal Medicine, 2023. doi.org/10.1001/jamainternmed.2023.1838 Held through its record or abstract; passage not matched.
  23. Noy and Zhang. Experimental evidence on the productivity effects of generative artificial intelligence. Science, 2023. doi.org/10.1126/science.adh2586
  24. Dell'Acqua et al.. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School, 2023. mitsloan.mit.edu/sites/default/files/2023-10/SSRN-id4573321.pdf
  25. Peng, Kalliamvakou, Cihon, and Demirer. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv, 2023. arxiv.org/abs/2302.06590
  26. Dahl, Magesh, Suzgun, and Ho. Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models. arXiv, 2024. arxiv.org/abs/2401.01301
  27. Magesh et al.. Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. arXiv, 2024. arxiv.org/abs/2405.20362
  28. Liu, Zhang, and Liang. Evaluating Verifiability in Generative Search Engines. Association for Computational Linguistics, 2023. arxiv.org/abs/2304.09848
  29. Acemoglu. The Simple Macroeconomics of AI. National Bureau of Economic Research, 2024. doi.org/10.3386/w32487
  30. Brynjolfsson, Rock, and Syverson. The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. American Economic Journal: Macroeconomics, 2021. doi.org/10.1257/mac.20180386
  31. Anders Humlum, Emilie Vestergaard. Large Language Models, Small Labor Market Effects (BFI Working Paper 2025-56). Becker Friedman Institute, University of Chicago, 2025-04. bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-3.pdf
  32. Anders Humlum, Emilie Vestergaard. Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI (NBER Working Paper 33777; previously 'Large Language Models, Small Labor Market Effects'). National Bureau of Economic Research, 2026-03. nber.org/papers/w33777
  33. Call Box. The Top 5 Phone Trends of 2020 and How Healthcare Practices Are Taking Action: Data Study. Call Box, 2020-02. assets-cdn.carwars.com/websites/callbox/resources/dental/CB+Healthcare Secondary source.
  34. Patient Prism. The Dental Patient Access Report. Patient Prism, 2025. patientprism.com/report/dental-patient-access-report/ Secondary source.
  35. American Medical Association, Center for Digital Health and AI. 2026 Physician Survey on Augmented Intelligence. American Medical Association, 2026. ama-assn.org/system/files/physician-ai-sentiment-report.pdf
  36. MGMA. Automatic for the people: AI for front office access. Medical Group Management Association, 2026. mgma.com/mgma-stat/automatic-for-the-people-ai-for-front-office-access
  37. Intuit. Intuit's 2026 Accountant Technology Survey: AI, fragmentation, and the future of the profession. Accounting Today (partner insights, Intuit), 2026. accountingtoday.com/partnerinsights/article/intuit/intuits-2026-accoun Secondary source.
  38. Klarna Communications. Klarna AI assistant handles two-thirds of customer service chats in its first month. Klarna International, 2024. klarna.com/international/press/klarna-ai-assistant-handles-two-thirds- Secondary source.
  39. Relatient. Relatient Voice AI Deflects More Than 60% of Routine Appointment Scheduling Calls. Relatient (via Business Wire), 2025. relatient.com/resources/press-release/relatient-voice-ai-deflects-more Secondary source.
  40. Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning, Daniel E. Ho. Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. arXiv (published in the Journal of Empirical Legal Studies, 2025), 2024. arxiv.org/pdf/2405.20362
  41. Andrew M. Bean et al.. Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nature Medicine, 2026. nature.com/articles/s41591-025-04074-y
  42. Solow. We'd better watch out. New York Times Book Review, 1987. nytimes.com/1987/07/12/books/we-d-better-watch-out.html Held through its record or abstract; passage not matched.
  43. Erik Brynjolfsson, Daniel Rock, Chad Syverson. The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. National Bureau of Economic Research (Working Paper 25148), 2018-10. nber.org/papers/w25148
  44. Erik Brynjolfsson, Lorin M. Hitt. Computing Productivity: Firm-Level Evidence. Review of Economics and Statistics 85(4): 793-808 (abstract via EconPapers/RePEc), 2003. doi.org/10.1162/003465303772815736
  45. Erik Brynjolfsson; Lorin M. Hitt; Heekyung Hellen Kim. Strength in Numbers: How Does Data-Driven Decisionmaking Affect Firm Performance?. International Conference on Information Systems (ICIS) 2011 Proceedings (AIS Electronic Library), 2011. doi.org/10.2139/ssrn.1819486
  46. Erik Brynjolfsson; Kristina McElheran. The Rapid Adoption of Data-Driven Decision-Making. American Economic Review: Papers and Proceedings (American Economic Association), 2005-2010. doi.org/10.1257/aer.p20161016
  47. Nicholas Bloom; John Van Reenen. Measuring and Explaining Management Practices Across Firms and Countries. LSE Centre for Economic Performance Discussion Paper (published in The Quarterly Journal of Economics, 2007), 2007. doi.org/10.1162/qjec.2007.122.4.1351
  48. Randall A. Lewis; Justin M. Rao. The Unfavorable Economics of Measuring the Returns to Advertising. The Quarterly Journal of Economics (Oxford University Press), 2015. doi.org/10.1093/qje/qjv023
  49. Brett R. Gordon; Florian Zettelmeyer; Neha Bhargava; Dan Chapsky. A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook. Marketing Science (INFORMS), 2019. doi.org/10.1287/mksc.2018.1135
  50. Lingli Wang, Ni Huang, Yili Hong, Luning Liu, Xunhua Guo, Guoqing Chen. Voice-based AI in call center customer service: A natural field experiment. Production and Operations Management (Wiley), via Crossref, 2023. doi.org/10.1111/poms.13953
  51. arXiv:2605.14830. Agentic AI and Human-in-the-Loop Interventions: Field Experimental Evidence from Alibaba's Customer Service Operations. arXiv (preprint), 2026-05. arxiv.org/abs/2605.14830
  52. Clio. The Science Behind Smarter Law: Clio's 2025 Legal Trends Report Reveals How Technology Is Rewiring the Way Lawyers Work. Clio, 2025. clio.com/about/press/the-science-behind-smarter-law-clios-2025-legal-t Secondary source.
  53. ADA Health Policy Institute. Dental practice research: Dental practice FAQs. American Dental Association, 2025. ada.org/resources/research/health-policy-institute/dental-practice-res
  54. Statista Research Department. Average number of yearly patient visits per U.S. dentist from 1990 to 2025. Statista, 2025. statista.com/statistics/965365/patient-visits-among-all-dentists-us-pe Secondary source.
  55. R. Henry Olaisen and Richard J. Manski. Dental Utilization and Expenditures, U.S. Civilian Noninstitutionalized Population Aged 2 and Older, 2019-2021 (Statistical Brief #555). Agency for Healthcare Research and Quality, 2021. meps.ahrq.gov/data_files/publications/st555/stat555.pdf
  56. Technical Marketing. HIPAA-ready programs. Technical Marketing, 2026-11. technicalmarketing.ai/hipaa/ Our own data.
  57. James B. Oldroyd; Kristina McElheran; David Elkington. The Short Life of Online Sales Leads. Harvard Business Review, 2011. hbr.org/2011/03/the-short-life-of-online-sales-leads
  58. SmartCompany. Are you letting online leads go cold? (quoting Oldroyd, McElheran and Elkington, The Short Life of Online Sales Leads, Harvard Business Review, March 2011). SmartCompany (Australia), 2011. smartcompany.com.au/marketing/20110704-are-you-letting-online-leads-go Secondary source.
  59. AMN Healthcare (formerly Merritt Hawkins). 2025 Survey of Physician Appointment Wait Times and Medicare and Medicaid Acceptance Rates. AMN Healthcare, 2025. amnhealthcare.com/siteassets/amn-insights/physician/ps-2025-physician-
  60. Karin V. Rhodes; Genevieve M. Kenney; Ari B. Friedman; Brendan Saloner; Charlotte C. Lawson; Daniel Chearo; Douglas Wissoker; Daniel Polsky. Primary care access for new patients on the eve of health care reform. JAMA Internal Medicine, 2012-11/2013-04. doi.org/10.1001/jamainternmed.2014.20
  61. Charles W. LeHew; Yi-Fan Chen; Patrick D. Smith; Brittaney B. Hill; Flavia F. Lamberghini; Alejandra Valencia. Access to restorative oral health care for children living in Illinois with Medicaid vs private dental insurance: An audit study. The Journal of the American Dental Association, 2023. doi.org/10.1016/j.adaj.2023.07.018
