Practitioner Standards for Verifying AI-Generated Tax Answers

Citation-backed AI tax research requires more than an answer with footnotes. For complex federal and state questions, practitioners need access to the underlying authority and a protocol for checking its meaning, relevance, and current status. OPR Alert 2026-19 created no new rules, but its mapping of existing professional obligations removes any credible excuse for delivering AI-generated tax answers without substantive verification.

What OPR Alert 2026-19 actually did — and what it did not do

The IRS Office of Professional Responsibility issued OPR Alert 2026-19 on June 24, 2026, formally mapping existing Circular 230 obligations to AI use for the first time. The OPR enforces Treasury Circular 230, 31 C.F.R. Part 10, and oversees CPAs, attorneys, enrolled agents, enrolled retirement plan agents, and enrolled actuaries who practice before the IRS.

The word “Introductory” in the alert’s title describes guidance directed at the current state of AI technology, with further guidance expected as tools and regulatory frameworks evolve. The alert distinguishes established research, document review, and analytics tools from generative AI systems that produce original documents, analysis, research, and tax advice with minimal human interaction. Its immediate effect is clarification of responsibility.

Why generative AI creates a different category of professional risk than earlier software

Generative AI can produce fluent, formatted, citation-bearing answers whether or not the underlying authorities, facts, or calculations are correct. A spreadsheet error may present a wrong number that a practitioner can spot. A generative system may instead present a confident paragraph citing a Revenue Ruling that does not exist. The distinction is not merely the speed of production, but the appearance of authority attached to an unsupported answer.

The OPR identifies fabricated outputs, bias, and lack of transparency as hazards that translate into professional risk when practitioners adopt the output. Its definition of AI as “machines that mimic human cognitive skills, including judgment, perception, and prioritization” explains why technological literacy belongs within competence. The tool imitates functions for which the practitioner remains responsible. Understanding its limits is therefore part of deciding whether its work is suitable for an engagement.

The OPR also identifies cross-contamination of client information as a distinct confidentiality hazard. Information generated for one client may be repurposed in an answer about another client or spilled into a related algorithm. When choosing a tool, firms must check how it keeps client documents apart as well as how well it answers questions. Among the purpose-built options, Marble is SOC 2 Type II certified and has strict client document context boundaries designed to keep data from spilling into other clients’ answers or related algorithms. Those controls address a specific hazard, but they do not replace the firm’s review of whether a tool is appropriate for its intended use.

The six Circular 230 provisions OPR Alert 2026-19 maps to AI outputs

The alert’s six-part framework covers diligence, competence, supervision, written advice, fees, and taxpayer data protection. More precisely, it brings together five Circular 230 provisions and IRC §§ 6713 and 7216. These obligations attach at different points in an engagement, so no single accuracy check can satisfy them all.

Section 10.22 requires due diligence in preparing or assisting in the preparation of, approving, and filing papers relating to Internal Revenue Service matters. It also covers the correctness of representations to the Department of the Treasury and to clients. The OPR makes clear that this federal standard cannot be delegated to an algorithmic process. For AI-generated research, practitioners must independently establish that each cited authority exists, supports the claimed proposition, and has not been superseded, overruled, or modified. They must also check calculations against source numbers and confirm that factual statements are accurate and include all material facts.

Section 10.35 requires the knowledge, skill, thoroughness, and preparation appropriate to the matter. The OPR applies that standard to technological literacy, including understanding how AI generates content and where errors or bias may arise. Practitioners must be able to assess whether an output is suitable for use in IRS matters. A lack of understanding that leads to improper advice or flawed filings is itself a competence problem.

Section 10.36 places responsibility for adequate firm procedures on leadership. Its application requires staff training, documented protocols for secure data handling and output monitoring, and vetting of third-party AI tools before deployment. Unsupervised use of consumer AI is therefore not solely an employee’s lapse. It can reveal a firm-level failure to define permitted tools, permitted uses, and required review.

Section 10.37 applies to advice in any written form, including client emails, letters, and opinion letters. Advice must rest on reasonable factual and legal assumptions that the practitioner independently verifies. Opaque AI reasoning that cannot be traced to verified authority may make reliance unreasonable. For example, an AI-drafted state income and sales tax nexus study requires state-by-state verification of sales thresholds, transaction thresholds, tax rates, and relevant effective dates. A tax opinion letter likewise requires review of the legal analysis and transaction facts, with fabricated, outdated, or inapplicable authorities replaced before delivery.

Section 10.27 prohibits an unconscionable fee in connection with a matter before the IRS. The OPR identifies double-billing for AI-assisted work as a potential violation, including charging as though displaced manual work had actually occurred. Practitioners should tell clients what work AI did and fairly credit the savings to the client’s account. Transparent billing also creates a sound basis for explaining the value of professional review and advisory work rather than charging for production time the tool eliminated.

IRC §§ 6713 and 7216 impose civil and criminal penalties for unauthorized disclosure of tax return information. Entering client data into an unapproved AI platform can create that exposure even when the prompt omits the client’s name, because combined figures and details may identify the taxpayer. The disclosure risk arises when protected information enters the tool. Before submission, firms must assess approval for the intended use, collection and retention, model-training practices, access rights, and confidentiality and security controls.

