AI and the IRS rules: what US accounting firms can and can't do

Short answer: yes, CPAs and tax preparers can use AI. The IRS has not banned it, no state board has banned it, and as of June 2026 the IRS has said in writing that AI can be a legitimate part of federal tax practice. The rules regulate two things: what happens to client tax return information when you feed it into someone else's software, and who is accountable for the output. Get those right and AI is fair game. Get them wrong and you are in criminal statute territory, not just an awkward peer review.

This piece maps the three layers that matter for a US firm: IRC Section 7216 (the criminal disclosure rules), Treasury Circular 230 (your professional duties before the IRS), and the AICPA and state board layer on top. One plain caveat: this is journalism, not legal advice, and the 7216 questions turn on regulations written before generative AI existed. Run your specific setup past a lawyer.

Section 7216: the rule most firms are quietly breaking

IRC Section 7216 makes it a federal crime for a tax return preparer to knowingly or recklessly disclose or use tax return information for any purpose other than preparing the return. Penalties run up to $1,000 per violation, up to a year in prison, or both. Its civil twin, Section 6713, adds $250 per prohibited disclosure, capped at $10,000 per calendar year. These are not AI rules, they date back decades. But they bite hardest right now on AI, because pasting a client's K-1 detail into a chatbot is exactly the kind of transfer the statute polices.

"Tax return information" is defined broadly: names, addresses, identifying numbers, income figures, even the fact that someone is your client. When it leaves your systems for an AI provider's servers, a disclosure has occurred, whether or not the provider retains, trains on, or instantly deletes the data.

The regulations do have exceptions, and this is where the honest uncertainty lives:

• Disclosures within the US for preparation assistance. Treas. Reg. 301.7216-2 permits disclosure to a US-based contractor or service provider that assists in preparing, processing, or e-filing the return, without client consent. This is why your tax software, hosted in a vendor's cloud, has never needed a signed consent form.
• Auxiliary services. Providers performing a direct, necessary role in return preparation sit inside the fence. Standard tax prep software almost certainly qualifies.
• The catch. The exception narrows sharply when the outside service makes substantive determinations: filing status, how income is reported, which deductions and credits are claimed. That kind of assistance requires consent. So the more you ask an AI to actually decide things about the return, the weaker your no-consent position gets.

Where does a general-purpose tool like ChatGPT or Claude land? Commentators including The Tax Adviser have concluded that general-purpose AI is unlikely to qualify as an auxiliary service, and the IRS has issued no guidance saying otherwise (as at July 2026). The conservative reading, and the one most firm risk committees are adopting, is: assume consent is required before client tax return information touches a general-purpose AI tool, or keep the information out of the tool entirely.

If you need consent, do it properly

A compliant 7216 consent is its own document, not a line in your engagement letter saying "we use technology." For 1040-series clients, Rev. Proc. 2013-14 prescribes the format and mandatory language, the consent must name the specific recipient ("various AI tools" fails), and it must be signed before the disclosure, not ratified after. A consent naming one provider also does not cover the day your team switches to another, which is a big part of why firms prefer the anonymization-first approach below. The drafting detail, including the engagement letter clause and a client-facing explainer, lives in our guide to client consent for AI.

Circular 230: the IRS told you the standard in June 2026

On June 24, 2026, the IRS Office of Professional Responsibility released Alert 2026-19, its first guidance on responsible AI use in federal tax practice. OPR did not write new rules. It said the existing ones already cover AI, and then spelled out how:

• Due diligence (Section 10.22). You must thoroughly review anything AI produced before it goes to a client or the IRS: facts, citations, calculations, all of it. The alert cites the growing pile of court sanctions over fabricated AI citations as the cautionary tale.
• Competence (Section 10.35). Competence now includes understanding the technology you use: how the system generates content, where it errs, where bias creeps in. "I didn't know it made things up" is not a defense, it is the violation.
• Firm procedures (Section 10.36). Partners and managers must have adequate procedures in place: staff training on AI risks, secure data handling protocols, and vetting of third-party AI tools before anyone uses them on client work.
• Fees (Section 10.27). A pointed one: you cannot bill hours for work AI did in seconds and present it as manual labor. OPR expects honesty about how the work was produced.

The through-line is accountability. The practitioner signs the return, the practitioner answers to OPR, and AI changes neither fact. That is the same conclusion we reached in no AI is ever going to jail for you, and it is now effectively the IRS's official position.