  62. Reddy S, Speer M, Saxon M, Ziegler M, Dedolph Z, Qaasim S. Evaluating network adequacy of oral health services for children on Medicaid in Arizona. AIMS Public Health (PMC8755958), 2021. pmc.ncbi.nlm.nih.gov/articles/PMC8755958/
  63. Clio. Highlights from the Legal Trends Report: the Legal Industry in 2024. Clio (a legal practice-management software vendor), 2024. clio.com/blog/highlights-from-2024-legal-trends-report/ Secondary source.
  64. Invoca. The Healthcare Lead Conversion Benchmarks Report 2026. Invoca, 2026. invoca.com/reports/the-invoca-healthcare-lead-conversion-benchmarks-re Secondary source.
  65. Kevin N Griffith et al.. Call Center Performance Affects Patient Perceptions of Access and Satisfaction. American Journal of Managed Care, 2014-10. pmc.ncbi.nlm.nih.gov/articles/PMC8177735/
  66. Clio. AI-powered legal practices surge: Clio's latest Legal Trends Report reveals major shift (press release). Clio (a legal practice-management software vendor), 2024. clio.com/about/press/clio-latest-legal-trends-report/ Secondary source.
  67. Gartner, Inc.. Gartner Survey Finds 64% of Customers Would Prefer That Companies Didn't Use AI For Customer Service. Gartner, Inc., 2023-12. gartner.com/en/newsroom/press-releases/2024-07-09-gartner-survey-finds
  68. A. Parasuraman; Valarie A. Zeithaml; Leonard L. Berry. A Conceptual Model of Service Quality and Its Implications for Future Research. Journal of Marketing (American Marketing Association), 1985. doi.org/10.1177/002224298504900403
  69. A. Parasuraman; Valarie A. Zeithaml; Leonard L. Berry. SERVQUAL: A Multiple-Item Scale for Measuring Customer Perceptions of Service Quality. Marketing Science Institute, 1986. msi.org/working-paper/servqual-a-multipleitem-scale-for-measuring-cust
  70. Aurelio Mauri; Roberta Minazzi; Simonetta Muccio. A Review of Literature on the Gaps Model on Service Quality: A 3-Decades Period: 1985-2013. International Business Research (Canadian Center of Science and Education), 1988. doi.org/10.5539/ibr.v6n12p134 Secondary source.
  71. Ellen Day and Hiram C. Barksdale Jr.. How firms select professional services. Elsevier / Industrial Marketing Management, 1992. doi.org/10.1016/0019-8501(92)90002-b Held through its record or abstract; passage not matched.
  72. Michael R. Darby; Edi Karni. Free Competition and the Optimal Amount of Fraud. The Journal of Law and Economics (University of Chicago Press), 1973. doi.org/10.1086/466756 Held through its record or abstract; passage not matched.
  73. U.S. Bureau of Labor Statistics. Table 4. Employer Costs for Employee Compensation for private industry workers by occupational and industry group [June 2026]. U.S. Bureau of Labor Statistics, 2026-06. bls.gov/news.release/ecec.t04.htm
  74. ADA Health Policy Institute. The State of the U.S. Dental Economy 3rd Quarter 2026 Update. American Dental Association, Q3 2026. ada.org/-/media/project/ada-organization/ada/ada-org/files/resources/r
  75. Clio (Themis Solutions Inc.). 2025 Legal Trends Report. Clio; PDF copy hosted by the Alaska Bar Association, 2025. alaskabar.org/wp-content/uploads/2025-Legal-Trends-Report.pdf Secondary source.
  76. Parviz Kheirkhah, Qianmei Feng, Lauren M Travis, Shahriar Tavakoli-Tabasi, Amir Sharafkhaneh. Prevalence, predictors and economic consequences of no-shows. BMC Health Services Research, 1997-2008. pmc.ncbi.nlm.nih.gov/articles/PMC4714455/
  77. Dantas LF, Fleck JL, Cyrino Oliveira FL, Hamacher S. No-shows in appointment scheduling - a systematic literature review. Health Policy (abstract via Europe PMC), 2018. ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:29482948%20AN
  78. MGMA. Phones are still a backlog costing medical practices time. Medical Group Management Association (MGMA Stat), 2026-03-10. mgma.com/mgma-stat/phones-are-still-a-backlog-costing-medical-practice
  79. National Center for Health Statistics. FastStats - Ambulatory Care Use and Physician office visits. Centers for Disease Control and Prevention, 2019. cdc.gov/nchs/fastats/physician-visits.htm
  80. Steven R. Machlin and Emily M. Mitchell. Statistical Brief #517: Expenses for Office-Based Physician Visits by Specialty and Insurance Type, 2016. Agency for Healthcare Research and Quality (MEPS), 2016. meps.ahrq.gov/data_files/publications/st517/stat517.shtml
  81. James W. Mold, George E. Fryer, A. Michelle Roberts. When Do Older Patients Change Primary Care Physicians?. Journal of the American Board of Family Practice, 1999-2001. jabfm.org/content/17/6/453.long
  82. Agency for Healthcare Research and Quality, Center for Financing, Access, and Cost Trends. MEPS HC-248G: 2023 Office-Based Medical Provider Visits. Agency for Healthcare Research and Quality (MEPS), 2023. meps.ahrq.gov/data_stats/download_data/pufs/h248g/h248gdoc.shtml
  83. Sunil Gupta; Donald R. Lehmann; Jennifer Ames Stuart. Valuing Customers. Journal of Marketing Research (American Marketing Association), 2004. doi.org/10.1509/jmkr.41.1.7.25084
  84. Frederick F. Reichheld; W. Earl Sasser, Jr.. Zero Defections: Quality Comes to Services. Harvard Business Review (excerpt hosted by Bain & Company), 1990. bain.com/insights/zero-defections-quality-comes-to-services-harvard-bu Secondary source.
  85. Reichheld. Zero defections: quality comes to services.. Harvard Business Review, 1990. eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pubmed&id=1010708
  86. Tarabichi Y, Higginbotham J, Riley N, Kaelber DC, Watts B. Reducing Disparities in No Show Rates Using Predictive Model-Driven Live Appointment Reminders for At-Risk Patients: a Randomized Controlled Quality Improvement Initiative. Journal of General Internal Medicine (PMC10150669), 2022. pmc.ncbi.nlm.nih.gov/articles/PMC10150669
  87. Jacqueline Mitchell. A Real-World Study of Patient-Facing AI in Primary Care. Harvard Medical School (news), 2025. hms.harvard.edu/news/real-world-study-patient-facing-ai-primary-care Secondary source.
  88. Clio (Themis Solutions Inc.). 2023 Legal Trends for Solo Law Firms. Clio, 2022. clio.com/wp-content/uploads/2022/10/2023-Legal-Trends-for-Solo-Law-Fir Secondary source.
  89. ADA Health Policy Institute. The State of the U.S. Dental Economy 4th Quarter 2025 Update. American Dental Association, Q4 2025. ada.org/-/media/project/ada-organization/ada/ada-org/files/resources/r
  90. Anthropic. Pricing - Claude API Docs. Anthropic, 2026-11-10. platform.claude.com/docs/en/about-claude/pricing
  91. OpenAI. Pricing - OpenAI API. OpenAI, 2026-11-10. platform.openai.com/docs/pricing
  92. Google. Gemini Developer API pricing. Google AI for Developers, 2026-11-10. ai.google.dev/gemini-api/docs/pricing.md.txt
  93. Retell AI. Pricing - Retell AI. Retell AI, 2026-11-10. retellai.com/pricing
  94. Vapi. Pricing - Vapi. Vapi, 2026-11-10. vapi.ai/pricing
  95. ElevenLabs. ElevenAgents pricing - ElevenLabs. ElevenLabs, 2026-11-10. elevenlabs.io/pricing/agents
  96. Deepgram. Pricing - Deepgram. Deepgram, 2026-11-10. deepgram.com/pricing
  97. Twilio. Voice Pricing - United States - Twilio. Twilio, 2026-10. twilio.com/en-us/voice/pricing/us
  98. U.S. Bureau of Labor Statistics. Receptionists, Occupational Outlook Handbook. U.S. Bureau of Labor Statistics, 2025-05. bls.gov/ooh/office-and-administrative-support/receptionists.htm
  99. U.S. Bureau of Labor Statistics. Customer Service Representatives, Occupational Outlook Handbook. U.S. Bureau of Labor Statistics, 2025-05. bls.gov/ooh/office-and-administrative-support/customer-service-represe
  100. U.S. Bureau of Labor Statistics. Employer Costs for Employee Compensation - June 2026. U.S. Bureau of Labor Statistics, 2026-06. bls.gov/news.release/ecec.nr0.htm
  101. Notable. 24/7 answers: Southwest General's success with HOPE AI voice. Notable Health, undated (page read 2026-11-10). notablehealth.com/customer-stories/southwest-general-health-center-hop Secondary source.
  102. Washington State Office of Public Defense. Private Representation Costs - Misdemeanors. Washington State Office of Public Defense, 2020. opd.wa.gov/sites/default/files/2023-08/00915-2020_MPrivateRates.pdf
