Circular 230 does not forbid a tax practitioner from using AI, but it does not let you rely on the output either. Section 10.22 requires due diligence in preparing and filing, and the IRS's June 2026 alert says practitioners cannot rely solely on AI. The practical answer is a written, per-return review with a named reviewer. This is not legal advice.
Below is that review as a checklist you can copy, with the section behind each item so you can read the primary text yourself. The aim is a record that shows, return by return, what the tool did, what a person verified and who signed. If a regulator ever asks how you used AI on a particular return, the answer should be a page, not a recollection.
What the primary sources actually say
Start with the text, not commentary. Section 10.22(a) says a practitioner must exercise due diligence in preparing or assisting in preparing, approving and filing tax returns and other papers relating to IRS matters, in determining the correctness of representations made to the Department of the Treasury, and in determining the correctness of representations made to clients about IRS matters. Section 10.22(b) adds a presumption: a practitioner is presumed to have exercised due diligence when relying on another person's work product, if the practitioner used reasonable care in engaging, supervising, training and evaluating that person. That paragraph is written about people, and it is subject to sections 10.34 and 10.37.
Section 10.34 sets standards for returns and documents. In summary, a practitioner may not, through willful, reckless or grossly incompetent conduct, sign or prepare a return that takes an unreasonable position, and under 10.34(d) may generally rely in good faith without verification on client-furnished information, but may not ignore the implications of information the practitioner has, and must make reasonable inquiries if information appears incorrect, inconsistent or incomplete. Note what that paragraph covers: information furnished by the client. It does not say a practitioner may rely on a drafting tool.
Then the alert. The IRS Office of Professional Responsibility issued "Introductory Guidelines for Responsible AI Use in Federal Tax Practice" on 24 June 2026 (alert 2026-19). It ties AI use to existing duties: verification of output under 10.22, competence under 10.35, confidentiality under the preparer-penalty provisions in sections 6713 and 7216 and 10.51(a)(15), supervision and firm procedures under 10.36, fees under 10.27(a), and written advice under 10.37. It describes AI-generated text as a draft, and recommends documenting AI use and verification. Read the full alert on irs.gov yourself; this tutorial summarises it and cannot replace it.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThe honest reading: why 10.22(b) does not cover a tool
It is tempting to treat an AI tool as "another person" whose work you may presume reliable under 10.22(b). The text talks about reasonable care in engaging, supervising, training and evaluating a person, which is a duty you cannot discharge for software in the same way. A model is not engaged, trained or evaluated by you in that sense. The safer reading, which matches the alert's emphasis on human scrutiny, is that the diligence duty in 10.22(a) stays with you for everything the tool produces. Ask an adviser how your firm should apply this; do not assume the presumption protects you.
What follows is that the quality of your process, not the quality of the tool, is what you can show. That is why the checklist below is built around evidence of review.
The per-return reliance checklist
Use one sheet per return, kept with the workpapers. Each line is a yes, no or not applicable, with initials and a date. Items that attract a "no" need a note and a fix before filing.
| # | Check | Why (source) |
|---|---|---|
| 1 | Tool, plan and version recorded; plan is a business or enterprise workspace approved by the firm | Competence 10.35; firm procedures 10.36 |
| 2 | What the tool was asked to do is listed (for example "draft explanation of a rental loss limit") | Documentation recommended in the alert |
| 3 | No return information entered unless the client consent and security steps are complete | Sections 6713, 7216; 10.51(a)(15) |
| 4 | Every number in the output traced to a source document or the preparation software | Verification, 10.22(a) |
| 5 | Every authority cited was opened in a primary source and the proposition confirmed | 10.22; 10.34(a) reasonable basis |
| 6 | Inconsistencies in client information queried, not smoothed over by the tool | 10.34(d) reasonable inquiries |
| 7 | Written advice rests on facts you verified, with no unreasonable reliance on AI statements | 10.37 |
| 8 | Fee reflects time and work genuinely done; any saving fairly credited | 10.27(a); alert |
| 9 | Reviewer, other than the person who prompted the tool where staff allow, has read the final text | Supervision, 10.36 |
| 10 | Sign-off block completed (below) | Your record |
Items 4 and 5 are where most of the time goes, and where most of the risk sits. Item 6 deserves a note: a tool that tidies a client's inconsistent figures into a neat schedule has hidden the very signal that 10.34(d) tells you to follow up. Keep the tool away from reconciling conflicting information; give it the query list instead.
