Run one hands-on session with made-up client files, then put a one-page policy in writing. Cover which tools are approved, what never goes in, how to check a number or a citation, and who reviews output before it leaves the firm. Repeat the session when you add a tool, and ask staff to bring examples of bad answers.
Why does the IRS care how you train people?
We are not lawyers. Read the rule yourself or ask your counsel how it applies to you.
The IRS Office of Professional Responsibility said in June 2026 that firm procedures under Circular 230 should cover staff training, secure data handling and AI accuracy monitoring, and that outsourced or third-party AI tools should be vetted1. Section 10.22(b) of Circular 230 already presumes diligence when a practitioner relied on another person's work after using reasonable care in engaging, supervising, training and evaluating that person2. The text names a person, not software. A written record of the session and of who attended is the evidence of the training you gave.
What goes in a 90-minute session?
| Segment | Minutes | What staff do |
|---|---|---|
| Policy walk-through | 15 | Read the one-page policy and sort real tasks into allowed and not allowed |
| Catch the error | 25 | Review an AI draft you have seeded with a wrong figure and a citation that does not support its sentence |
| Practice with invented files | 25 | Draft an email, a formula and a procedure, with names and amounts replaced |
| Review and escalation | 15 | Say who approves what, and what to do when unsure |
| Questions and sign-in | 10 | Raise cases from their own desks; everyone signs the policy |
Build the seeded draft yourself from a real assistant's output. The exercise works because staff see for themselves that the text reads fine and is wrong in places.
Why make checking a habit?
The time saving is the easy part to teach. In a randomized experiment on mid-level professional writing tasks, access to ChatGPT cut completion time by 37%, about 10 minutes per task3.
Checking is the hard part. In a 2023 evaluation of generative search engines, only 74.5% of citations supported the sentence they were attached to4. Tell staff to open every cited source and find the passage. A citation that cannot be found, or does not say what the sentence says, goes back to the person who asked for it.
What belongs on the one-page policy?
- The tools staff may use, and for what.
- What never goes into a tool: return information, Social Security numbers, bank details, until counsel clears the tool.
- Who reviews output before it reaches a client or the IRS, by role.
- When a client's written consent is needed.
- Where staff report a wrong answer or a mistaken paste.
Keep it to one page so that people read it. Put the date on it and keep old versions.
How do I keep the training current?
Rerun the session whenever you approve a new tool, since each has different data terms. Hold a thirty-minute review each quarter where staff bring the worst answers they met, and add the best of them to the seeded exercise.
Keep a shared log of wrong answers with the date, the tool and what was wrong. After a few months it shows which tasks the tools handle well enough for less review and which still need a second pair of eyes.
People also ask
Can an AI agent do accounting work for me?
Agents can run short chains of steps, such as requesting a document, but they act with whatever access you give them. Train staff on supervising one before you grant it: narrow task, least access, approval before anything is sent.
What is the best AI for accounting?
No one tool wins every task. Match kind to job, then pick on data terms. Your trainees can help by scoring candidates on the same ten tasks from a closed year with names changed.
Are free AI tools for accounting safe to use?
They are fine for practice with invented data. Training should state plainly that client records stay out of any tool whose terms you have not read, free or paid, until your counsel has cleared it.
Is ChatGPT good at accounting?
It drafts and explains well and can state a tax rule wrongly with confidence. The seeded-error exercise exists for this reason: staff should verify each rule and figure against the IRS or the client's documents.
Is CPA still worth it with AI?
The IRS says AI is a tool, not a substitute for professional judgment, and that final decisions rest with qualified professionals1. The signature and the duty of diligence remain with the practitioner, which is why training time is spent on review.
Sources
- 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
- 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
- Noy and Zhang. Experimental evidence on the productivity effects of generative artificial intelligence. Science, 2023. doi.org/10.1126/science.adh2586
- Liu, Zhang, and Liang. Evaluating Verifiability in Generative Search Engines. Association for Computational Linguistics, 2023. arxiv.org/abs/2304.09848