  103. Insurance Research Council. Insurance Research Council Study Finds Rising Medical Costs and Attorney Involvement Are Driving Up Auto Injury Claim Costs. Insurance Research Council (The Institutes), 2022. insurance-research.org/news/insurance-research-council-study-finds-ris
  104. State Bar of California, Standing Committee on Professional Responsibility and Conduct. Arbitration Advisory No. 2025-01: Determination of a Reasonable Fee. State Bar of California, 2025. calbar.ca.gov/sites/default/files/portals/0/documents/mfa/2025/2025-01
  105. CPA.com and AICPA Private Companies Practice Section (PCPS). Findings and insights from the 2024 CPA.com & AICPA PCPS client advisory services (CAS) benchmark survey. CPA.com / AICPA, 2024. cpa.com/sites/cpa/files/2024-12/2024-CAS-Benchmark-Survey.pdf
  106. CPA Practice Advisor. Ignition Report Shows Shift in Pricing for Accounting Firms. CPA Practice Advisor (reporting Ignition's benchmark), 2025. cpapracticeadvisor.com/2025/08/21/ignition-report-shows-shift-in-prici Secondary source.
  107. Lisa S Rotenstein, Nate Apathy, Susan Edgman-Levitan, Bruce Landon. Comparison of Work Patterns Between Physicians and Advanced Practice Practitioners in Primary Care and Specialty Practice Settings. JAMA Network Open, 2021. ncbi.nlm.nih.gov/pmc/articles/PMC10265293/
  108. Santo L, Kang K. National Ambulatory Medical Care Survey: 2019 National Summary Tables. National Center for Health Statistics, 2019. stacks.cdc.gov/view/cdc/123251/cdc_123251_DS1.pdf
  109. State Bar of Michigan. 2023 Economics of Law Survey. State Bar of Michigan, 2023. michbar.org/file/pmrc/pdfs/EoL_Report23.pdf
  110. OWASP GenAI Security Project. LLM10:2025 Unbounded Consumption. OWASP Foundation, 2025. genai.owasp.org/llmrisk/llm102025-unbounded-consumption/
  111. Entrepreneur. Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, According to the Company's CEO. Entrepreneur Media, 2025-05. entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-mo Secondary source.
  112. Pew Research Center. Do people click on links in Google AI summaries?. Pew Research Center, 2025. pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to
  113. Fishkin. 2024 Zero-Click Search Study: For every 1,000 EU Google Searches, only 374 clicks go to the Open Web. In the US, it’s 360.. SparkToro, 2024. sparktoro.com/blog/2024-zero-click-search-study-for-every-1000-us-goog Secondary source.
  114. Pew Research Center. Americans and AI 2026: Chatbots, Smart Devices and Views on Impact. Pew Research Center, 2026. pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-sma
  115. KFF. Poll: 1 in 3 Adults Are Turning to AI Chatbots for Health Information, Equaling the Share Who Use Social Media for Health. KFF, 2026. kff.org/health-information-trust/poll-1-in-3-adults-are-turning-to-ai-
  116. BrightLocal Research Team. Local Consumer Review Survey 2025. BrightLocal, 2025. brightlocal.com/research/local-consumer-review-survey-2025 Secondary source.
  117. Rosie Murphy. Local Consumer Review Survey 2026: Star Ratings Keep Rising, Old Reviews Don't Cut It. BrightLocal, 2026. brightlocal.com/research/local-consumer-review-survey/ Secondary source.
  118. Illinois Supreme Court Commission on Professionalism / Clio. Clio Legal Trends Report Finds Law Firms Aren't Responsive to Potential Clients. Illinois Supreme Court Commission on Professionalism, 2019. 2civility.org/clio-legal-trends-report-finds-law-firms-arent-responsiv Secondary source.
  119. Ryan Law. Update: AI Overviews Reduce Clicks by 58%. Ahrefs, 2025. ahrefs.com/blog/ai-overviews-reduce-clicks-update/ Secondary source.
  120. Seer Interactive. AIO Impact on Google CTR: September 2025 Update. Seer Interactive, 2025. seerinteractive.com/insights/aio-impact-on-google-ctr-september-2025-u Secondary source.
  121. Mehrzad Khosravi, Hema Yoganarasimhan. Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia. arXiv (cs.CY preprint, not peer reviewed), 2026. arxiv.org/abs/2602.18455
  122. Liz Reid. AI in Search is driving more queries and higher quality clicks. Google (The Keyword), 2025. blog.google/products/search/ai-search-driving-more-queries-higher-qual
  123. Miriam Ellis. New Research: The Prevalence of AI Overviews in Local Search. Whitespark, 2025. whitespark.ca/blog/case-study-the-prevalence-of-ai-overviews-in-local- Secondary source.
  124. Conductor. Health Care Industry: 2026 AEO / GEO Benchmarks. Conductor, 2025. conductor.com/academy/health-care-aeo-geo-benchmarks Secondary source.
  125. SE Ranking. AI Overviews Research: Analyzing Google's approach to YMYL topics. SE Ranking, 2024. seranking.com/blog/ai-overviews-and-ymyl-topics-research/ Secondary source.
  126. Conductor. Financials Industry: 2026 AEO / GEO Benchmarks. Conductor, 2025. conductor.com/academy/financials-aeo-geo-benchmarks/ Secondary source.
  127. Conductor. The 2026 AEO / GEO Benchmarks Report. Conductor, 2025. conductor.com/academy/aeo-geo-benchmarks-report/ Secondary source.
  128. Microsoft Clarity. AI Traffic Converts at 3x the Rate of Other Channels (Study). Microsoft, 2025. clarity.microsoft.com/blog/ai-traffic-converts-at-3x-the-rate-of-other Secondary source.
  129. OpenAI. Setting up your business on ChatGPT. OpenAI Help Center, 2026. help.openai.com/en/articles/20001635-setting-up-your-business-on-chatg
  130. Vivek Pandya. AI Traffic Drives 83% Higher Revenue Per Visit. Adobe, 2026. business.adobe.com/blog/ai-traffic-higher-revenue-per-visit-retail-hol Secondary source.
  131. Pew Research Center. From Diagnoses to Treatments, Why Americans Use AI Chatbots for Health. Pew Research Center, 2026. pewresearch.org/wp-content/uploads/sites/20/2026/08/PS_2026.8.25_AI-ch
  132. Bain & Company, Global Consumer Lab. Why Consumers Choose an AI Chatbot Over a Search Engine (snap chart). Bain & Company, 2025. bain.com/insights/why-consumers-choose-an-ai-chatbot-over-a-search-eng Secondary source.
  133. FindLaw (Thomson Reuters). The 2024 U.S. Consumer Legal Needs Survey: What attorneys need to know. FindLaw, 2024. findlaw.com/lawyer-marketing/blog/the-2024-u-s-consumer-legal-needs-su Secondary source.
  134. MouthWatch. Dental Practice Marketing in 2026: What 448 Practices Say. MouthWatch, 2026. mouthwatch.com/resources/dental-practice-marketing-2026 Secondary source.
  135. Daniel Hood. Half of firms' clients have been with them 8+ years. Accounting Today (Arizent), 2019. accountingtoday.com/list/half-of-firms-clients-have-been-with-them-8-y Secondary source.
  136. Press Ganey. High standards, high stakes: How healthcare provider reviews drive patient choice. Press Ganey, 2023. pressganey.com/?p=19081 Secondary source.
  137. Google Search Central. AI features and your website. Google, 2025. developers.google.com/search/docs/appearance/ai-features
  138. Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, Ameet Deshpande. GEO: Generative Engine Optimization. arXiv, 2023. arxiv.org/abs/2311.09735
  139. OpenAI. Searching the web with ChatGPT. OpenAI Help Center, 2026. help.openai.com/en/articles/9237897-chatgpt-search
  140. Rand Fishkin. NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility. SparkToro, 2025. sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recom Secondary source.
  141. Anindya Ghose, Avi Goldfarb, and Sang Pil Han. How Is the Mobile Internet Different? Search Costs and Local Activities. Information Systems Research / INFORMS, 2013. doi.org/10.1287/isre.1120.0453
  142. Luca. Reviews, Reputation, and Revenue: The Case of Yelp.Com. Elsevier BV, 2016. doi.org/10.2139/ssrn.1928601
  143. Anderson and Magruder. Learning from the Crowd: Regression Discontinuity Estimates of the Effects of an Online Review Database. Oxford University Press (OUP), 2012. doi.org/10.1111/j.1468-0297.2012.02512.x
  144. Judith A. Chevalier; Dina Mayzlin. The Effect of Word of Mouth on Sales: Online Book Reviews. Journal of Marketing Research (American Marketing Association), 2006. doi.org/10.1509/jmkr.43.3.345
