AI, due diligence and the new IRS

The Internal Revenue Service that practitioners face in 2026 is not the IRS of 2024. The agency started 2025 with about 102,000 employees. It finished the year with about 74,000. That is a 27% reduction, concentrated in experienced enforcement and technical staff. The National Taxpayer Advocate confirmed those numbers in her 2025 Annual Report to Congress.

The Independent Office of Appeals lost more than a quarter of its people. Back-office employees with no tax-administration experience were involuntarily detailed into frontline filing-season roles. IRS CEO Frank Bisignano told the Senate Finance Committee on April 15 that the 2026 filing season met its targets with "less people and better results." He committed the agency to a digital-first model. Twelve days later, the House Appropriations Committee advanced a smaller IRS budget for fiscal 2027 and expanded the agency's use of AI and data analytics for enforcement.

At the same time, the One Big Beautiful Bill Act, signed July 4, 2025, made more than 100 Code changes. Many are retroactive to 2025. SALT caps, no-tax-on-tips and overtime deductions, expanded child tax credits, and modified thresholds all require new forms, new systems and new guidance. The Treasury Inspector General for Tax Administration flagged the combination in its October 2025 Major Management Challenges report.

Practitioners are turning to AI to manage the load. That is rational. It is also where the ethics line moves.

The rules already exist

Tax professionals do not need new ethics rules to govern AI. The existing ones do the work.

Circular 230 Section 10.22 requires due diligence in preparing returns, advising clients and representing taxpayers before the IRS. That includes the correctness of representations made to the Treasury. When an AI tool generates a citation, a procedural recommendation, or a legal analysis, the output is a representation the practitioner intends to rely on. Sending it to the IRS without independent verification is a §10.22 failure. The technology does not change that.

Section 10.22(b) is more pointed. A practitioner may rely on the work of another only by exercising reasonable care in engaging, supervising, training and evaluating that work. The Office of Professional Responsibility has said for years this is an affirmative duty. Willful blindness does not satisfy it.

The Treasury and the IRS proposed substantial Circular 230 amendments on Dec. 20, 2024 (REG-116610-20). Those amendments would modernize the competence and technology provisions. They remain proposed as of this writing. Track the rulemaking. Do not wait for it. The current rules are sufficient.

Other provisions reinforce the framework. Section 10.35 requires the legal knowledge, skill and thoroughness needed for the matter. Section 10.37 prohibits reliance on written advice the practitioner knows or should know is unreliable. Sections 10.33 and 10.36 require firm-level quality controls. AICPA Statements on Standards for Tax Services Section 1.4, effective Jan. 1, 2024, address tool reliance directly. Software does not relieve professional responsibility. The member must evaluate whether the output fits the engagement.

Apply Section 10.22(b) to the tool, not just the person

The four-part test translates cleanly to AI.

Engaging: Evaluate the system's knowledge architecture before deployment. A tool that pulls from a single source, or from opaque training data with no verification layer, fails this element. Reasonable care requires evidence that the system cross-references curated tax content against authoritative government sources in real time.

Supervising: Spot-checking conclusions is not enough. The tool must distinguish verified primary authority from unverified, training-derived content. Without tiered source labeling, supervision is an empty exercise.

Training: Confirm that the tool verifies every IRC section, regulation and Internal Revenue Manual cited against an official source before presenting it. The tool must also detect when prior authority has been voided or modified. A system that surfaces a Treasury Decision without recognizing that Congress disapproved it is not just outdated. It is dangerous.

Evaluating: Demand an explicit authority hierarchy. The Code, Treasury Regulations and case law are binding. IRS publications, practitioner commentary and AI-generated reasoning are not. A tool that treats all sources as equivalent fails Section 10.22(b) no matter how confident the prose.

Where AI fails the standard

The risks are not theoretical.

Courts have sanctioned lawyers for AI-fabricated citations. Mata v. Avianca was the first widely reported case. The list has grown, and the consequences have escalated. 

On May 5, 2026, the Supreme Court of Georgia suspended Clayton County Assistant District Attorney Deborah Leslie for six months after a filing in a murder appeal was found to contain five citations to cases that do not exist, five more unsupported citations, and three fabricated quotations. Leslie initially claimed the filing had been altered, then admitted she had used AI. The Clayton County District Attorney filed a grievance against her own prosecutor with the State Bar of Georgia. The Leslie suspension is reported to be the first suspension in the United States tied directly to AI-generated fabrications in a court filing. It will not be the last. 

AI-related sanctions across U.S. courts have been reported at roughly $145,000 in the first quarter of 2026. Oregon courts have begun assessing $500 per fabricated citation. In Whiting v. City of Athens, Sixth Circuit counsel were sanctioned more than $30,000 for fake AI-generated citations. A reported 61.6% of federal judges now use AI tools themselves, which means they know exactly what a hallucinated case looks like. In tax practice, a fabricated authority used to support a return position or a representation to the IRS triggers IRC Section 6694 preparer penalties and Circular 230 §10.51(a)(13) for false opinions through gross incompetence. The National Taxpayer Advocate has told practitioners not to rely solely on AI-generated tax advice. The IRS has added misleading AI-generated content to the Dirty Dozen.

