The Practitioner Review Layer That Determines Whether AI-Drafted IRS Responses Hold Up Under Scrutiny

AI can draft client memos and IRS notice responses faster, but those documents need the practitioner’s review to stand up to close checks. For tax firms, the relevant tool criteria are traceable citations, verified tax sources, confidentiality safeguards, and fit with IRS correspondence work. Those features support review rather than replace it, because professional judgment must catch errors that a fluent draft can conceal.

Why both sides of an IRS enforcement encounter are now AI-assisted

AI now operates on both sides of an IRS enforcement encounter. The IRS uses it to flag compliance issues and generate notices, while practitioners use it to draft responses. That creates a feedback loop in which machine-generated correspondence can receive a machine-generated answer. The practitioner review layer is the human checkpoint between them. Its function is to decide whether the response addresses a sound enforcement premise, accurately represents the client, and advances a supportable position.

The resource asymmetry makes that checkpoint more consequential. AI flags compliance issues faster than ever before in IRS history, while the staff and support available to resolve them have shrunk. An automated flag does not replace an expert’s examination of books and records or discussion with a taxpayer. Faster drafting on the practitioner’s side likewise does not establish that a proposed answer resolves the issue the IRS raised.

A response can be polished yet accept a mistaken premise in the notice or answer a question the IRS did not ask. Review must therefore connect the two documents to the underlying records and law. Without that connection, speed can accelerate an error and delay its resolution.

What Circular 230 actually requires when AI drafts the response

The IRS Office of Professional Responsibility’s OPR Alert 2026-19 applies existing Circular 230 obligations to AI-assisted work. Under §10.22, due diligence requires thorough review of AI-created documents before delivery to a client or submission to the IRS, including verification of facts, citations, and calculations. The duty remains with the practitioner who uses the output. AI authorship does not transfer responsibility to the software. A document’s apparent completeness is therefore no substitute for examining its foundation.

A meaningful examination requires competence. Section 10.35 requires practitioners to understand how AI develops content, recognize potential bias and errors, and evaluate whether the output is suitable for an IRS matter. Technological illiteracy can itself create disciplinary exposure. Section 10.37 then governs the substance of written advice by requiring independent verification of AI projections, representations, and citations. Where the system’s underlying logic is opaque, reliance on its output may itself be unreasonable.

Section 10.36 reaches leadership that willfully, recklessly, or through gross incompetence fails to establish adequate compliance procedures, including staff training, internal protocols, and third-party vendor vetting. Procedures make competent review a consistent requirement. Section 10.27(a) extends the obligation to billing, because charging for manual labor time that AI eliminated may constitute an unconscionable fee. The alert states that cost savings should be passed on openly through billing practices that reflect the efficiencies gained.

What hallucinations look like in an actual IRS proceeding

In documented hallucination incidents, correspondence and filings presented made-up legal authorities with the same confidence as real ones. The defects were identified in the correspondence and filings only after submission. In Clinco, the review failure compounded when the attorney continued after the IRS flagged the problem. The initial error exposed inadequate verification. Continuing after notice of the defect turned a missed error into a failure to respond to a direct warning.

Judge Holmes used the phrase “like an overmixed soufflé.” The image was apt for a matter involving an attorney who also owned a restaurant, but its professional significance lies in the collapse under scrutiny. A convincing draft can hold its shape until someone asks whether its authorities are real and back up its claims.

The alert cites real-world hallucination incidents as warnings about professional responsibility, although Clinco was a Tax Court conduct issue rather than an OPR disciplinary proceeding. Those are different ways to hold practitioners to account for the same failure. Fabrication and accuracy can look identical in AI output. Independent verification requires a reviewer with access to primary sources and the judgment to assess what they establish.

How can tax firms use AI to respond to IRS notices faster without narrowing review to citations?

Review the draft against the authorities, records, and original notice as well as its own internal logic. Every cited case, revenue ruling, Treasury regulation, or IRS publication must be pulled and read. Existence is only the first question to resolve. A real authority can still be used for the wrong proposition, so the reviewer must establish that it supports the sentence in which it appears. Citation verification without that substantive comparison leaves the legal argument untested.

The same discipline applies to the factual account. The reviewer must compare the draft with the client’s actual records, including any calculations, rather than accept a plausible account assembled from the inputs. A plausible account is not the client’s account. The original notice or examination request letter must also be read to establish precisely what the IRS seeks. A response based on a familiar notice template can miss a specific request even when what it says about tax in general is correct.