  145. Gao, McCullough, Agarwal, Jha. A Changing Landscape of Physician Quality Reporting: Analysis of Patients’ Online Ratings of Their Physicians Over a 5-Year Period. JMIR Publications Inc., 2012. doi.org/10.2196/jmir.2003
  146. BrightLocal. Local Consumer Review Survey 2023. BrightLocal, 2023. brightlocal.com/research/local-consumer-review-survey-2023/ Secondary source.
  147. Adam E. Block, Medha Reddy and Ella Corkum. The Impact of 1-Star Physician Ratings on New Patient Volume. The American Journal of Managed Care, 2015-01-01 to 2018-04-25. ajmc.com/view/the-impact-of-1-star-physician-ratings-on-new-patient-vo
  148. Michael Luca. Reviews, Reputation, and Revenue: The Case of Yelp.Com. Harvard Business School Working Paper 12-016, 2011. econpapers.repec.org/RePEc:hbs:wpaper:12-016
  149. Sarah Yang (UC Berkeley Media Relations). Yelp reviews boost restaurant business. Berkeley News, University of California, Berkeley, 2012. news.berkeley.edu/2012/09/04/yelp-reviews-boost-restaurant-business/ Secondary source.
  150. MGMA. Document, schedule, communicate: AI tools. Medical Group Management Association, 2025. mgma.com/mgma-stat/document-schedule-communicate-ai-tools
  151. Randall A. Lewis, Justin M. Rao, David H. Reiley. Here, there, and everywhere: correlated online behaviors can lead to overestimates of the effects of advertising. Proceedings of the 20th International World Wide Web Conference (WWW 2011), 2011. gwern.net/doc/economics/advertising/2011-lewis.pdf
  152. Brett R. Gordon, Robert Moakler, Florian Zettelmeyer. Close Enough? A Large-Scale Exploration of Non-Experimental Approaches to Advertising Measurement. Marketing Science, 2023. econpapers.repec.org/RePEc:inm:ormksc:v:42:y:2023:i:4:p:768-793
  153. Thomas Blake; Chris Nosko; Steven Tadelis. Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment. Econometrica (Econometric Society), 2014. doi.org/10.3982/ECTA12423
  154. Ron Berman. Beyond the Last Touch: Attribution in Online Advertising. Marketing Science (INFORMS), 2018. doi.org/10.1287/mksc.2018.1104
  155. PatientNews. Mastering Dental Phone Conversion. PatientNews, undated. patientnews.com/article/mastering-dental-phone-conversion Secondary source.
  156. Jay Geier. A Goldmine of New Patient Opportunity. Inside Dentistry, Dentalaegis (Volume 11, Issue 2, September 2015, pp. 4-5), 2014-2015. dentalaegis.com/id/special-issues/2015/09/a-goldmine-of-new-patient-op Secondary source.
  157. Christi Crane (Ohio State News), on research led by Simon Haeder and Wendy Xu. Having insurance doesn't mean you can get a mental health appointment. The Ohio State University, 2025-10 to 2026-03. news.osu.edu/having-insurance-doesnt-mean-you-can-get-a-mental-health-
  158. 411 Locals. Small Business Owners Don't Answer 62% Of Phone Calls. 411 Locals, 2016. 411locals.us/small-business-owners-dont-answer-62-of-phone-calls/ Secondary source.
  159. HHS Office for Civil Rights. Use of Online Tracking Technologies by HIPAA Covered Entities and Business Associates. U.S. Department of Health and Human Services, 2024-06-26. hhs.gov/hipaa/for-professionals/privacy/guidance/hipaa-online-tracking
  160. United States District Court for the Northern District of Texas. American Hospital Association v. Becerra, No. 4:23-cv-01110-P (N.D. Tex.), Opinion and Order, ECF No. 67. Court order, as hosted by the American Hospital Association, 2024-06-20. aha.org/system/files/media/file/2024/06/opinion-order-in-aha-et-al-v-x
  161. United States Congress. 18 U.S.C. 2511: Interception and disclosure of wire, oral, or electronic communications prohibited. Legal Information Institute, Cornell Law School (US Code text), 2026. law.cornell.edu/uscode/text/18/2511
  162. Washington State Legislature. RCW 9.73.030: Intercepting, recording, or divulging private communication - Consent required - Exceptions. Washington State Legislature, 2026. app.leg.wa.gov/rcw/default.aspx?cite=9.73.030
  163. HHS Office for Civil Rights. Business Associates Frequently Asked Questions (FAQs). U.S. Department of Health and Human Services, 2026. hhs.gov/hipaa/for-professionals/privacy/guidance/business-associates/i
  164. Fin. Facts about Fin. Fin (fin.ai), undated (page read 2026-11-10). fin.ai/ai-instructions Secondary source.
  165. Consumer Financial Protection Bureau. Chatbots in consumer finance. Consumer Financial Protection Bureau, 2023-06-06. consumerfinance.gov/data-research/research-reports/chatbots-in-consume
  166. National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. National Institute of Standards and Technology, 2023-01. nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
  167. National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1. National Institute of Standards and Technology, 2024-07. nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
  168. OWASP GenAI Security Project. OWASP Top 10 for LLM Applications 2025. OWASP Foundation, 2025. genai.owasp.org/llm-top-10/
  169. International Organization for Standardization / IEC. ISO/IEC 42001:2023 - AI management systems. International Organization for Standardization, 2023-12. web.archive.org/web/2025/https://www.iso.org/standard/81230.html
  170. OWASP GenAI Security Project. LLM01:2025 Prompt Injection. OWASP Foundation, 2025. genai.owasp.org/llmrisk/llm01-prompt-injection/
  171. OWASP GenAI Security Project. LLM06:2025 Excessive Agency. OWASP Foundation, 2025. genai.owasp.org/llmrisk/llm062025-excessive-agency/
  172. Utah Legislature. Utah Code 13-77-103 Required disclosures. Utah State Legislature (le.utah.gov), 2025-05-07. le.utah.gov/xcode/Title13/Chapter77/C13-77-S103_2025050720250507.html
  173. Utah Legislature. Utah Code 13-77-104 Safe harbor. Utah State Legislature (le.utah.gov), 2025-05-07. le.utah.gov/xcode/Title13/Chapter77/C13-77-S104_2025050720250507.html
  174. Federal Communications Commission. Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts, Declaratory Ruling, FCC 24-17. Federal Communications Commission, 2024-02-08. docs.fcc.gov/public/attachments/FCC-24-17A1.pdf
  175. Federal Communications Commission. Implications of Artificial Intelligence Technologies on Protecting Consumers From Unwanted Robocalls and Robotexts (notice of proposed rulemaking), 89 Fed. Reg. 73321. Federal Register / U.S. Government Publishing Office, 2024-09-10. govinfo.gov/content/pkg/FR-2024-09-10/html/2024-19028.htm
  176. United States Congress. 47 U.S.C. 227: Restrictions on use of telephone equipment. Legal Information Institute, Cornell Law School (US Code text), 2026. law.cornell.edu/uscode/text/47/227
  177. California Legislature. California Penal Code section 632. California Legislative Information (leginfo.legislature.ca.gov), 2026. leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=PE
  178. Florida Legislature. Florida Statutes 934.03 Interception and disclosure of wire, oral, or electronic communications prohibited. The Florida Legislature (Online Sunshine), 2024. flrules.org/gateway/statute.asp?id=934.03
  179. Maryland General Assembly. Maryland Code, Courts and Judicial Proceedings, section 10-402. Maryland General Assembly, 2026-11-10. mgaleg.maryland.gov/mgawebsite/Laws/StatuteText?article=gcj&section=10
  180. Massachusetts General Court. Massachusetts General Laws, Part IV, Title I, Chapter 272, Section 99. Massachusetts Legislature, 2026-11-10. malegislature.gov/Laws/GeneralLaws/PartIV/TitleI/Chapter272/Section99
  181. Texas Legislature. H.B. No. 149 (enrolled), Texas Responsible Artificial Intelligence Governance Act. Texas Legislature Online (capitol.texas.gov), 2026-01-01. capitol.texas.gov/tlodocs/89R/billtext/html/HB00149F.htm
  182. Colorado General Assembly. Colorado Session Laws 2026, Chapter 131 (SB26-189), Automated Decision-Making Technology. Colorado General Assembly (leg.colorado.gov), 2027-01-01. leg.colorado.gov/laws/session-laws/SB26-189/131/download
  183. Utah Legislature. Utah Code 13-77-101 Definitions. Utah State Legislature (le.utah.gov), 2026-05-06. le.utah.gov/xcode/Title13/Chapter77/C13-77-S101_2026050620260506.html