The sign-off block
A reviewer's initials are weak evidence. A short block with specific statements is better. Here is an illustrative filled-in version for a small business return:
AI USE RECORD | Return: [client ref] | Tax year: [year]
Tool/plan: business workspace, approved list v3, [date]
Task given to tool: first draft of the cover note explaining
the home-office deduction and the vehicle-use split.
Sources opened to verify: client mileage log (rows 1-212),
prior-year return, primary text of the cited rule.
Figures changed after review: vehicle business-use %
(tool 78%, log shows 71%).
Open queries sent to client: 2 (supplier payment dated
after year end; unexplained transfer of [amount]).
Reviewed by: [name], [date]. I have read the final text
against the sources above and am satisfied it is accurate.
Preparer sign-off: [name], [date].The "figures changed" line is the most valuable. It proves the review found something, and it gives you a measure of how often the tool errs, which feeds your decision on whether to keep using it for that task.
Worked example: a cover note for a small business return
Take an illustrative practice of three preparers. A client runs a one-person design business and has a home office and a vehicle used partly for work. The preparer asks the approved business-plan tool to draft a short cover note explaining the two deductions, giving it only anonymised figures and the firm's own wording template.
The draft states that the home office deduction is "capped at a fixed percentage of gross income" and quotes a specific cents-per-mile rate for the vehicle. Neither statement came from the file. Item 5 sends the preparer to the primary sources: the preparer cannot locate the cap as a rule, and the mileage rate has to be checked against the rate the IRS has published for that tax year. The reviewer deletes the first claim, replaces the second with the verified rate, and records both under "figures changed". Total time added by the review: about fifteen minutes, against roughly twenty-five saved on drafting. The saving is real but small, which is the point: the value of a tool in this work is modest, and the review is what makes it safe.
Four more situations, each with the item that bites
Research memo drafted by AI
The tool cites a ruling by number. Check item 5 first: find the ruling on the IRS site or in a research service and read the passage. If you cannot find it, it does not go in the memo. Court sanctions over fabricated citations are what the alert itself points to.
Client email explaining a notice
Written advice under 10.37 must relate law to facts and not rely unreasonably on representations. Use the tool to simplify your own explanation, never to originate it, and check the dates, amounts and deadlines against the notice itself.
Bank-statement categorisation
Here item 6 is critical: a categorisation that puts a transfer into "sales" because it looks like income may be wrong. Sample thirty lines against the statement, record the error rate, and query anything unexplained with the client.
Engagement-letter wording about AI
A paragraph telling clients that AI-assisted tools may be used in preparing drafts, always with human review, supports transparency. Have a qualified adviser approve the wording, because the alert discusses disclosure and fee credit without a separate rule number.
Firm-level duties: who must set the procedures
Section 10.36 puts a duty on those with principal authority over a firm's practice to take reasonable steps to ensure the firm has adequate procedures for compliance. The alert lists what that can look like: staff training, internal rules for secure data handling and accuracy monitoring, vetting of third-party AI tools, and documentation. For a small practice, that translates into four documents: an approved-tools list, a one-page AI use rule, the per-return sheet above and a quarterly review of errors found.
Two related tutorials cover the data side in depth: Section 7216 consent for AI tools sets out what clients must sign, and adding AI tools to a written information security plan covers the security programme. For due diligence on credits and filing status, Form 8867 diligence with AI notes is relevant. For prompt wording, see ChatGPT prompts for tax preparers, and for season planning, how small tax practices use AI through the busy season.
