IRS AI Guidelines for Tax Professionals: Five Duties Firms Still Miss

On June 24, 2026, the IRS Office of Professional Responsibility told tax practitioners how to use AI. Alert 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice, applies existing Circular 230 duties to generative tools: verify every fact, citation, and calculation; bill for the work actually done; keep taxpayer data on secure, firm-approved systems; and write the procedure down. A chatbot does not carry those duties for you.

Last verified: September 29, 2026. Primary sources: OPR Alert 2026-19 (June 24, 2026) and Treasury Department Circular No. 230 (31 C.F.R. Part 10). The cybersecurity audit cited below is TIGTA Report 2026-200-053 (September 15, 2026).

The section-by-section reading of the same alert is in IRS Circular 230 AI guidance — what it actually means for CPA firms. What follows is the operational version: five duties, the slip under each, and what to do with client files.

What Alert 2026-19 is

Alert 2026-19 is OPR guidance on duties already in Circular 230. The OPR says virtually all professional tax firms already use some form of AI, including research and document tools they may not label as AI. Generative systems add fabricated output ("hallucinations"), bias, and opaque logic. They also create a confidentiality failure the alert describes directly: data entered for one client can be reused when the model answers a later question about a different client.

The alert points to court sanctions against lawyers for fake citations, and to a July 2025 Australian government report prepared by Deloitte Australia that contained invented quotes and nonexistent sources, after which Deloitte reportedly refunded part of its fee. The OPR includes those cases so firms treat model output as a draft.

The five requirements

1. Due diligence before the paper leaves the firm (§10.22)

Circular 230 §10.22 requires due diligence in preparing, approving, and filing returns and other papers relating to IRS matters, and in what you tell the Treasury and your clients. Applied to generative AI, that means a full review of every AI-created document and every AI-written passage before it goes to a client or to the IRS. Verify facts, citations, and calculations. The slip is shipping the draft because it looks finished: a memo with plausible case names, a workpaper with a clean calculation, or a client email summarizing a credit. Once you adopt the text, §10.22 applies to it. The alert's instruction is to treat AI text as a first draft.

2. Fees that match the time the tool saved (§10.27)

Section 10.27(a) bars an unconscionable fee in any matter before the IRS. Generative tools can cut research and drafting time. Charging for manual hours that were not spent, or billing twice for the same AI-assisted task, may violate §10.27 depending on the facts. The OPR's examples are a pattern across clients, or a large gap between time billed and time spent. Cost savings should be passed on openly. Disclose the AI work in general or specific terms as needed, and fairly credit the client's account for the reduction. The alert does not set a required discount percentage. The bill has to match the work.

3. Written advice you have actually checked (§10.37)

Section 10.37 covers written advice on a federal tax matter, including advice sent electronically. The advice has to rest on reasonable factual and legal assumptions, account for facts you know or should know, tie the law to those facts, and refuse unreasonable reliance on someone else's projections. You also may not weigh the advice by the chance the return will not be audited.

Generative projections, forecasts, citations, and formulas meet that standard only after a person confirms them. If the system's logic is opaque, reliance may be unreasonable under §10.37. Open the citations. Read the cases. Recompute inputs and formulas. The slip is sending the model's memo because it sounds like a tax opinion.

4. Taxpayer data stays off unsecured tools (§7216, §6713)

This is the duty that governs where client files go, and the alert states it as a prohibition. Internal Revenue Code §7216 makes knowing or reckless unauthorized use or disclosure of tax return information a criminal offense. Section 6713 is the related civil penalty. Circular 230 §10.51(a)(15) separately disciplines willful unauthorized disclosure or use.

Treasury Regulation §301.7216-1(b)(3) defines tax return information broadly. The alert quotes it: the term includes a taxpayer's name, address, or identifying number furnished in any form for, or in connection with, preparing that taxpayer's return. Workpapers, organizer answers, a wage statement in a chat window, and a pasted K-1 fall under that definition when they were gathered to prepare the return. The alert does not invent a separate workpaper rule. It applies the disclosure statutes to the place you type.

The instruction: use only secure, enterprise-approved AI, with confidentiality safeguards in place, and never upload sensitive data to unsecured sites. The slip is a staff member pasting return information into a consumer chatbot because it was faster than the approved process. Willful mishandling can bring Circular 230 discipline on top of the statute.

Two publications already sit under this duty. IRS Publication 4557 is the preparer's baseline for safeguarding taxpayer data. IRS Publication 5708 is the template for the Written Information Security Plan required by the FTC Safeguards Rule. See also Section 7216 explained and WISP requirements for tax firms.

5. A written firm procedure people actually follow (§10.36)

Section 10.36 puts the person with principal authority over the firm's IRS practice on the hook for adequate procedures, and for making sure people follow them. The alert says those procedures have to cover AI:

• Train staff on AI risks and on the Circular 230 requirements.
• Set protocols for secure data handling and for monitoring accuracy.
• Vet third-party AI tools, and document the vetting.
• Record how the firm uses AI and how it verifies output.
• Keep a procedure for breaches and errors.
• Stay current on the federal and state AI law that applies to the practice.

The slip is a verbal rule ("don't put client data in a public chatbot") with no training record, no approved-tool list, and no note of what happens when someone breaks it. Competence belongs in the same file. Under §10.35, competent practice means the knowledge, skill, thoroughness, and preparation the matter requires, and the OPR reads that to include how the AI system produces content, where it fails, and whether the output is fit for an IRS matter.