  184. Utah Legislature. Utah Code 13-77-102 Liability for violation of consumer protection law. Utah State Legislature (le.utah.gov), 2026-05-06. le.utah.gov/xcode/Title13/Chapter77/C13-77-S102_2026050620260506.html
  185. Utah Legislature. S.B. 332 Artificial Intelligence Revisions, Enrolled Copy. Utah State Legislature (le.utah.gov), 2025. le.utah.gov/Session/2025/bills/enrolled/SB0332.pdf
  186. Federal Trade Commission. FTC Announces Crackdown on Deceptive AI Claims and Schemes. Federal Trade Commission, 2024-09-25. ftc.gov/news-events/news/press-releases/2024/09/ftc-announces-crackdow
  187. Federal Trade Commission. FTC Finalizes Order with DoNotPay (FTC press release page). Federal Trade Commission, 2025-01-16. ftc.gov/node/87474
  188. Federal Trade Commission. FTC Reopens and Sets Aside Rytr Final Order in Response to the Trump Administration's AI Action Plan. Federal Trade Commission, 2025-12-22. ftc.gov/news-events/news/press-releases/2025/12/ftc-reopens-sets-aside
  189. Federal Trade Commission. FTC Sues to Stop Air AI from Using Deceptive Claims about Business Growth, Earnings Potential, and Refund Guarantees to Bilk Millions from Small Businesses. Federal Trade Commission, 2025-08-25. ftc.gov/news-events/news/press-releases/2025/08/ftc-sues-stop-air-ai-u
  190. Federal Trade Commission. The Consumer Reviews and Testimonials Rule: Questions and Answers. Federal Trade Commission, 2024-10-21. ftc.gov/business-guidance/resources/consumer-reviews-testimonials-rule
  191. Federal Trade Commission. Federal Trade Commission Announces Final Rule Banning Fake Reviews and Testimonials. Federal Trade Commission, 2024-08-14. ftc.gov/news-events/news/press-releases/2024/08/federal-trade-commissi
  192. Federal Trade Commission. FTC Warns 10 Companies About Possible Violations of the Agency's New Consumer Review Rule. Federal Trade Commission, 2025-12-22. ftc.gov/news-events/news/press-releases/2025/12/ftc-warns-10-companies
  193. Federal Trade Commission. Civil Penalty Inflation Adjustments (Federal Trade Commission notice, document 2026-18853). Federal Register (public inspection copy), 2026. public-inspection.federalregister.gov/2026-18853.pdf
  194. Federal Trade Commission. Keep your AI claims in check. Federal Trade Commission (Business Blog), 2023-02-27. web.archive.org/web/2024/https://www.ftc.gov/business-guidance/blog/20
  195. U.S. Department of Health and Human Services. 45 CFR 164.502 (Code of Federal Regulations, current text as reproduced by Cornell LII). Legal Information Institute, Cornell Law School (copy of the CFR), 2024. law.cornell.edu/cfr/text/45/164.502
  196. HHS Office for Civil Rights. Are we required to “certify” our organization’s compliance with the standards of the Security Rule?. U.S. Department of Health and Human Services, 2013-07-26. hhs.gov/hipaa/for-professionals/faq/are-we-required-to-certify-our-org
  197. Federal Trade Commission. FTC Enforcement Action to Bar GoodRx from Sharing Consumers’ Sensitive Health Info for Advertising. Federal Trade Commission, 2023-02-01. ftc.gov/news-events/news/press-releases/2023/02/ftc-enforcement-action
  198. Federal Trade Commission. FTC to Ban BetterHelp from Revealing Consumers’ Data, Including Sensitive Mental Health Information, to Facebook and Others for Targeted Advertising. Federal Trade Commission, 2023-03-02. ftc.gov/news-events/news/press-releases/2023/03/ftc-ban-betterhelp-rev
  199. Federal Trade Commission. Health Breach Notification Rule (89 FR 47028, final rule), 16 CFR 318.1(a). Federal Register (govinfo.gov), 2024-07-29. govinfo.gov/content/pkg/FR-2024-05-30/pdf/2024-10855.pdf
  200. HHS Office for Civil Rights. Dental Practice Pays $10,000 to Settle Social Media Disclosures of Patients’ Protected Health Information. U.S. Department of Health and Human Services, 2019-10-02. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/elit
  201. HHS Office for Civil Rights and Elite Dental Associates. Resolution Agreement and Corrective Action Plan, Elite Dental Associates (HHS TN 16-240465). U.S. Department of Health and Human Services, 2016-06-04. hhs.gov/sites/default/files/elite-dental-ra-cap.pdf
  202. HHS Office for Civil Rights. Dr. U. Phillip Igbinadolor, D.M.D. & Associates, P.A. (UPI) HIPAA Enforcement Action. U.S. Department of Health and Human Services, 2022-03-28. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/upi/
  203. HHS Office for Civil Rights. HIPAA Security Rule To Strengthen the Cybersecurity of Electronic Protected Health Information (notice of proposed rulemaking, 90 FR 898). Federal Register (govinfo.gov), 2025-01-06. govinfo.gov/content/pkg/FR-2025-01-06/pdf/2024-30983.pdf
  204. HHS Office for Civil Rights. HHS/OCR RIN 0945-AA22: HIPAA Security Rule to Strengthen the Cybersecurity of Electronic Protected Health Information (Unified Agenda entry). Reginfo.gov (OIRA / GSA), 2026-11-10. reginfo.gov/public/do/eAgendaViewRule?pubId=202510&RIN=0945-AA22
  205. ABA Standing Committee on Ethics and Professional Responsibility. ABA Formal Opinion 512: Generative Artificial Intelligence Tools. American Bar Association, 2024-07-29. americanbar.org/content/dam/aba/administrative/professional_responsibi
  206. American Bar Association. Rule 1.6: Confidentiality of Information (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  207. American Bar Association. Rule 5.3: Responsibilities Regarding Nonlawyer Assistance (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  208. American Bar Association. Rule 1.18: Duties to Prospective Client (ABA Model Rules of Professional Conduct). American Bar Association, 2020. americanbar.org/groups/professional_responsibility/publications/model_
  209. Jim Ash. Board of Governors adopts ethics guidelines for generative AI use. The Florida Bar News, 2024-01-19. floridabar.org/the-florida-bar-news/board-of-governors-adopts-ethics-g
  210. The Florida Bar. Florida Bar Ethics Opinion 24-1 (January 19, 2024). The Florida Bar (copy hosted by University at Buffalo School of Law CLE materials), 2024-01-19. law.buffalo.edu/content/dam/law/content/cle/24mar27-materials/fl-bar-e
  211. State Bar of California, Committee on Professional Responsibility and Conduct. Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law (Executive Summary). State Bar of California, 2026-05-14. calbar.ca.gov/Portals/0/documents/ethics/Generative-AI-Practical-Guida
  212. Professional Ethics Committee for the State Bar of Texas. Opinion 705. Texas Center for Legal Ethics, 2025-02. legalethicstexas.com/resources/opinions/opinion-705/
  213. NYSBA Task Force on Artificial Intelligence. Report and Recommendations of the New York State Bar Association Task Force on Artificial Intelligence. New York State Bar Association, 2024-04-06. nysba.org/wp-content/uploads/2022/03/2024-April-Report-and-Recommendat
  214. United States Congress (Internal Revenue Code). 26 U.S. Code 7216 - Disclosure or use of information by preparers of returns. Legal Information Institute, Cornell Law School (U.S. Code text), 2026-11-10. law.cornell.edu/uscode/text/26/7216
  215. United States Congress (Internal Revenue Code). 26 U.S. Code 6713 - Disclosure or use of information by preparers of returns. Legal Information Institute, Cornell Law School (U.S. Code text), 2026-11-10. law.cornell.edu/uscode/text/26/6713
  216. U.S. Department of the Treasury / Internal Revenue Service. 26 CFR 301.7216-3 - Disclosure or use permitted only with the taxpayer's consent. Legal Information Institute, Cornell Law School (e-CFR text), 2026-11-10. law.cornell.edu/cfr/text/26/301.7216-3
  217. U.S. Department of the Treasury / Internal Revenue Service. 26 CFR 301.7216-1 - Penalty for disclosure or use of tax return information. Legal Information Institute, Cornell Law School (e-CFR text), 2026-11-10. law.cornell.edu/cfr/text/26/301.7216-1
  218. Internal Revenue Service. Publication 5708: Creating a Written Information Security Plan for your Tax & Accounting Practice (Rev. 8-2024). Internal Revenue Service, 2024-08. irs.gov/pub/irs-pdf/p5708.pdf
  219. Federal Trade Commission. 16 CFR 314.4 - Elements. Legal Information Institute, Cornell Law School (e-CFR text), 2024. law.cornell.edu/cfr/text/16/314.4
