Responsible AI Use for Tax Professionals: Turning OPR’s AI Alert Into Firm Policy

AI is already inside tax practice, including research platforms many firms license. On June 24, 2026, the IRS Office of Professional Responsibility (OPR) issued Introductory Guidelines for Responsible AI Use in Federal Tax Practice, identified as OPR Alert, Issue 2026-19. OPR Alert 2026-19 does not create a separate AI rulebook. It explains how existing Circular 230 duties and taxpayer-information rules apply to AI-assisted federal tax practice, including due diligence, competence, fees, written advice, firm procedures, and confidentiality.

It’s easy to read the alert as a caution list. The OPR alert applies existing duties involving due diligence, competence, confidentiality, written advice, and fees to AI-assisted work. Section 10.36 makes documented firm-level governance the operating center. A living AI policy turns those duties into procedures staff can follow, document, and update as tools and guidance change. What follows is what the guidance says, how to adopt AI inside it, and how to build that policy.

How Is AI Already Changing Tax Practice?

AI has moved from experiment to infrastructure in tax practice and adoption is broad and accelerating. 60% of tax professionals use AI for tax research at least weekly, up from 33% in 2025, according to a 2026 Blue J and CPA.com survey of 1,000 accountants. The IRS treats the presence of AI as settled fact. Its alert observes that virtually all professional tax firms use some form of AI, whether they are aware of it or not.

The same shift is happening inside the IRS. The agency uses AI and machine learning to sort large volumes of data, identify patterns, and prioritize returns for human review. Tax professionals are now working in an environment where AI influences both how firms serve clients and how the IRS identifies returns that warrant closer scrutiny.

The economics are following. In the same survey, 69% anticipate moving to value-based, hybrid, or fixed-fee billing as AI commodifies routine tax knowledge.

These introductory guidelines are the agency’s response to adoption that is already changing work, risk, and pricing. Firm owners now have to govern how staff use it, what data enters it, and how its work is checked.

What Does OPR Alert, Issue 2026-19 Actually Say?

The guidance applies five Circular 230 provisions plus the IRC confidentiality sections to AI workflows. It names three limitations of generative AI:

• Fabricated facts, citations, or calculations.
• Biased output.
• Opaque reasoning or sources.

The alert notes that courts have increasingly sanctioned lawyers for improper use of AI, mainly fake citations and other hallucinations in legal filings, with penalties from financial sanctions and public censure to removal from representation. Most sanctions so far involve attorneys, though other tax-professional cases are beginning to appear. The guidance’s own example is Deloitte Australia, which delivered a 230-plus-page government report in July 2025 with invented judge quotes and references to nonexistent reports, and reportedly agreed to refund part of its fee.

The confidentiality stakes are civil and criminal. §6713 imposes civil penalties for each unauthorized use or disclosure, and §7216 carries up to a $1,000 fine, up to one year imprisonment, or both, rising to as much as $100,000 when the disclosure or use connects to an identity-theft crime.

Why AI Governance Is a Firm Obligation Under §10.36

Section 10.36 turns the alert’s cautions into a firm operating obligation. The practitioner who holds or shares principal authority for the firm’s tax practice must take reasonable steps to ensure adequate procedures exist and are followed. Discipline under §10.36(b) requires more than an isolated policy gap. The rule ties liability to willfulness, recklessness, or gross incompetence plus a pattern or practice of noncompliance, or to knowing (or having reason to know) of that pattern and failing to correct it promptly.

Compare two firms. One relies on a senior practitioner’s careful habits. The other uses written procedures a first-year preparer can follow.

When the senior practitioner retires, the first firm can show only what that person usually did. A new hire can undo that informal control in one prompt. The second firm can show the approved tools, prohibited inputs, review steps, training records, and escalation path that should have governed the work.

The difference is durability. Caution stored in one practitioner’s memory disappears with staff turnover and weakens during busy season. A written procedure can survive both and produce evidence of the firm’s controls.

How Should a Tax Firm Adopt AI Responsibly?

Responsible adoption requires separate controls for verification, client data, and fees.

