The IRS and SCDOR Just Turned AI Audit Systems On. AI Agent Operators Should Pay Attention.

The March 2026 GAO report that found the IRS had lost much of its AI enforcement capacity was not wrong. It was just early. On April 16, 2026 — one day after the filing deadline — the IRS began running AI-assisted processing against this season's returns to flag discrepancies and route high-risk filers for further review. The same week, the South Carolina Department of Revenue confirmed that it will start using AI to help select business returns for audit in 2026, with the option to expand to individual returns and to reach back up to three years of prior filings.

That is a meaningful shift in the enforcement picture, and it happened faster than most commentators expected. I wrote three weeks ago that enforcement capacity would eventually return and that companies which banked on perpetual non-enforcement were making a bet on the future I was not prepared to make for them. The return has already started. This post walks through what is actually in place now, what it means for AI agent operators, and why the state-level developments matter more than the federal ones for most of our readers.

What the IRS deployed on April 16

The IRS's post-filing processing system is now crosschecking returns against the agency's own data holdings — bank interest reported by financial institutions, earnings reported by gig platforms and payment apps, 1099-DA proceeds from digital asset brokers, and information returns from every other source the IRS already receives. The AI layer identifies returns where the filer's numbers do not match what third parties reported, and it routes flagged returns for further review.

This is not new in concept. The IRS has been running matching programs for decades. What is new is the scale and the speed. The system runs continuously against the 2026 filing-season intake and can flag discrepancies within weeks rather than months. It also handles a broader set of information-return types, including the 1099-DA gross-proceeds reporting that started for 2025 transactions. Digital asset brokers furnished those forms for the first time this season, and the matching now runs against them.

For AI agent operators, the specific sources being matched matter:

• 1099-DA gross proceeds on digital asset transactions, including stablecoin transfers reported by custodial brokers

• 1099-NEC payments above the new $2,000 OBBBA threshold, reported by vendors that received large cumulative payments from an agent operator

• 1099-K payments processed through third-party payment networks, reverted to the $20,000 and 200-transaction threshold

• Bank and brokerage information returns that include digital asset and stablecoin activity on platforms that comply with 1099-DA

If your agent operates buyer accounts, seller accounts, or treasury functions that produce any of those reporting events, the IRS's AI layer is now matching them against whatever you filed as an entity. The matching is not optional and it does not discriminate between human-controlled and agent-controlled transactions. The EIN on the account is the EIN that takes the hit.

What South Carolina deployed

SCDOR's program is narrower in one respect and broader in another. Narrower: it is South Carolina only, and it starts with business returns. Broader: it is explicitly audit selection, not just return processing. The system runs inside SCDOR's closed-loop integration between FAST Enterprises' GenTax system and the FAS audit review platform. Nothing goes to external AI services, according to SCDOR. Returns flagged by the AI layer are routed to revenue agents for audit, and flagged businesses can be looked back up to three years.

This is the template more states will follow. The FAST Enterprises stack is in production at a majority of US state revenue agencies, and the closed-loop AI audit selection pattern scales cheaply from one state to the next. Once a state DOR sees an audit-yield uplift from AI-assisted selection, the budget case for expansion writes itself. Texas, California, New York, and Washington all have the underlying infrastructure to do what South Carolina just did, and all have active digital services tax programs that would benefit from better audit targeting.

Why this matters more at the state level than the federal level

State sales and use tax is where most AI agent operators have their real exposure, and state DORs are staffed and funded to enforce it. Washington's SB 5814 expansion, Maryland's 3% tech tax, Illinois's revenue-only economic nexus threshold, Maine's digital services expansion, and Chicago's 15% Personal Property Lease Transaction Tax are all in effect right now, and all of them apply to transactions that AI agents can easily trigger in the normal course of operation.