  220. Federal Trade Commission. 16 CFR 314.5 - Effective date. Legal Information Institute, Cornell Law School (e-CFR text), 2024-05-13. law.cornell.edu/cfr/text/16/314.5
  221. U.S. Department of the Treasury. 31 CFR 10.22 - Diligence as to accuracy. Legal Information Institute, Cornell Law School (e-CFR text), 2026-11-10. law.cornell.edu/cfr/text/31/10.22
  222. Internal Revenue Service, Office of Professional Responsibility. Introductory Guidelines for Responsible AI Use in Federal Tax Practice (OPR Alerts, Issue Number 2026-19). Internal Revenue Service (GovDelivery bulletin), 2026-06-24. content.govdelivery.com/accounts/USIRS/bulletins/41d6e70
  223. American Institute of CPAs. AICPA Code of Professional Conduct, 1.700.001 Confidential Client Information Rule. AICPA (copy hosted by the Minnesota Board of Accountancy; text bears © 2026 AICPA), 2026. mn.gov/boa/assets/AICPA%20Code%20of%20Professional%20Conduct_tcm1233-7
  224. Kelly D. Mullins (AICPA Professional Ethics Division). Accounting ethics in the age of AI. Journal of Accountancy (AICPA & CIMA), 2026-08-05. journalofaccountancy.com/news/2026/aug/accounting-ethics-in-the-age-of
  225. Vals AI. VLAIR - Legal Research. Vals AI, 2025. vals.ai/industry-reports/vlair-10-14-25 Secondary source.
  226. Marc Eulerich, Aida Sanatizadeh, Hamid Vakilzadeh, David A. Wood. Is it all hype? ChatGPT's performance and disruptive potential in the accounting and auditing industries. Review of Accounting Studies (Springer), 2024. link.springer.com/10.1007/s11142-024-09833-9
  227. Sina Gogani-Khiabani, Ashutosh Trivedi, ShinPing Chyi, Saeid Tizpaz-Niari. Performance of LLMs on VITA test: potential for AI-assisted tax returns for low income taxpayers. Artificial Intelligence and Law (Springer); record in the NSF Public Access Repository, 2025. par.nsf.gov/biblio/10649839
  228. Utah Legislature. Utah Code 13-77-105 Enforcement. Utah State Legislature (le.utah.gov), 2025-05-07. le.utah.gov/xcode/Title13/Chapter77/C13-77-S105_2025050720250507.html
  229. U.S. Department of Health and Human Services. Annual Civil Monetary Penalties Inflation Adjustment (Federal Register document 2026-01688). Federal Register, 2025 adjustment, published 2026-01-28. public-inspection.federalregister.gov/2026-01688.pdf
  230. U.S. Department of Health and Human Services. Notification of Enforcement Discretion Regarding HIPAA Civil Money Penalties (84 FR 18151). Federal Register (govinfo.gov), 2019-04-30. govinfo.gov/content/pkg/FR-2019-04-30/pdf/2019-08530.pdf
  231. Christopher C. Rivers, Tribunal Member, Civil Resolution Tribunal (British Columbia). Moffatt v. Air Canada, 2024 BCCRT 149. Civil Resolution Tribunal of British Columbia, 2024-02-14. decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do
  232. Judge P. Kevin Castel, U.S. District Court, S.D.N.Y.. Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions (Doc. 54). U.S. District Court for the Southern District of New York (RECAP/CourtListener copy of the filed order), 2023-06-22. storage.courtlistener.com/recap/gov.uscourts.nysd.575368/gov.uscourts.
  233. U.S. District Court for the District of Wyoming. Wadsworth v. Walmart Inc., No. 2:23-CV-118-KHR, Order on Sanctions and Other Disciplinary Action (Doc. 181). U.S. District Court for the District of Wyoming (RECAP/CourtListener copy of the filed order), 2025-02-24. storage.courtlistener.com/recap/gov.uscourts.wyd.64014/gov.uscourts.wy
  234. U.S. District Court for the Northern District of Alabama. Johnson v. Dunn, No. 2:21-cv-1701-AMM, Sanctions Order (Doc. 204). U.S. District Court for the Northern District of Alabama (Docket Alarm copy of the filed order), 2025-07-23. docketalarm.com/cases/Alabama_Northern_District_Court/2--21-cv-01701/J
  235. Mark Bassingthwaighte, Risk Manager. Insurance Coverage Issues for Lawyers in the Era of Generative AI. ALPS Corporation, 2025. alpsinsurance.com/blog/insurance-coverage-issues-for-lawyers-in-the-er
  236. Independent Insurance Agents & Brokers. Verisk to Roll Out New General Liability Exclusions for Generative AI Exposures. Big I (independentagent.com), 2026. independentagent.com/vu_resource/verisk-to-roll-out-new-general-liabil Secondary source.
  237. American Medical Association. Augmented Intelligence Development, Deployment, and Use in Health Care (AMA policy). American Medical Association, 2024-11-12. ama-assn.org/system/files/ama-ai-principles.pdf
  238. CBS News (MoneyWatch). Eating disorder helpline shuts down AI chatbot that gave bad advice. CBS News, 2023-05. cbsnews.com/news/eating-disorder-helpline-chatbot-disabled Secondary source.
  239. Nation's Restaurant News. McDonald's is ending its AI drive-thru test with IBM. Nation's Restaurant News, 2024-06-17. nrn.com/quick-service/mcdonald-s-is-ending-its-ai-drive-thru-test-with Secondary source.
  240. Colin Lecher and Katie Honan, THE CITY. Mamdani to kill the NYC AI chatbot we caught telling businesses to break the law. The Markup, 2026-01-30. themarkup.org/artificial-intelligence/2026/01/30/mamdani-to-kill-the-n Secondary source.
  241. Information Age (Australian Computer Society). CBA reverses AI-driven job cuts, admits error. Australian Computer Society (Information Age), 2025-08. ia.acs.org.au/article/2025/cba-reverses-ai-driven-job-cuts--admits--er Secondary source.
  242. Damien Charlotin. AI Hallucination Cases Database. Damien Charlotin (personal research database), 2026-11-10. damiencharlotin.com/hallucinations/ Secondary source.
  243. U.S. District Court for the Northern District of Texas. Local Civil Rules, N.D. Tex., LR 7.2(f) Disclosure of Use of Generative Artificial Intelligence. U.S. District Court for the Northern District of Texas, 2024-09-03. txnd.uscourts.gov/sites/default/files/documents/CIVRULES.pdf
  244. Chief Judge David C. Godbey, U.S. District Court for the Northern District of Texas. Special Order No. 2-98 (adopting new rules LR 7.2(f) and LCrR 47.2(e)). U.S. District Court for the Northern District of Texas, 2024-09-03. txnd.uscourts.gov/sites/default/files/orders/2-98.pdf
  245. U.S. District Court for the Northern District of Texas. Judge Brantley Starr: Specific Requirements (Certificate regarding Artificial Intelligence). U.S. District Court for the Northern District of Texas, 2026-11-10. txnd.uscourts.gov/judge/judge-brantley-starr/
  246. Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond. Generative AI at Work. National Bureau of Economic Research (Working Paper 31161), 2023. nber.org/papers/w31161
  247. National Center for Health Statistics. FastStats: Oral and Dental Health. Centers for Disease Control and Prevention, 2025. cdc.gov/nchs/fastats/dental.htm
  248. Gurol-Urganci I, de Jongh T, Vodopivec-Jamsek V, Atun R, Car J. Mobile phone messaging reminders for attendance at healthcare appointments. Cochrane Database of Systematic Reviews, 2013. cochrane.org/evidence/CD007458_mobile-phone-messaging-reminders-attend
  249. U.S. Department of Health and Human Services. 45 CFR 164.501 (Code of Federal Regulations, current text as reproduced by Cornell LII). Legal Information Institute, Cornell Law School (copy of the CFR), 2024. law.cornell.edu/cfr/text/45/164.501
  250. U.S. Department of Health and Human Services. 45 CFR 164.508 (Code of Federal Regulations, current text as reproduced by Cornell LII). Legal Information Institute, Cornell Law School (copy of the CFR), 2024. law.cornell.edu/cfr/text/45/164.508
  251. ADA Health Policy Institute. Dentists AI Usage and Attitudes. American Dental Association, 2026. ada.org/resources/research/health-policy-institute/dental-practice-res
  252. HHS Office for Civil Rights. HHS Civil Rights Office Enters Settlement with Dental Practice Over Disclosures of Patients’ Protected Health Information. U.S. Department of Health and Human Services, 2022-12-14. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/new-
  253. American Dental Association. ADA Principles of Ethics and Code of Professional Conduct: Veracity (section 5.F Advertising). American Dental Association, 2026-11-10. ada.org/about/principles/code-of-ethics/veracity