Verification comes first. Treat AI output like a new preparer’s draft. A qualified professional must check the facts, authorities, calculations, assumptions, and fit to the client’s circumstances before the work leaves the firm.

Data handling creates a separate analysis. OPR directs practitioners to use only secure, enterprise-approved AI for client data. The alert’s own example of the risk is cross-client leakage: data generated for one client repurposed to answer an inquiry about another. The alert stops at security, though. Whether entering tax return information into even an approved tool is a permitted disclosure is a separate question under the §7216 regulations. Those regulations define disclosure as making tax return information known to any person in any manner (Treas. Reg. §301.7216-1(b)(5)) and list specific uses and disclosures that do not require consent. The IRS has not issued AI-specific guidance on how general-purpose or enterprise AI tools fit those exceptions.

The shortcut is usually innocent. A resolution specialist pastes a client’s notice into a consumer chatbot to draft a communication to their client. The input may constitute a disclosure even if the provider says it retains nothing. When no §301.7216-2 exception clearly applies, obtain valid taxpayer consent before the information is entered. The consent requirements are covered in the policy section below.

Billing remains unsettled. OPR says cost savings should be passed on openly and fairly credited to the client. It warns that billing for manual labor or time not actually spent may violate the provision depending on the facts, such as a noticeable pattern across clients or the size of the billing differentials. The AICPA calls the pass-through expectation an overly simplistic view that ignores the full economics of AI adoption, including software licensing, implementation, governance, and training costs. The alert also directs practitioners to disclose the AI activities performed on the engagement, in general or specific terms as needed. Revisit the policy when the IRS clarifies its position.

DO

• Verify every AI-drafted citation and calculation before it leaves the firm.
• Map each planned use of tax return information to a §301.7216-2 exception or valid taxpayer consent before use or disclosure.

DON’T

• Assume a vendor’s privacy promises make an input safe.
• Bill hours that weren’t worked, with or without a written pricing rationale.

How Do You Build an AI Use Policy That Stays Current?

A living policy needs clear boundaries, assigned controls, and a defined update process. To cover the full scope of the guidance, your AI governance framework should connect three components that address different risks:

• Written Information Security Plan (WISP): Security controls, access, approved systems, service providers, data handling, retention, and incident response.
• AI use policy: Approved uses and tools, prohibited inputs, human review, training, documentation, billing, and accountability.
• §7216 decision process: The forms and workflow for obtaining taxpayer consent before tax return information enters an AI tool.

What the WISP should cover

The FTC Safeguards Rule requires a documented plan to protect client data, and the Written Information Security Plan (WISP) your firm already maintains is the natural home for AI security controls. Your WISP should state which AI tools are approved, who is authorized to use them, and what data can be entered.

OPR reinforced that obligation in Issue 2026-22 on written information security plans. That July 22 guidance ties data security to §10.35 and §10.36 and says a WISP should be updated as business or technology changes dictate.

The WISP should not carry the entire governance burden because §10.36 also reaches competence, verification, training, documentation, written advice, and firm procedures.

For the security side of the work, use the IRS Solutions guide to building and maintaining a WISP. It links to IRS Publications 5708, 5709, and 4557 and explains how to adapt the plan to your firm’s workflow, technology, and staffing.

What the AI use policy should cover

• Scope, approved business uses, and an inventory of approved AI tools.
• Data classifications, prohibited inputs, access controls, and links to the WISP and the §7216 consent process.
• Human verification and approval standards for facts, citations, calculations, and written advice.
• Usage records that show the tool, purpose, reviewer, verification steps, and final disposition.
• Vendor diligence covering security, data location, retention, model training, subcontractors, and contract terms.
• Role-based staff training, testing, and escalation procedures.
• Client disclosure and billing standards for AI-assisted work.
• Incident response for errors, unauthorized disclosures, client notification, correction, and tool suspension.
• A policy owner, approval authority, effective date, review history, and change log.

What §7216 means for AI use

Before entering client information into an AI tool, determine whether §7216 applies to the firm and the information, whether an exception permits the use or disclosure, and, if not, whether taxpayer consent is required.