The five duties at a glance

Duty	Authority	What the alert requires	Common slip
Due diligence	§10.22	Verify facts, citations, and calculations before anything goes to a client or the IRS	Shipping the model's draft because it reads cleanly
Fees	§10.27	Reflect real time savings; disclose AI work; credit the client for efficiencies	Billing the old manual hours
Written advice	§10.37	Confirm assumptions, authorities, and math; opaque logic can make reliance unreasonable	Treating a generated memo as an opinion
Confidentiality	§7216, §6713, §10.51(a)(15)	Use only secure, enterprise-approved systems; never upload sensitive data to unsecured sites	Pasting return information into a public chatbot
Firm procedures	§10.36, with §10.35	Written training, access rules, accuracy checks, vendor vetting, and a record of all four	A hallway policy nobody can produce

The September 2026 TIGTA report is about the IRS's own systems

On September 15, 2026, TIGTA issued Report 2026-200-053, The IRS's Cybersecurity Program Was Not Effective for Fiscal Year 2026. The program was rated not effective because three NIST Cybersecurity Framework functions — Identify, Protect, and Detect — fell below the FISMA bar of maturity Level 4. Govern, Respond, and Recover were rated effective. In the Protect findings, 86 percent of the sampled systems (6 of 7) still had critical vulnerabilities open past the 30-day fix window. TIGTA's warning is about the agency: if those gaps stay open, taxpayer data could be vulnerable to inappropriate and undetected use, modification, or disclosure.

That audit covers the IRS under the Federal Information Security Modernization Act. It leaves Circular 230 and your §7216 and §10.22 duties where they were. Taxpayer data is exposed when the system holding it is the wrong one, or when nobody checks the output.

What to put in the file this week

1. Name the AI tools the firm actually uses, including research platforms staff already have open.
2. Split that list into approved and prohibited. Public chat tools stay prohibited for anything that meets the §301.7216-1 definition of tax return information.
3. Write who checks facts, citations, and calculations, and where that check is recorded, before AI-assisted text reaches a client or a return.
4. Align invoices with §10.27 so AI-reduced time shows up in the bill.
5. Add the AI procedure to the WISP and to the §10.36 file: training, vendor review, access control, and the steps after an error or a breach.
6. Move document collection onto a channel the firm controls. A walkthrough is in how to send tax documents securely.

Where secure client communication fits

The alert's confidentiality line is about the system that holds the file. XyloDocs is an AI client communication platform for tax firms: a client portal, document intake, e-signature, and bilingual calling, SMS, and WhatsApp, with AES-256 encryption, access controls, and an audit trail built to align with IRS Publication 4557 and IRS Publication 1075. The communication platform is available now, on its own. Client accounts are included.

That is the path client documents and messages travel. Citation checks, fee decisions, and the Written Information Security Plan stay with the firm, which is what §10.22, §10.27, §10.36, and §10.37 already require. For the question of which AI may see a return, start from the approved-tool list in your procedures, then read what the IRS expects when you use AI with client data.

Frequently asked questions

What are the IRS AI guidelines for tax professionals in 2026?

OPR Alert 2026-19, issued June 24, 2026, explains how Circular 230 applies when a practitioner uses AI. Verify AI output before it reaches a client or the IRS, bill fairly for AI-assisted work, keep taxpayer data on secure enterprise-approved systems, and adopt written procedures for training, vendor review, and accuracy checks.

Does Circular 230 allow AI on tax workpapers?

Yes, as a draft aid. Section 10.22 still requires you to verify facts, citations, and calculations before the workpaper is used on a return or sent to a client. If the workpaper contains tax return information — including a name or identifying number gathered to prepare the return — §7216 and §6713 also restrict where that file can be uploaded.

Can a tax practitioner paste client data into a public AI tool?

The alert says to use secure, enterprise-approved AI and to keep sensitive data off unsecured sites. Unauthorized disclosure of tax return information can bring a civil penalty under §6713, a criminal penalty under §7216, and discipline under Circular 230 §10.51(a)(15).

What does firm leadership have to document under these guidelines?

Under §10.36, the person responsible for the firm's IRS practice needs procedures staff actually follow: AI training, secure data handling, accuracy monitoring, a record of third-party AI vendors you have vetted, and a way to handle breaches or errors.

Does the 2026 TIGTA cybersecurity report change AI rules for preparers?

Report 2026-200-053 rates the IRS's own cybersecurity program. Identify, Protect, and Detect were not effective for fiscal year 2026. Preparer duties under Circular 230 and §7216 are unchanged.

Sources and related reading

• OPR Alert 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice (June 24, 2026)
• Treasury Department Circular No. 230 (31 C.F.R. Part 10), especially §§10.22, 10.27, 10.35, 10.36, 10.37, and 10.51
• 26 U.S.C. §7216 and 26 U.S.C. §6713; Treas. Reg. §301.7216-1
• IRS Publication 4557, Safeguarding Taxpayer Data
• IRS Publication 5708, Written Information Security Plan
• TIGTA Report 2026-200-053 (September 15, 2026)
• IRS Circular 230 AI guidance for CPA firms
• What the IRS expects when you use AI with client data