Somewhere inside the IRS, a machine-learning model is comparing the income you reported on your 2025 tax return against every W-2, 1099, and third-party record the agency can match to your Social Security number. If the numbers don’t line up, another model scores the discrepancy. A third may check whether your return resembles patterns linked to identity theft rings. None of these systems existed a decade ago. Today, the IRS operates more than 125 of them.
That figure comes from a Government Accountability Office audit completed in late 2025 and published as GAO report 26-107522. Auditors cataloged 126 active AI use cases across the agency, more than double the roughly 54 tracked in federal AI inventories just two years earlier under the reporting requirements of Executive Order 13960. The systems screen returns for fraud, flag underreported income, route taxpayer correspondence, and forecast call-center demand. Together, they form the most extensive automated filter the IRS has ever applied to the approximately 150 million individual returns filed each year.
At the same time, multiple congressional proposals for fiscal year 2026 would strip roughly $1 billion more from the IRS budget, extending a pattern of cuts that began when lawmakers started clawing back Inflation Reduction Act enforcement funding in 2023. The agency’s own FY 2026 budget documents request stable funding, but as of late May 2026, appropriations bills carrying the proposed reductions had not reached final passage. The trajectory is clear even if the final number is not: the IRS is deploying more automated screening tools than ever while facing the prospect of fewer people to oversee what those tools produce.
What the AI models actually do
A companion GAO inventory released under identifier GAO-26-108418 breaks the non-classified use cases into categories. The largest cluster falls under compliance and fraud detection, the functions most likely to touch ordinary filers. Some models run automated matching, comparing reported wages and investment income against employer and brokerage filings to spot underreporting before a human reviewer ever opens the case. Others use pattern recognition to identify suspicious clusters of returns, a technique the IRS has relied on to dismantle identity-theft and refund-fraud rings that file thousands of bogus returns in a single season.
A smaller set supports internal operations: prioritizing case assignments so that the most complex returns reach experienced examiners first, routing taxpayer correspondence to the right department, and forecasting how many calls will hit IRS phone lines on a given Monday in April.
What the public inventory does not reveal matters just as much. A subset of AI systems remains classified, meaning the specific algorithms that influence audit selection, their error rates, and the data inputs they consume are shielded from outside review. That opacity is significant because false positives, returns incorrectly flagged for scrutiny, impose real costs on taxpayers who must then respond to notices, assemble documentation, and sometimes wait months for resolution. The National Taxpayer Advocate has repeatedly warned in annual reports to Congress that automated notices already generate confusion and burden, particularly for lower-income filers who lack access to professional tax help.
The governance framework on paper
The IRS has built a formal structure to manage its expanding AI portfolio. Internal Revenue Manual section 10.24.1 requires every business unit to maintain inventory entries for each model and dataset tied to an AI use case. The same manual section lays out how the agency reports its AI systems upward to the Treasury Department and the Office of Management and Budget. Treasury’s own budget and performance reporting portal shows the pipeline is active, with consolidated AI use-case data updated as recently as January 2026.
On paper, the framework looks thorough. In practice, the GAO found gaps. The audit’s full title tells the story: “Artificial Intelligence: IRS Actions Needed to Address Skills Gaps, Information Quality, and Strategic Management.” Auditors concluded that the agency lacks enough staff with the technical expertise to evaluate whether its own models are performing accurately. Some inventory records contained incomplete or outdated information about the data feeding those models, raising questions about whether anyone inside the IRS can say with confidence how well a given algorithm is working or whom it is flagging most often. The GAO issued formal recommendations; the IRS agreed to address them but, as of May 2026, has not published a timeline for doing so.
Budget pressure and the staffing bind
The skills-gap warning hits harder against the budget backdrop. The IRS received a historic $80 billion funding boost through the Inflation Reduction Act in 2022, but Congress began pulling that money back almost immediately. By early 2025, lawmakers had redirected or rescinded more than $20 billion of the original allocation, according to Congressional Budget Office estimates and reporting by the Associated Press. The current round of proposed cuts, totaling roughly $1 billion for FY 2026, would further shrink the agency’s capacity to hire and retain the data scientists, engineers, and compliance officers needed to keep human judgment in the loop as AI screening scales up.
No public statement from IRS leadership directly connects a specific budget scenario to a specific AI deployment outcome. That silence is worth noting. The agency’s Budget in Brief describes planned spending. The IRM governance manual describes how models are tracked. Neither document addresses what happens to AI oversight if the workforce contracts. The GAO’s warnings remain the closest thing to an official acknowledgment that the current path carries risk, and those warnings come from an outside auditor, not from the agency building and running the systems.
Who is most likely to feel the effects
The IRS has not published demographic or income-bracket breakdowns of which returns its AI models flag most frequently, and the classified portion of the inventory makes independent analysis difficult. But the general mechanics of automated matching offer some clues. Filers with complex returns, multiple income streams, gig-economy earnings reported on numerous 1099-K and 1099-NEC forms, or large itemized deductions generate more data points for models to compare and, statistically, more opportunities for a mismatch to surface. Self-employed taxpayers and small-business owners, whose income is not withheld and verified at the source the way a salaried worker’s paycheck is, have historically faced higher correspondence-audit rates, a pattern that AI-driven screening is likely to reinforce rather than correct.
At the other end of the spectrum, the IRS has said publicly that IRA funding was partly intended to increase audit coverage of high-income individuals and large corporations, groups that had seen audit rates plummet over the prior decade. Whether AI models are meaningfully shifting enforcement attention toward those filers, or simply adding volume to the existing pool of automated notices sent to lower- and middle-income households, is a question the current public data cannot answer.
What to do if an AI-flagged notice lands in your mailbox
For anyone filing a 2026 return, the practical reality is this: a wider, more automated filter is scanning returns than at any prior point in IRS history. More models reviewing more data means more returns will pass through some form of algorithmic screening before a human ever looks at them. That does not automatically translate into more full-scale audits. Many flagged returns are resolved through automated matching or a simple correspondence notice, often a CP2000, asking the filer to confirm or explain a specific line item.
But the volume of those notices could rise, and the agency’s capacity to resolve disputes quickly could fall if staffing does not keep pace with the technology. If you receive unexpected IRS correspondence, verify its authenticity through irs.gov or by calling the number printed on the notice. Respond within the stated deadline. Keep organized records of income, deductions, and supporting documents throughout the year, not just at filing time. And understand that the system generating these contacts is growing faster than the human workforce assigned to review its output. That imbalance, a government agency automating enforcement while losing the staff to supervise the automation, is the core tension neither the IRS nor Congress has resolved.