Practitioners who specialize in IRS representation stress the need for a partner to review the draft because the wrong word can cause an audit to expand. The reviewer must examine language that could open unintended issues and look for omissions, including client-specific facts and arguments that should appear but were never requested from the AI and therefore never entered the draft.

Confidentiality reaches back to the inputs, before any response exists. IRC §§6713 and 7216(a) and Circular 230 §10.51(a)(15) make uploading client data to unsecured or public AI platforms a potential violation. A clean final document cannot repair that exposure. The reviewer must also check that the drafting tool is a secure, enterprise-approved system with strong confidentiality safeguards. The alert treats an AI-drafted document as a starting point.

How the IRS's own review failures complicate the practitioner's position

The strongest objection is that practitioners face rigorous review duties while the agency’s AI-assisted enforcement systems have documented review gaps. The March 2026 GAO report found that 61% of the IRS’s 126 AI applications were still in development and identified unintentional bias as a risk. Those findings warrant scrutiny of enforcement outputs. They do not establish that every notice is flawed or that a particular notice resulted from biased processing. They do establish why the agency’s position cannot substitute for independent examination of the client’s facts.

That uncertainty increases the work required of the practitioner. Review must test the incoming position as well as the outgoing response, because a well-supported answer may need to challenge the notice’s premise when addressing its requests. The agency’s gaps are not a reciprocal exemption. A practitioner who documents a defect and responds with a substantiated position has exercised judgment that is absent when someone simply forwards an AI draft. The latter conduct creates Circular 230 exposure regardless of the quality of the IRS’s review.

Practitioners can use paper trails from the agency’s internal procedures proactively when examining the basis of an AI-assisted enforcement position. Using those paper trails to develop a supported response keeps the focus on the evidence. A challenge grounded in the record gives the practitioner something concrete to defend, while a general objection to IRS use of AI does not resolve the client’s issue.

What distinguishes a review process that provides real protection from one that provides the appearance of it

A protective process records what was verified, who verified it, and which primary sources supported the decision at the time of review. A later note that someone reviewed the document provides much less evidence of professional judgment. Specificity gives the review log its value. If a position is questioned, the record should show the basis for the human decision rather than merely indicate that an approval step occurred. Documentation cannot replace sound analysis, but it can distinguish that analysis from unexamined acceptance of machine output.

The drafting and reviewing functions should also be separated. The practitioner who selected an AI draft should not be its sole reviewer, because familiarity with that choice can make genuine scrutiny harder. A second review must still be substantive. Its purpose is to test the authorities, factual account, scope, language, and omissions rather than endorse the first practitioner’s selection. That separation supports the genuine review that firm procedures under §10.36 must establish.

Under the same provision, firms must examine third-party AI tools before use, and a purpose-built tax workflow tool grounded in IRS publications, Treasury regulations, and verified case law with traceable citations is better positioned to support defensible review than a general-purpose tool trained on open internet data. Traceable source material gives the reviewer a basis for verification, while substantive judgment remains necessary to decide whether an authority applies to the client’s facts.

Leadership needs a record that staff have been trained on AI risks, limitations, and the firm’s review protocols, because §10.36 can reach failures to establish those procedures. An undocumented expectation is not a dependable process. Several courts have also issued standing orders that require attorneys to certify that they independently checked AI-generated research against authoritative sources. That judicial pattern reinforces the practical need for evidence that verification occurred, rather than an assurance that the output looked reliable.

The process must finally account for AI’s contribution to the bill. Practitioners should record what AI did and base fees on the time actually spent, leaving out charges for manual labor the software took over. This is a separate compliance question from document quality. A thoroughly verified response does not resolve a fee problem created by billing for eliminated work.

Why the review layer is where AI in tax practice either earns its place or fails the practitioner

The defensible gain from AI is faster drafting with retained professional accountability. A practitioner who builds review into the engagement can reduce routine work while preserving a documented basis for the position taken. A cursory approval leaves the practitioner responsible for errors with no evidence that someone independently checked the authorities, facts, and choices about how to respond.

As AI handles more intake, document categorization, and initial drafting, the practitioner’s time concentrates on the decisions those tasks cannot settle. Those decisions include whether an authority fits, whether a notice is based on an error, and whether wording that is accurate broadens the scope more than needed. This is the work review must preserve. Building review competency now supports the complex advisory judgment that remains human, rather than allowing drafting speed to displace the very work for which the client needs a practitioner.