A one-page firm rule, filled in
The per-return sheet needs a firm rule behind it. Here is an illustrative version for a three-preparer practice. Adapt it with an adviser; do not paste it in unchanged.
AI USE RULE (reviewed quarterly; owner: [name])
1. Approved tools: [list]. Business workspaces only. No personal accounts.
2. No return information is entered until the client's written
consent is on file and the tool is on the security plan.
3. AI output is a draft. Nothing reaches a client, the IRS or a
workpaper as final until a reviewer has verified figures and
authorities against sources.
4. Every use on a return is logged on the AI use record.
5. Any error that reaches a client is reported to the owner the same
day and added to the error log.
6. Fees follow work actually done; savings are reflected in fixed-fee
pricing, not hidden.
7. New tools or new uses need the owner's written approval and a
five-return dry run.Notice that rule 7 builds in the dry run described later. Without it, the approved list tends to grow by habit, and the firm drifts into uses nobody evaluated.
The error log, with illustrative numbers
A review step is only as good as its feedback loop. Keep a simple monthly log of what reviewers caught. Suppose the practice used the tool on 140 returns across a month and the reviewers logged the following, as an illustration:
| Error type | Count | Caught at |
|---|---|---|
| Wrong figure in a cover note | 11 | Item 4, reviewer |
| Authority cited but not found | 4 | Item 5, preparer |
| Condition or limit omitted | 7 | Item 9, reviewer |
| Client inconsistency smoothed over | 2 | Item 6, preparer |
| Error that reached a client | 1 | Client reply |
Twenty-five issues in 140 returns, with one escaping, tells you two things. The checklist works because it caught 24 of 25, and the escape shows where to strengthen it. Perhaps the one that escaped was a deadline in an email, so you add "dates and deadlines" to item 4. If the "authority not found" count is rising, restrict the tool from citing anything and use it only for structure. Decisions like that are exactly what "competence" under 10.35 looks like in practice: knowing the tool's limits from evidence, not from marketing.
When something goes wrong: a short incident routine
The alert recommends preparing breach and error procedures, and a small firm can do this in a page. If an error reaches a client or return information goes somewhere it should not, the sequence is: stop using the tool for that task, tell the owner, preserve the prompt and output as they were, assess what information was involved, and take advice on any notification or correction duties. Do not delete the evidence. Then fix the cause, whether that is the prompt, the checklist or the approved list, and record the change in the log.
A realistic example: a preparer pastes a full client statement into a personal chatbot account to summarise it, because the approved tool was slow. Nothing malicious happened, but return information has gone into an unapproved account. The routine says: close the chat as far as the settings allow, record exactly what was entered, take advice about whether any consent or disclosure issue arises, and tell the client if advised. Then ask why the approved tool was slow, because a workaround usually points to a process problem you can fix.
A fee example with numbers
Take an illustrative hourly rate of $150 and a task that used to take 40 minutes to draft. A tool produces a first draft in five minutes, and review takes 15, so the work now takes about 20 minutes. Billing 40 minutes would charge for time not spent, which is the situation the alert warns about. Billing 20 minutes at $150 gives $50. A fixed fee of $75 for the document, priced on value and agreed in advance, is a different matter and can be reasonable, but the engagement letter should say how fees are set. A quick rule is to ask whether the client would be comfortable seeing how the bill was calculated. If not, change the bill or the explanation, and take advice if you are unsure.
Training staff so the sheet is not just paperwork
A checklist that people tick without reading protects nobody. Run a thirty-minute session each quarter built on real, anonymised examples from the error log. Show the draft, ask the team to find the problem, and then show how the checklist would have flagged it. The best exercise is a draft with three planted errors: a wrong figure, an invented citation and a smoothed-over inconsistency. Staff who find all three in under ten minutes understand what the checks are for.