  254. ADA Health Policy Institute. The U.S. Dentist Workforce - 2026 update. American Dental Association, 2025. ada.org/-/media/project/ada-organization/ada/ada-org/files/resources/r
  255. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 6212 Offices of dentists (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/6212.csv
  256. Chris Harrop, MGMA. Operating costs keep climbing for medical practices in 2026 (MGMA Stat). Medical Group Management Association, 2026-06-23. mgma.com/mgma-stat/operating-costs-keep-climbing-for-medical-practices
  257. HHS Office for Civil Rights. Manasa Health Center resolution agreement (HHS HIPAA enforcement page). U.S. Department of Health and Human Services, 2023-06-05. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/mana
  258. HHS Office for Civil Rights and Manasa Health Center. Resolution Agreement and Corrective Action Plan, Manasa Health Center, LLC (HHS TN 20-380176). U.S. Department of Health and Human Services, 2023-03-28. hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements/mana
  259. U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Physicians and Surgeons. U.S. Bureau of Labor Statistics, 2025. bls.gov/ooh/healthcare/physicians-and-surgeons.htm
  260. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 6211 Offices of physicians (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/6211.csv
  261. American Medical Association (Policy Research Perspectives). Physician Practice Characteristics in 2024: Private Practices Account for Less Than Half of Physicians in Most Specialties. American Medical Association, 2024. ama-assn.org/system/files/2024-prp-pp-characteristics.pdf
  262. Clio (Themis Solutions Inc.). Legal Trends Report 2019. Clio; PDF copy hosted by the District of Columbia Bar, 2019. dcbar.org/getmedia/a7bbe4b0-eec7-4fc7-8c85-485853c235ab/Clio-2019-Lega Secondary source.
  263. American Bar Association. Rule 7.1: Communications Concerning a Lawyer's Services (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  264. American Bar Association. Rule 7.3 Solicitation of Clients (ABA Model Rules of Professional Conduct). American Bar Association, 2024. americanbar.org/groups/professional_responsibility/publications/model_
  265. American Bar Association. ABA Profile of the Legal Profession. American Bar Association, 2025. americanbar.org/news/profile-legal-profession/
  266. U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Lawyers. U.S. Bureau of Labor Statistics, 2025. bls.gov/ooh/legal/lawyers.htm
  267. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 54111 Offices of lawyers (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/54111.csv
  268. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter by establishment size class 1, all industries and areas. U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/size/1.csv
  269. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter, NAICS 54111 Offices of lawyers (all areas). U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/industry/54111.csv
  270. Mark Calaguas. 2024 Artificial Intelligence TechReport. American Bar Association, Law Practice Division (ABA TechReport), 2024. americanbar.org/groups/law_practice/resources/tech-report/2024/2024-ar
  271. AICPA. Accounting Firms Say Hiring Outlook Is Strong, AICPA Report Finds (news release on 2025 Trends). AICPA & CIMA, 2023-24. aicpa-cima.com/news/article/accounting-firms-report-strong-hiring-outl
  272. U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Accountants and Auditors. U.S. Bureau of Labor Statistics, 2025. bls.gov/ooh/business-and-financial/accountants-and-auditors.htm
  273. Sweeney, J. T.; Summers, S. L.. The Effect of the Busy Season Workload on Public Accountants' Job Burnout. Behavioral Research in Accounting (American Accounting Association), 2002. doi.org/10.2308/bria.2002.14.1.223
  274. American Institute of CPAs. CPA Firms Report Steady Growth in Revenue and Profit, AICPA Research Finds. AICPA, FY2024. aicpa-cima.com/news/article/cpa-firms-report-steady-growth-in-revenue-
  275. Federal Trade Commission. FTC Safeguards Rule: What Your Business Needs to Know. Federal Trade Commission, 2026-11-10. ftc.gov/business-guidance/resources/ftc-safeguards-rule-what-your-busi
  276. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 541211 Offices of certified public accountants (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/541211.csv
  277. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 541213 Tax preparation services (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/541213.csv
  278. U.S. Bureau of Labor Statistics. QCEW open data: 2024 annual averages, NAICS 541219 Other accounting services (all areas). U.S. Bureau of Labor Statistics, 2024. data.bls.gov/cew/data/api/2024/a/industry/541219.csv
  279. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter, NAICS 541211 Offices of certified public accountants (all areas). U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/industry/541211.csv
  280. Intuit QuickBooks. Accountants embrace AI and strategic advisory services to fuel growth, despite tech and talent challenges, according to 2025 Intuit QuickBooks Survey. Intuit (accounting software vendor), 2025-04. quickbooks.intuit.com/r/news/intuit-quickbooks-tech-survey/ Secondary source.
  281. AICPA and CIMA. U.S. Accounting Firms Show Strong Growth in Profit and Revenue, AICPA and CIMA. AICPA & CIMA, FY2022. aicpa-cima.com/news/article/u-s-accounting-firms-show-strong-growth-in
  282. National Association of Tax Professionals. NATP TAXPRO guide to the 2025 Fee Study (TAXPROPages). National Association of Tax Professionals, 2025. dam.natptax.com/m/39796eb5ee56a576/original/TAXPROPages.pdf
  283. U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Dentists. U.S. Bureau of Labor Statistics, May 2025. bls.gov/ooh/healthcare/dentists.htm
  284. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter by establishment size class 2, all industries and areas. U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/size/2.csv
  285. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter, NAICS 6212 Offices of dentists (all areas). U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/industry/6212.csv
  286. U.S. Bureau of Labor Statistics. QCEW open data: 2024 first quarter, NAICS 6211 Offices of physicians (all areas). U.S. Bureau of Labor Statistics, 2024-Q1. data.bls.gov/cew/data/api/2024/1/industry/6211.csv
  287. AMA Council on Ethical and Judicial Affairs. AMA Code of Medical Ethics, Opinion 9.6.1 Advertising & Publicity. American Medical Association, 2026-11-10. code-medical-ethics.ama-assn.org/ethics-opinions/advertising-publicity
  288. Oldroyd. The Short Life of Online Sales Leads. Harvard Business Review, 2011. thedenmangroupselling.wordpress.com/wp-content/uploads/2011/03/the-sho
  289. Federal Trade Commission. 16 CFR 465.2 Fake or false consumer reviews, consumer testimonials, or celebrity testimonials. Legal Information Institute (copy of the e-CFR text), 2024-10-21. law.cornell.edu/cfr/text/16/465.2
  290. Federal Trade Commission. 16 CFR 465.4 Buying positive or negative consumer reviews. Legal Information Institute (copy of the e-CFR text), 2024-10-21. law.cornell.edu/cfr/text/16/465.4
  291. Google. Tips to improve your local ranking on Google (Google Business Profile Help). Google Business Profile Help, 2026-11. support.google.com/business/answer/7091?hl=en
  292. Matt Lawson (Think with Google). I-Want-to-Go Moments: From Search to Store. Google, 2015. thinkwithgoogle.com/_qs/documents/992/i-want-to-go-micro-moments.pdf
  293. Henry Schein One. What Are Good No-Show and Cancellation Rates for a DSO or Multi-Location Dental Practice?. Henry Schein One, 2026. henryscheinone.com/insights/blogs/dso-dental-practice-no-show-cancella Secondary source.
  294. Alec Tyson, Giancarlo Pasquini, Alison Spencer and Cary Funk. 60% of Americans Would Be Uncomfortable With Provider Relying on AI in Their Own Health Care. Pew Research Center, 2022-12. pewresearch.org/science/2023/02/22/60-of-americans-would-be-uncomforta