The AICPA and state board layer

For CPAs there is a third layer. The AICPA Code of Professional Conduct's confidentiality rule requires client consent before disclosing confidential client information, and submitting client data to a third-party AI platform is a disclosure under that rule too, independent of 7216. The revised Statements on Standards for Tax Services, effective January 1, 2024, reinforce the diligence and quality-control expectations that AI-assisted work must clear.

State boards of accountancy largely adopt versions of the AICPA Code, so a confidentiality breach through careless AI use is also a potential licensing matter. As at July 2026, no state board appears to have issued AI rules that go materially beyond the AICPA position. The high-level takeaway: satisfy 7216 and Circular 230, treat client data as confidential by default, and you will generally satisfy your board too. Check your own state's board for anything specific.

The do and don't list

You can, without client consent:

• Use AI for research, drafting, and technical questions that contain no client-identifying information. "Explain the Section 179 limits for 2026" discloses nothing.
• Use AI on genuinely anonymized scenarios. Strip names, SSNs, addresses, employer names, and unusual figures that could identify the taxpayer. Be honest with yourself about whether the residue is really anonymous.
• Use AI features inside your existing tax and practice software where the vendor sits within the preparation-assistance exception. Confirm this in the vendor's terms rather than assuming it. The same applies to AI features in your practice management platform, whether that is FYI, Kloud Connect, or XPM.
• Use AI for firm-side work that involves no tax return information at all: marketing, internal templates, workflow documentation, staff training material.

You cannot, or should not:

• Paste identifiable client tax return information into a consumer AI tool with no consent in place. This is the clearest 7216 exposure in the whole discussion.
• Rely on a vague "we use AI" clause as consent. Rev. Proc. 2013-14 requires specificity for 1040-series clients.
• Let AI make substantive return determinations and treat it as mere processing assistance. Substantive decisions push you out of the no-consent exceptions.
• File or send anything AI-drafted without human verification of every fact, figure, and citation. That is now an explicit OPR expectation, not just good practice.
• Use free-tier tools that train on your inputs for anything touching client work. Enterprise agreements with no-training commitments are the baseline. Our breakdown of which AI tools are safe for client data covers the specific vendor terms.

What a compliant firm setup actually looks like

None of this argues for avoiding AI. The gains are real and the rules are workable. The firms doing this well converge on four moves: enterprise-grade AI with written no-training and confidentiality terms, an anonymize-first habit for anything client-related, a short written policy so staff are not improvising (our one-page AI policy guide is a starting template, and OPR's firm-procedures expectation makes having one close to mandatory), and a named human reviewer on any AI-assisted output that leaves the building. Tool choice matters less than the data terms behind it; our Claude vs ChatGPT vs Copilot comparison covers where each vendor stands.

The rules predate generative AI, and the IRS has chosen to interpret rather than rewrite them. Good news: the compliance path is the one you already know. Protect the data, verify the work, own the output. For the broader picture, start with our AI for accounting firms pillar.

Frequently asked questions

Can a CPA legally use ChatGPT for client tax work?

Yes, with conditions. Using it for research or anonymized scenarios is fine. Entering identifiable client tax return information likely requires prior written consent under Section 7216, on the conservative reading that general-purpose AI falls outside the preparation-assistance exceptions. Many firms sidestep the question by anonymizing inputs or using enterprise tools vetted under their AI policy.

Does the IRS have to approve an AI tool before my firm uses it?

No. There is no IRS approval, certification, or registration process for AI tools (as at July 2026). The burden runs the other way: your firm must vet the tool, and under Circular 230 Section 10.36 the IRS expects documented procedures showing you did.

What is the actual penalty if my firm mishandles client data in an AI tool?

Section 7216 is criminal: up to $1,000 per violation and up to a year in prison for knowing or reckless disclosure. Section 6713 adds a civil $250 per disclosure, capped at $10,000 a year. Beyond the statutes, you face OPR discipline under Circular 230, state board action, and the client trust damage that outlasts all three.

Is AI inside my tax software treated differently from a standalone chatbot?

Generally yes. US-based software that assists in preparing, processing, or e-filing returns sits within the regulatory exceptions tax software has always relied on, so consent is usually not required. The distinction weakens if the AI feature makes substantive determinations, so read the vendor's terms and ask where your data goes.

Do these rules apply to unenrolled preparers or just CPAs and EAs?

Section 7216 and its civil twin apply to anyone who prepares returns for compensation, credentialed or not. Circular 230 governs those who practice before the IRS: CPAs, EAs, and attorneys. Unenrolled preparers largely escape OPR's reach, but not the criminal disclosure statute, which is the sharper edge anyway.