New staff should complete the exercise before their first live use of any tool, and the owner should record attendance. That record, together with the logs and the sign-off blocks, makes up the evidence that the firm trains and supervises, which is the substance of 10.36. Keep copies of everything for the same period you keep workpapers, and ask your adviser how long that should be.
Fees and clients: the part people forget
Section 10.27(a) prohibits unconscionable fees in IRS matters. The alert flags that billing for time not actually spent, or charging twice for AI-assisted work, may fall foul of it depending on the facts, and that clients should be fairly credited with cost reductions. If a tool cuts drafting from forty minutes to fifteen, billing forty minutes at your hourly rate for that task invites a question. Many firms respond by moving to fixed fees for defined work and being open about how the work is done. Take advice on your own approach and put it in your engagement letter.
Testing before you trust a tool on live returns
Run a dry month before using a tool on real files. Take five completed, anonymised returns from a prior year and ask the tool to produce the same cover notes and schedules. Compare against the finished work and log every error: wrong figure, missing condition, invented authority. Count how many a reviewer following your checklist would have caught. If any slip through the checklist, tighten it before go-live.
Then set triggers for re-testing: a change of plan or model, a new use case, any error that reached a client, and the start of each season. Keep the logs. They are the firm's evidence of competence under 10.35 as well as a basis for training.
What this checklist does not do
It is not legal advice, and it does not guarantee compliance. It does not cover state rules, which differ and may include their own AI laws; the alert itself notes that several states have enacted AI governance laws. It does not decide whether your client consent forms are adequate. It turns the published guidance into a routine you can show. Have a qualified adviser review your version, and re-read the alert whenever the IRS updates it. For the broader picture of AI in a small firm, see the AI guide.
Circular 230 and AI questions
Did the IRS create new rules for AI?
No. The alert itself says the ethical obligations of competence, diligence and confidentiality remain unchanged. The Office of Professional Responsibility's June 2026 alert applies existing Circular 230 duties, such as due diligence, competence and supervision, to AI-assisted work. Read the alert on irs.gov and take advice from a qualified adviser on how it applies to your practice.
Can I rely on AI output if it looks right?
The alert says practitioners cannot rely solely on AI and that human scrutiny and editing are essential. Treat output as a draft, verify facts, citations and calculations against primary sources, and record who did that and when.
Do I have to tell clients I use AI?
The alert discusses disclosing AI activities as needed and crediting clients with cost savings, but it gives no separate rule number for it. Decide your approach with an adviser, then state it in your engagement letter so clients are not surprised.
Does a business AI plan solve the confidentiality issue?
It helps, because business plans do not train on your content by default, but it does not replace consent and security duties around return information. Check the rules on use and disclosure of return information, and see the consent and security tutorials linked here.
Further reads
- Can AI Do My Business Taxes? What It Can Safely Prepare — What AI can safely prepare and what it cannot.
- How Accountants Use AI to Explain Tax to Clients in Plain English — Explaining tax to clients without losing accuracy.
- Tax Software With AI: What a Small Business Should Check — Questions to ask a tax software vendor about AI.
- How Much Does AI Cost a Small Accounting Firm? — What AI costs a small accounting firm.
- AI Receptionist for Accounting Firms: Tax-Season Calls and Cost — Handling tax-season calls without exposing return data.
- AI Readiness Checklist for Accountants, Solicitors, Consultants — Test whether your firm is ready for wider AI use.
- AI Tax Return Prep for Small Practices: Extract, Review, Sign Off — A three-stage workflow for AI-assisted return prep: extract with page references, tie out with a spreadsheet, then review and sign off in writing.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Treasury Circular 230, 31 CFR sections 10.22, 10.27, 10.34, 10.35, 10.36, 10.37 and 10.51 (Cornell Legal Information Institute text); IRS Office of Professional Responsibility issue 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice, 24 June 2026 (PDF on irs.gov). This is not legal advice.