Cite this paper

Mike Golliet. "How technical marketing systems make AI profitable." Technical Marketing, November 10, 2026. https://technicalmarketing.ai/papers/how-technical-marketing-systems-make-ai-profitable/

You may quote from this paper with a link to this page. It describes rules and research in general terms and is not legal, medical, tax or financial advice. Found an error? Tell us and we will correct it and note the change here.

Short answers from the paper.

Is AI profitable?

It can be, and the evidence that it is for most firms is thin. Controlled studies show gains on tasks: in one, customer support workers with an AI tool resolved 14% more issues an hour on average2. Evidence about whole organizations is scarcer and less encouraging. An MIT Project NANDA report says 95% of organizations are getting zero return from generative AI3, and Chapter 1 sets out how that figure was produced and what it does not show. The paper's answer is that AI is profitable when it adds customers, frees paid time or keeps customers by more than it costs to run, and that a firm can know this only if it measures.

Full answer

How does a service business make money with AI?

In the paper's model, through four mechanisms: capture (answering contacts the firm now misses), conversion (turning more answered contacts into customers), capacity (freeing paid staff time) and retention (keeping customers longer). Each has conditions under which it earns nothing. An agent that answers calls the front desk was already answering captures nothing, and freed time is worth money only if it is used. The plan that feeds the agent earns through the demand it produces, which the model counts on its own line. Chapter 4 sets out the model, and Chapter 11 works it through for a dental, a medical, a legal and an accounting practice on cited figures.

How much does an AI receptionist cost to run?

Ours is priced at $299 a month with 250 call minutes included, and $0.39 a minute after that18. For what a call costs to run, Chapter 5 builds three configurations from vendors' published prices and finds that a three-minute call costs between about $0.18 and $0.42 to run, computed from those prices and stated assumptions. For comparison, the median hourly wage of a receptionist was $18.27 in May 202598, and a wage is not the full cost of an employee. Cost is seldom the deciding number. What decides profit is how many customers the agent adds.

What is the ROI of AI for a small business?

No study we found measures it for small service firms, and the paper says so. The controlled evidence comes from large employers and platforms. What a small firm can do is compute its own. Chapter 5 gives the break-even on cited client values: the agent added to a plan, at $299 a month with 250 call minutes included, is covered by about 1.0 DUI matters a year at a law firm, or by 0.6 to 1.2 bookkeeping clients a year at an accounting firm. The hard part is showing that the client would not have come anyway, and Chapter 7 shows how to count that with the tools a small office already has.

What should a business have in place before it buys an AI agent?

In the paper's argument, three things: demand for the agent to act on, a process for it to sit inside, and a count of what it did. Chapter 13 gives a maturity scale, a 90-day sequence and a list of questions to put to any vendor. They include what the agent says when asked whether it is a person, who signs a business associate agreement if patient information is involved, and how usage is billed and what limits a spike.

Find out what AI says about you.

A free 20-minute call. We ask the assistants your customers' questions and show you where you come up today.

No pitch unless you ask for one.