Congress Wants the IRS to Audit With AI: What H.R. 9501 Means for Agent Operators

On July 1, 2026, the House Committee on Ways and Means favorably reported H.R. 9501, the "AI Tax Integrity Act of 2026," by a vote of 40-0. Introduced by Rep. Vern Buchanan (R-FL), the bill directs the Treasury Department to stand up an artificial intelligence pilot program to identify inaccurate tax returns. It is a small bill with a large signal: the enforcement apparatus is being wired for AI faster than the substantive rules governing AI-driven commerce are settling.

What the bill actually requires

The operative text is narrow and administrative. Treasury must establish a pilot program — within 180 days of enactment — that uses AI to identify inaccurate returns. The categories named in the bill are specific: returns tainted by identity theft, fraudulent claims for credits, deductions, or refunds by individual or business taxpayers, and returns that were improperly prepared by third-party return preparers. The pilot runs for no less than 18 months and no more than two years, and the Government Accountability Office must report back to the House Ways and Means Committee and the Senate Finance Committee on how it performed — including the aggregate amount of improper refunds detected and the accuracy of the tool.

The Joint Committee on Taxation scored it as having a negligible effect on federal receipts through 2036. That is the tell. Congress is not expecting this pilot to raise revenue directly. It is building capability and a reporting record — the scaffolding a larger deployment sits on later. H.R. 9501 did not move alone. It cleared committee inside a seven-bill package on taxpayer rights and IRS administration, alongside H.R. 9499, the "Protecting Taxpayers from Ghost Preparers Act," which also passed 40-0. Read together, the package points at the same target: the correctness and provenance of return preparation, at machine scale.

Why an agent operator should read a return-fraud bill

At first glance this has nothing to do with sales and use tax on autonomous software. It is an income-tax-administration measure aimed at refund fraud and bad preparers. But the relevance is not in the subject matter — it is in the direction of travel.

We wrote earlier this year about the GAO's finding that the IRS had fallen behind on AI-driven enforcement — that the agency had the data but not the tooling to act on it at scale. H.R. 9501 is the legislative answer to that gap. It is Congress instructing Treasury to close it, on a clock, with a mandated report card. Pair it with the IRS Office of Professional Responsibility's Circular 230 AI guidelines and the state-level AI audit deployments already underway, and a pattern is unmistakable: revenue authorities are systematically acquiring the ability to review returns, positions, and preparer conduct with automated tools.

For anyone operating an AI agent, an agentic workflow, or an autonomous commerce product, that matters for two reasons.

First, your own filings are about to be read by a machine. If your operation is generating income-tax returns, credit claims, or refund positions — directly or through a preparer — the "inaccurate return" net this bill authorizes is being cast at exactly the kind of high-volume, automated activity that agent businesses produce. The improperly-prepared-return category and the companion Ghost Preparers Act are aimed at preparation that lacks a clear, accountable human signer. AI-assisted preparation without proper attribution is precisely the fact pattern in the crosshairs. If an agent or a downstream tool is touching return preparation, the identity of the preparer and the diligence behind the position need to be documented, not assumed.

Second, the asymmetry is widening. The substantive law governing agent commerce — whether an x402 micropayment is a sale, whether an agent operator is a marketplace facilitator, how a bundled digital-plus-services offering is classified state by state — remains genuinely unsettled. We have said so repeatedly, and nothing in the last two weeks changed it. What H.R. 9501 illustrates is that the enforcement side does not wait for the substantive side to resolve. The government is building AI to detect discrepancies and improper claims now, while the operator still bears the ambiguity about what the correct position even is. You do not get to plead that the rules were unclear if the tooling flags an underpayment; you get an assessment, and you argue about it afterward. That is the worst side of the ledger to be on.

The practical compliance move

None of this is cause for alarm, but it is cause for discipline. Three concrete steps for agent operators:

Keep a defensible record of your tax positions. When you take a classification position on an ambiguous transaction — an API micropayment, a bundled onboarding fee, an agent-mediated sale — write down why, and cite the authority you relied on. Automated review rewards clean, documented positions and punishes the ones that look like a guess. A position you can explain survives a flag; a position you cannot explain becomes an adjustment.

Know who your preparer is, in the legal sense. If any part of your return preparation runs through automated tooling, make sure a competent human is accountable for the output and that the preparer attribution is correct. The ghost-preparer and improperly-prepared-return provisions are aimed squarely at preparation that no identifiable, responsible person stands behind.

Reserve conservatively on the unsettled items. When the law is genuinely open, the presence of AI-assisted enforcement is one more reason to set aside the higher of the plausible outcomes rather than the lower. Automated systems do not extend the benefit of the doubt; they surface the discrepancy and leave the doubt for you to litigate. Reserving for the conservative answer is cheaper than financing an assessment later.

What to watch next

H.R. 9501 still has to clear the full House, move through the Senate Finance Committee, and be reconciled before it is law — a 40-0 committee vote signals momentum, not enactment. Watch for floor action and for whether the Senate attaches anything broader to the vehicle. The more consequential signal is the mandated GAO report: 18 to 24 months after the pilot launches, we will have the first federal data on how accurately AI flags improper returns and how much it recovers. If those numbers are good, the pilot stops being a pilot, and the "not less than 18 months" language becomes the opening chapter of permanent AI-driven return review. Agent operators should assume the machine reads everything, and file accordingly.

See how AgentTax computes and documents per-state, per-transaction tax positions for AI agents — so when the enforcement tooling reads your record, there is a defensible position on the page, not a guess.

This analysis is for informational purposes only and does not constitute legal or tax advice. Consult a licensed tax professional for compliance decisions.