For tax resolution firms, this can require additional analysis. Information obtained in connection with preparing a return may be tax return information even when it is later used in resolution or representation work. Information obtained independently for representation is not automatically covered merely because it concerns the same taxpayer. Mixed engagements should be evaluated based on the facts.

De-identification alone does not necessarily resolve the issue. Nor does a vendor’s promise that it does not retain or train on client data. If no applicable exception permits the disclosure or use, valid taxpayer consent may be required before the information is entered into the AI tool.

When consent is required for a Form 1040-series taxpayer, Rev. Proc. 2013-14 sets requirements for its content, format, timing, and signature. Firms should have counsel review their §7216 analysis and decision process before allowing client information to be used with AI systems.

How the policy stays current

Assign one owner to monitor the policy and coordinate updates with the firm’s WISP, engagement documents, vendor review, and training.

• Review the policy annually as a firm practice, or more often when the firm’s risk profile requires it.
• Trigger an interim review after a change in law or guidance, an approved tool or vendor change, a new use case, an incident, or a material control failure.
• Record the change, approver, effective date, affected procedures, and required staff communication or training.

Neither the alert nor Circular 230 prescribes an annual AI policy review. Annual review is a practical default, paired with event-based triggers and OPR’s direction to stay current.

How Can Tax Firms Govern AI as Guidance Evolves?

The governance environment is still developing. On September 22, 2026, the AICPA announced the Council on AI Risk in Tax, which it will lead with former IRS Commissioner Danny Werfel. The council will refine the AI Risk Framework for Tax and develop practical resources for responsible AI use in tax administration and practice. Its formation reinforces the need for a living AI policy that can accommodate new tools, risks, and professional guidance without forcing firms to start over.

Responsible AI adoption is an ongoing governance responsibility for tax firms. The IRS alert offers a starting point for managing AI-assisted research and workflows through human review, client-data safeguards, and documented procedures. Firms that assign ownership, define review intervals, and establish triggers for policy updates will be better prepared to gain efficiency while managing compliance risk as the technology and guidance continue to evolve.

Frequently Asked Questions

Does the guidance apply if we have never deliberately adopted AI?

Yes. The alert says virtually all professional tax firms use some form of AI, including through licensed research platforms, and calls AI ubiquitous in modern law and accounting offices. For a firm that never made a deliberate AI decision, existing professional duties still apply to AI embedded in software the firm already licenses.

What happens if a firm has no AI procedures?

A policy gap alone does not trigger discipline. Under §10.36(b), the practitioner with principal authority for the firm’s tax practice is subject to discipline when, through willfulness, recklessness, or gross incompetence, they fail to take reasonable steps to put adequate procedures in place or to see that they are followed, and firm members engage in a pattern or practice of noncompliance. The same standard reaches a practitioner who knows or should know of such a pattern and fails to correct it promptly.

What counts as a disclosure into an AI tool?

Entering tax return information into a third-party AI tool makes it known to the provider, which fits the regulatory definition of disclosure (Treas. Reg. §301.7216-1(b)(5)). The open question is whether an exception covers it. Section 301.7216-2 contains exceptions, including limited disclosures for qualifying auxiliary services, but the IRS has not said how AI tools fit them. When no exception clearly applies, obtain consent first.

How often should we revisit our AI use policy?

The guidance sets no revisit frequency, so anchor the default where your firm already has a calendar. The natural fit is the annual security check on your written information security plan, or WISP, paired with the standing triggers, a guidance change, a tool change, or an incident. Scope matters as much as schedule. The guidance’s stay-updated directive reaches beyond Circular 230, and it names four states that have already enacted AI governance legislation, California, Colorado, Illinois, and Utah, alongside ABA Formal Opinion 512 on generative AI tools.

Disclaimer: This article is provided for general educational and informational purposes only. It is not legal, tax, accounting, or professional advice and should not be relied on as a substitute for guidance based on a firm’s specific circumstances. Laws, regulations, professional standards, and IRS guidance concerning AI use and taxpayer information may change. Tax professionals should review the cited authorities and consult qualified legal or compliance counsel before adopting policies, disclosing taxpayer information to an AI provider, or relying on an exception or consent process. Links are provided for reference, and third-party content may change after publication.