ChatGPT is good at explaining a concept, drafting a client email and writing a spreadsheet formula. It is unreliable as a source of tax rules or exact figures, because it can state something wrong in a confident tone. Use it to draft and find leads, then confirm every rule and number against the IRS, the standards or client documents.
Where does it help?
Drafting is the clearest case. In a randomized experiment on mid-level professional writing tasks, access to ChatGPT cut completion time by 37%, about 10 minutes against a 27-minute average1. Writing tasks are not accounting tasks, and the study did not test accuracy of tax content. We expect a similar saving on emails, engagement letters and procedure write-ups, where you edit the draft in your own voice.
Formulas are a second case. You can test a formula on rows where you already know the answer, which makes the check cheap.
What do accounting exam studies show?
| Study | Model and setup | Questions | Result |
|---|---|---|---|
| Wood and co-authors (2023) | ChatGPT as of January 2023 | 28,085 accounting assessment questions | 47.5% without partial credit, 56.5% with it; students averaged 76.7%2 |
| Eulerich and co-authors (2024) | ChatGPT 3.5 | CPA, CMA, CIA and EA exam questions | 53.1% average, passed none3 |
| Eulerich and co-authors (2024) | ChatGPT 4 with few-shot training and a calculator | Same exams, multiple choice | 85.1% average, passed all3 |
| Gogani-Khiabani and co-authors (2025) | GPT-4 and Claude 3.5 Haiku; GPT-3.5 | IRS-endorsed VITA certification questions, basic multiple choice | 83.33%; GPT-3.5 reached 54.17%4 |
Two things stand out. Scores rose by about 32 points when the model and setup improved (85.1 - 53.1), so a result from an older version tells you little about today's. And the exams were multiple choice. The Eulerich authors state that passing them does not by itself show AI can do the work of accountants. A client file has missing pages, odd facts and a deadline, and no answer key.
Why do confident wrong answers happen?
NIST calls it confabulation: the production of confidently stated but erroneous or false content5. The wording of a wrong answer is as smooth as the wording of a right one, so tone tells you nothing.
Citations need the same suspicion. In an evaluation of generative search engines, only 74.5% of citations supported their associated sentence6. A source link proves a page exists. It does not prove the page says what the answer claims.
What should I confirm every time?
We are not lawyers. Read the rule yourself or ask your counsel how it applies to you.
Check tax rules, deadlines, thresholds and penalties against the IRS or state source. Tie every figure to a source document. Open every citation and find the sentence. The IRS Office of Professional Responsibility wrote in June 2026 that Circular 230 due diligence requires verifying the accuracy of facts, citations and calculations produced by AI, and that practitioners cannot rely solely on it7.
Keep client details out of the chat unless you have read the terms and your consent rules allow it. The preparer who signs the return answers for it, whatever drafted the first version.
People also ask
Can an AI agent do accounting work for me?
It can take over a few repetitive steps, such as chasing documents or preparing a reconciliation for review, if you give it access and approve its output. It cannot sign, and the diligence stays with you.
What is the best AI for accounting?
There is no single winner, and we have not tested enough products to rank them. Choose the kind of tool for the job, put data terms ahead of features, and run each candidate on a closed year before you trust it.
Are free AI tools for accounting safe to use?
With invented data, yes. With client records, read the terms: a free plan may retain inputs or offer little administrator control. Section 7216 and the AICPA confidentiality rule both turn on consent, so keep records out until you can say where they go.
How do I train my staff to use AI?
Run a hands-on session with made-up files that includes a planted wrong figure and a citation that does not hold up. Add a one-page policy naming approved tools, banned inputs and reviewers, and rerun the session when tools change.
How is AI used in accounting?
Mostly for drafts and shortcuts: client letters, engagement-letter wording, spreadsheet formulas, meeting notes and account-code suggestions that a person approves. Work that decides a tax position or goes out under a signature stays with the practitioner.
Which AI is best for CPAs?
Comparisons on equal tasks are NF. CPAs usually combine a general assistant for drafting, the assistant in their practice software for records, and a publisher's research product for starting points, with every output checked by a person.
How to read this page: NF not found, ( ) our arithmetic on cited figures, and a small number points to a source below.
Sources
- Noy and Zhang. Experimental evidence on the productivity effects of generative artificial intelligence. Science, 2023. doi.org/10.1126/science.adh2586
- David A. Wood and co-authors (Issues in Accounting Education 38(4): 81-108). The ChatGPT Artificial Intelligence Chatbot: How Well Does It Answer Accounting Assessment Questions?. Issues in Accounting Education (American Accounting Association); abstract on the Penn State research portal, 2023-01. pure.psu.edu/en/publications/the-chatgpt-artificial-intelligence-chatb
- 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
- 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
- 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
- Liu, Zhang, and Liang. Evaluating Verifiability in Generative Search Engines. Association for Computational Linguistics, 2023. arxiv.org/abs/2304.09848
- 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