The problem is not that attorneys are using AI. The problem is they are using consumer-grade AI with no verification layer. ChatGPT is not a legal research system. It is an autocomplete engine with a law school vocabulary. The sanctioned practitioners are not bad lawyers. They are practitioners without a process.

The risk is bigger than fabrication. It includes misidentified authority. The IRS DeFi broker reporting regulations, T.D. 10021, were issued in late 2024. Congress disapproved them under the Congressional Review Act through Pub. L. 119-5, signed April 10, 2025. A research tool that cites those regulations as operative law today is producing advice grounded in a rule that no longer exists. Tools without a mechanism to track legislative and administrative voidance cannot satisfy the evaluating element. Period.

Data privacy is the next exposure. Section 7216 of the Internal Revenue Code prohibits unauthorized disclosure or use of tax return information. There is a criminal penalty of up to $1,000 per offense and up to one year imprisonment, and a civil penalty under Section 6713 of $250 per disclosure, capped at $10,000 per year per preparer. Entering identifiable client data into a general-purpose AI platform may trigger both provisions if the platform's terms permit input data to train the model. The distinction between secured, purpose-built environments and general consumer platforms is not technical. It is ethical.

There is a subtler failure mode. Call it the competence illusion. General-purpose AI produces fluent, authoritative-sounding text whether the analysis is right or wrong. The practitioner feels competent. The practitioner is violating Section 10.35. A junior associate's errors look like errors. AI errors look like expertise. That is what makes them more dangerous, not less.

Controversy is where the stakes peak

Advisory and return-prep mistakes can usually be amended. Controversy mistakes often cannot.

A missed 30-day Appeals window forecloses Appeals. A failure to raise an issue at exam can waive it. An inadequate Collection Due Process hearing request can cost the taxpayer judicial review. An untimely response to a statutory notice of deficiency ends the Tax Court door. Procedural errors in controversy are structurally more severe than errors anywhere else in tax practice. AI overreliance compounds that risk.

Appeals is itself under strain. The National Taxpayer Advocate's Fiscal Year 2026 Objectives Report devotes an entire objective to Appeals independence. The report calls for training that emphasizes a judicial and impartial approach. It warns that compliance-oriented performance pressures threaten to turn Appeals into an extension of exam. Practitioner commentary expects the IRS to bypass Appeals more often by issuing statutory notices of deficiency, pushing taxpayers into Tax Court.

Congress is responding. On May 18, 2026, the House passed a package of eight bipartisan tax administration bills, including H.R. 6506, the Taxpayer Due Process Enhancement Act. The bill strengthens CDP, protects refunds, and expands judicial review of tax liability claims. The Senate has not acted. The direction is clear. The IRS is leaning harder on automation, including improved identity-theft filters announced May 18, 2026. Congress is reinforcing procedural rights at the back end. The controversy practitioner sits in the middle.

The ethics analysis turns here. When IRS personnel are undertrained, when automated notices proliferate, and when Appeals is understaffed and under throughput pressure, the practitioner who refuses to use tools that track deadlines, reconcile notice language against the Internal Revenue Manual, or surface overlooked procedural options may fail the reasonable care standard. The duty runs in both directions. Do not rely blindly on AI. Do not ignore tools that demonstrably improve diligence when the system around the client is degraded.

Purpose-built controversy platforms have emerged. TaxAdvisor 360, CoCounsel Tax and Blue J are examples of domain-specific tools built around IRS procedural workflows rather than general-purpose chat. The ethics question for any of them is the same. Does the architecture meet §10.22(b)? Does the source transparency meet SSTS 1.4? Do the security controls meet §7216? Answer those before signing the contract.

What practitioners should do now

First, write a firm AI policy. Identify approved tools. Identify prohibited tools. Identify what client data may go where. Document it.

Second, apply the §10.22(b) four-part test to every tool. Engaging, supervising, training, evaluating. If a tool fails one element, do not use it for client work.

Third, verify every AI-generated authority against a primary source before it leaves the firm. Every citation. Every quote. Every Treasury Decision. Every IRM section. No exceptions.

Fourth, segregate controversy work. Procedural deadlines and notice responses require human sign-off. AI can draft. The practitioner files.

Fifth, train the firm. Section 10.36 requires it. Document the training.

Sixth, monitor the rulemaking. REG-116610-20 will likely become final. The Taxpayer Due Process Enhancement Act may pass the Senate. The Circular 230 floor and the procedural ceiling are both moving.

AI is a powerful leverage tool for tax professionals. It is not a substitute for them. The IRS is smaller, more automated and less forgiving than it was two years ago. The client needs the practitioner's judgment more than ever. The ethics rules already say so. Apply them.