The IRS AI processing adds pressure on federal reporting — 1099-NEC, 1099-DA, income return accuracy. That pressure matters, but the federal income tax layer for most agent operators is lower-stakes than the sales tax layer. A company that miscategorizes a single SaaS subscription on its federal return absorbs a relatively small penalty. A company that fails to collect sales tax on thousands of agent-initiated transactions across multiple states is looking at cumulative assessments that can exceed the underlying revenue.

State AI-audit adoption changes the detection probability for exactly that kind of sustained under-collection. Before AI audit selection, a state DOR had to decide which businesses to audit from among thousands of candidates with limited staff. A business with agent-initiated transactions that crossed an economic nexus threshold but never registered could sit under the radar for years. After AI audit selection, the DOR scans every business return every cycle and surfaces anomalies automatically. Thresholds crossed without a registration event become a flag. Revenue growth without a corresponding sales tax filing becomes a flag. Large B2B categories with low collected-tax ratios become flags. The economics of ignoring sales tax shift.

The enforcement asymmetry is closing

Three weeks ago the narrative was that the IRS had dismantled its AI enforcement capability and that state DORs were the only serious threat. That narrative is already out of date. The IRS has reorganized and pushed AI processing into production, even if its long-term AI strategy is still constrained by the staff cuts the GAO documented. State DORs are adopting AI audit selection at a pace that was not visible a quarter ago. The combined effect is that the window in which AI agent operators could rely on enforcement latency as a compliance strategy is narrowing quickly.

This does not mean enforcement is perfect. It is not. The IRS AI system runs on known information-return feeds, which means transactions that never produce a 1099 are still invisible to the matching layer. State AI audit selection is as good as the data the state already has, and agent-initiated transactions that never cleared through a reportable channel still depend on the operator's own record-keeping. The enforcement layer has improved; it has not become omniscient.

But the direction of travel is one-way. Every quarter that passes, state DORs add more data feeds, more matching rules, and more audit-selection sophistication. The IRS rebuilds its AI team on the budget cycles that are already in motion. The gap between legal obligation and practical enforcement risk is closing, not widening. Operators who built their compliance posture around a temporary enforcement gap are about to find that the posture does not hold up.

What this means for AI agent builders specifically

Three concrete takeaways.

First, your reporting posture on the information-return layer — 1099-NEC, 1099-DA, 1099-K — needs to be defensible today. The IRS AI layer is matching this filing season. If your agent paid a vendor $2,001 in 2025 and you did not issue a 1099-NEC, the vendor's tax return will create a gap the matching system is designed to flag. The counterparty's filing is what triggers the mismatch, not yours, so you do not get to control whether it surfaces.

Second, your sales tax collection posture needs to hold up to an AI-selected audit rather than to an under-resourced human one. State AI audit selection favors businesses with the largest gap between expected and collected sales tax, and agent-initiated transactions are exactly the kind of activity that produces large expected-collection numbers. A platform that issues invoices through an agent, or accepts payment on an agent's behalf, or operates a marketplace for agent-to-agent commerce is generating the data that will train the next round of state audit models.

Third, voluntary disclosure windows become more valuable and probably shorter-lived. Most states offer voluntary disclosure programs that limit lookback periods and waive penalties for operators who come forward before the state identifies a problem. Those programs rely on the state not yet knowing about the operator. As AI audit selection improves, the window in which a VDP is available is going to shrink. Operators considering whether to register in a state where they have accumulated exposure should factor in the rising detection curve.

The bottom line

The IRS enforcement gap I wrote about three weeks ago is real, but it is shorter than many commentators assumed. AI audit deployment is happening simultaneously at the federal and state levels, and the state level is where the practical exposure sits for AI agent operators.

The compliance math has always favored getting ahead of enforcement rather than betting against it. That math has not changed. What has changed is the size of the bet and the time you have to make it.

AgentTax is the tax compliance layer for AI agents. Calculate transaction tax in real time across all 51 US jurisdictions. Sign up free or explore the AI agent sales tax landing page for state-by-state coverage.

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