What the Tax Court's first AI-hallucination case tells practitioners about real enforcement exposure

Clinco, T.C. Memo. 2026-16, is the Tax Court’s first case addressing potential AI hallucinations. The taxpayer’s attorney cited four cases to challenge the validity of a notice of deficiency, and the court described three as likely large language model hallucinations. Judge Mark V. Holmes warned that hallucinated authorities could cause an argument’s persuasiveness to collapse “like an overmixed soufflé”. The court stopped short of formal sanctions, but its public warning treated fabricated citations as a professional conduct problem.

In Fletcher v. Experian Information Solutions, No. 25-20086 (5th Cir., Feb. 18, 2026), the Fifth Circuit sanctioned counsel for unverified AI-assisted drafting and misleading responses to a show-cause order. It declined to adopt a special AI rule because existing sanctions doctrines already required accuracy and verification. In Withers v. City of Aberdeen (N.D. Mississippi, June 8, 2026), both sides’ pro hac vice lead counsel submitted fabricated AI-generated citations. The judge canceled the trial, revoked their admissions, barred them from the district for two years, and disqualified and fined local counsel.

These decisions converge on responsibility for the submitted work, regardless of the technology used to produce it. For tax practitioners, fabricated authority can also implicate IRC § 6694 preparer penalties for understatements on returns or refund claims. Circular 230 § 10.51(a)(13) separately addresses false opinions given knowingly, recklessly, or through gross incompetence. Depending on what the practitioner did, the risk may reach beyond a flawed filing and affect their professional standing.

Why the governance gap between AI adoption and firm policy is the immediate practical problem

The IRS has noted that virtually all professional tax firms use some form of AI, whether they recognize it or not. Leadership must account for the tools used in actual work and the controls governing them. Adoption can be visible in completed drafts while the absence of supervision remains hidden.

Under Section 10.36, firms must address this gap now to meet their duties. A firm must do more to meet its duty to supervise than expect each practitioner to use sound judgment on their own. Staff need defined rules on approved tools, client information, output review, and documentation of verification. Without those procedures, leadership has no adequate basis for treating AI-assisted work as a controlled firm process. The verification protocol is therefore an organizational requirement as well as an individual professional discipline.

The unresolved question practitioners are watching — client disclosure and the limits of current guidance

OPR Alert 2026-19 does not resolve whether practitioners must affirmatively tell clients that AI assisted in preparing their returns. IRC § 7216 governs disclosure of protected information to other parties, but the last formal IRS guidance predates the generative AI era. Practitioners consequently face a distinction between disclosing the use of a production tool and obtaining permission to share protected data.

The AICPA advises preparers to err on the side of caution and disclose AI use voluntarily. The counter-position holds that AI within secure, enterprise-approved infrastructure may resemble other third-party processing tools that have not historically required special disclosure. Under that view, the relevant event is sharing protected information. Neither position makes client disclosure a substitute for verification or adequate data controls.

There is also a limit to what disclosure alone can accomplish. An AI system can generate erroneous positions at scale, a problem that individual client notices may not address. That concern underlies calls for Treasury to clarify that the “reliance on others” defense under § 10.22(b) does not extend to generative AI. Practitioners should document AI use, obtain client consent before entering protected information into an AI tool, and disclose AI involvement when asked or when its role in producing advice is material.

Building a verification protocol that maps to the six obligations

An effective protocol follows the engagement from tool approval through billing rather than treating verification as a final proofreading pass. Before work begins, competence and supervisory review under §§ 10.35 and 10.36 establish which tools may be used and for what purposes. That review covers retained data, training use, access, and security controls, with consumer-grade tools prohibited for work involving client information. At intake, the firm confirms the approved use and obtains specific consent where § 7216 requires it before submitting protected information.

Marble is one tool built to answer complex federal and state tax questions with cited sources. It provides answers grounded in primary authority with every claim cited and can turn research into memos and client communications. Its intake, document review, and compliance-check functions also place tool assessment within the wider engagement workflow. The relevant selection criteria are cited authority, client-context controls, and suitability for tax research and drafting. Purpose-built design can reduce the review burden, but it does not transfer responsibility for the answer.

The next stage is substantive review under § 10.22. Embedded citations let reviewers move from a proposition to its supporting source without reconstructing the research from scratch. The practitioner still retrieves and reads each authority, tests its relevance, and establishes its current status. In parallel, calculation review reconciles the output to source numbers, while factual review tests both accuracy and completeness.

Before sending written advice, practitioners review it under § 10.37 to check that it rests on facts they have verified and legal reasoning they can support. This applies equally to a short email and a formal opinion letter. Before invoicing, § 10.27 review checks that charges reflect work actually performed and fairly credit AI-driven cost reductions. Documenting AI use and the review steps for each matter supports firm procedures and the practitioner’s defense if the work is later examined. These practices lower exposure and clarify professional value, with less emphasis on routine production and more room for advisory judgment.