What did the IRS say about AI in tax practice?

On June 24, 2026, the IRS Office of Professional Responsibility issued its first guidance on artificial intelligence in federal tax practice. The short version: your existing Circular 230 duties already cover AI, and none of them get easier because a machine did the first draft. The guidance (OPR Alert 2026-19, "Introductory Guidelines for Responsible AI Use in Federal Tax Practice") does not create new rules. It applies the rules you already work under to a new tool. Its own summary line is the one worth pinning above your desk: technology is a powerful tool, not a substitute for professional judgment.

For most firms this is not a warning about some future workflow. You almost certainly use AI already, inside research platforms and document-review tools, whether or not anyone called it that. What changed is that the OPR has now said, on the record, who owns the output when that tool gets something wrong. You do.

Does the IRS allow tax professionals to use AI?

Yes. The IRS does not prohibit AI in tax practice, and the OPR guidance is explicit that AI can be used responsibly. What it prohibits is treating AI output as finished work. Every duty below is a version of the same idea: the tool can do the labor, but a practitioner still signs, and the signature still means what it always meant.

The six Circular 230 duties, applied to AI

The guidance maps to six provisions you already know. Here is what each one now requires when generative AI is in the workflow.

Due diligence (§10.22). You must verify the accuracy of every fact, citation, and calculation an AI produces before it goes to a client or the IRS. Due diligence cannot be delegated to an algorithm. This is the provision behind the headlines about attorneys sanctioned for fabricated case citations: a generative model will invent a confident, well-formatted source that does not exist, and the person who filed it is accountable, not the model.

Competence (§10.35). Competence now includes understanding the AI you use, its limitations, and how it can fail. You do not need to build the model, but you are expected to know enough about how it works to judge when its output is unreliable. "I did not know it could hallucinate" is not a defense the guidance leaves open.

Fees (§10.27). When AI cuts the time a task takes, your billing has to reflect the time you actually spent. Charging a client for hours of manual research that the tool did in minutes can run into the unconscionable-fee rule. The efficiency is real; the guidance says the client should see it.

Compliance procedures (§10.36). Firm leadership is responsible for having AI policies in place: staff training on the risks, documented protocols for secure data handling and accuracy checks, and a process for vetting third-party tools before anyone uses them. This one lands on partners and firm owners, not individual preparers.

Written advice (§10.37). Advice to clients must rest on verified information, not on an unconfirmed AI output. If a recommendation traces back to something the model asserted and no one checked, the §10.37 standard is not met.

Confidentiality (§6713, §7216). You cannot put identifiable client data into an unsecured or public AI system. IRC §7216 and §6713 already penalize unauthorized disclosure of taxpayer information, and typing a client's return into a consumer chatbot is a disclosure. The guidance says client data belongs only in secure, enterprise-grade tools with real confidentiality safeguards.

The part the guidance implies but does not spell out: vet the tool

The OPR tells you to vet third-party AI tools before use and to keep client data inside secure, enterprise-approved systems. It does not tell you how to tell one from the other. That is the gap every firm now has to close, because the duty is yours whether or not the vendor makes it easy.

A tool that meets the Circular 230 bar has to answer five questions cleanly. Ask them of anything you are already using and anything a vendor is selling you.

1. Where does my client data go, and who can see it? The answer has to be specific: named infrastructure, no training on your inputs, no shared context between unrelated client matters. "It's secure" is not an answer. Data spilling between clients is one of the failure modes the OPR names directly.

2. Is it built for §7216 and §6713, or retrofitted? Enterprise tax tools should treat taxpayer confidentiality as the design constraint, not a settings toggle. If the security story is a paragraph on a pricing page, keep asking.

3. Can I verify every output back to a source? Due diligence under §10.22 means checking facts, citations, and calculations. A tool that shows its work, cites the actual IRS provision, and lets you trace a number to its origin makes that duty doable. A tool that hands you a confident paragraph with no source makes you do the verification blind.

4. Does it keep a human in the loop by design, or does it market autonomy? The guidance is clear that the review step is not optional. A vendor promising to remove the practitioner from the workflow is selling you the exact thing the OPR says you cannot delegate. The right tool speeds up your review; it does not replace it.

5. Does the vendor overclaim? A tool that promises to do things it cannot yet do is a compliance risk before you even turn it on, because your written advice (§10.37) can end up resting on capability that was never there. Straight answers about what a tool does today, versus what is on the roadmap, are themselves a signal you are dealing with a serious vendor.

At SignalsHQ we build tax AI to sit under a practitioner, not over one. Our research layer (Tax Assist) cites the IRS provision it relied on so you can check it, and the review layer surfaces the workpaper and QC flags for a human to sign off. That is a deliberate design choice, and after this guidance it is also the only design that fits Circular 230. If you are working through how to bring AI into your firm without tripping these duties, we are happy to talk it through.

A short checklist before tax season

• Write a one-page AI acceptable-use policy and put a named owner on it.

• Keep a register of approved tools with a note on each one's data handling.

• Log the human verification step so due diligence is documented, not assumed.

• Train every staff member who touches AI on hallucination, confidentiality, and what they must check.

• Confirm no client data ever enters a consumer or public AI tool.

None of this is new obligation. It is the obligation you already had, written down for a tool you are already using.

Frequently asked questions

Can my firm use AI for tax work under Circular 230? Yes. The IRS OPR guidance from June 24, 2026 permits AI use and does not create new rules. It requires that your existing Circular 230 duties (due diligence, competence, confidentiality, and fees) are met when AI is in the workflow, which means a practitioner verifies every output before it goes out.

Do I have to tell my clients I used AI? The OPR guidance does not impose a specific client-disclosure requirement. It focuses on the duties you owe regardless: verifying accuracy under §10.22, protecting client data under §7216, and billing for time actually spent under §10.27. Firm policy and client engagement terms are where disclosure practices are set.

What counts as an "enterprise-approved" AI tool? The guidance does not certify specific tools. In practice, an enterprise-approved tool keeps client data in secure named infrastructure, does not train on your inputs, does not share context between unrelated clients, lets you verify every output to a source, and keeps a human in the review loop. Consumer chatbots do not meet this bar because entering client data into them can be an unauthorized disclosure under §7216.

Who is responsible if an AI tool produces a wrong result? The practitioner is. The OPR guidance states that due diligence cannot be delegated to an algorithmic process, so the person who signs and submits the work is accountable for its accuracy, not the AI vendor or the model.

Source: IRS Office of Professional Responsibility, Alert 2026-19, "Introductory Guidelines for Responsible AI Use in Federal Tax Practice," issued June 24, 2026; Treasury Department Circular 230 §§10.22, 10.27, 10.35, 10.36, 10.37; Internal Revenue Code §6713 and §7216.