If you earned seven figures last year and filed a clean return, you might still hear from the IRS. The agency is now running 125 artificial intelligence models to screen tax returns for potential problems, more than double the 54 it operated just two years ago, according to a Government Accountability Office inventory analysis published in early 2026. Each model scores, sorts, or flags returns before a human examiner ever opens the file.
That expansion is landing alongside a deliberate push to audit more high-income taxpayers. The IRS’s fiscal year 2024 Data Book confirmed that examination activity rose during the year, and a separate Treasury Inspector General for Tax Administration (TIGTA) audit report verified that high-income individual examinations specifically increased in FY 2024. For filers across the income spectrum, the math has changed: more algorithms watching, more agents following up.
How 54 models became 125
Under Section 10.24.1 of the Internal Revenue Manual, the IRS catalogs every model, algorithm, and dataset tied to an AI use case. That governance framework also restricts how personally identifiable information and federal tax information can move through automated systems. But the GAO found that the agency’s model count more than doubled without a proportional increase in the data-science workforce needed to manage those tools. The report flagged skills gaps, information-quality shortfalls, and weak strategic oversight as risks that could undermine the accuracy of automated screening.
Critically, the agency has not published model-by-model performance data or error rates. That means outside observers, including Congress, cannot yet tell whether the jump to 125 models has sharpened audit targeting or simply increased the volume of flagged returns without a corresponding gain in precision.
High-earner audits: what the data actually shows
The trend line for wealthy filers has been dramatic. A GAO analysis (GAO-24-106112) documented how audit rates for taxpayers earning $1 million or more fell sharply after 2010, dropping from roughly 8 percent to below 2 percent as the agency lost thousands of experienced revenue agents to attrition and budget cuts.
Rebuilding that capacity became a stated priority after Congress approved roughly $80 billion in new IRS funding through the Inflation Reduction Act in August 2022. But the path has not been smooth. The Fiscal Responsibility Act of 2023 clawed back a portion of that funding, and workforce disruptions in early 2025, including agency-wide buyout offers and organizational restructuring, further complicated hiring timelines.
Still, the numbers are moving. “We increased examinations of high-income individuals in fiscal year 2024, and we intend to sustain that trajectory,” an IRS spokesperson said in a statement accompanying the Data Book release. TIGTA confirmed the uptick, though it also found that the IRS had not fully defined key terms and methodologies required by a 2022 Treasury directive on wealthy-filer compliance, making precise comparisons to pre-2010 rates difficult.
Tax professionals say the shift is already tangible. “I have three clients with seven-figure adjusted gross incomes who received correspondence audits in the last six months, after years of seeing almost none,” said a CPA at a mid-size firm who asked not to be named because of ongoing client matters. The practitioner added that two of the three audits focused on Schedule E rental-loss deductions and one on charitable contribution substantiation, categories that align with known IRS enforcement priorities for high-income filers.
Current rates have not matched the 2010 peak, and the exact gap depends on income thresholds and counting methods that remain unsettled. But the direction is clear: examination rates for high earners are climbing back from their post-cut lows.
What the IRS still has not disclosed
For all the new firepower, several important pieces of the picture remain outside the public record as of mid-2026:
- Model specifics. The IRS has not explained what each of the 125 AI models does, which return types they target, or how they interact with programs focused on high-wealth filers, partnerships, or large corporations.
- Accuracy metrics. No error rates, false-positive rates, or outcome data have been released for any individual model. Without these, it is impossible to assess whether flagged returns actually yield additional tax owed.
- Staffing breakdown. The Data Book shows overall headcount recovering from earlier lows, but the agency has not disclosed how many data scientists, machine-learning engineers, and senior revenue agents it currently employs versus how many it needs to responsibly manage 125 models.
- Routing of flagged returns. Whether AI-flagged returns are reviewed by experienced examiners or routed through less-experienced staff is not addressed in any public document.
These gaps matter because the GAO’s central concern is not that the IRS is using AI. It is that the agency is scaling AI faster than it can demonstrate responsible management of those tools.
What about everyone else?
High earners are the stated priority, but AI models do not screen only seven-figure returns. The IRS has historically used automated tools to flag discrepancies on returns claiming the Earned Income Tax Credit, and research from Syracuse University’s Transactional Records Access Clearinghouse (TRAC) has shown that low-income EITC claimants have at times faced audit rates comparable to those of top earners. Whether the new models shift that balance, for better or worse, is an open question the agency has not addressed publicly.
For any filer, the most reliable defense against an automated flag remains clean documentation. The models are trained on patterns of discrepancy between what a return reports and what the IRS already knows from W-2s, 1099s, and other information returns. Matching every line on a filed return to supporting records reduces the chance of a false-positive selection. Taxpayers who receive an IRS notice or audit letter can check their account status or request help through the agency’s online account portal.
A bigger arsenal, an unfinished rulebook
The IRS is deploying AI at a pace that, by the GAO’s own assessment, exceeds its demonstrated ability to govern those systems. If the agency closes the skills and quality gaps that auditors have identified, the result could be a more efficient and more equitable enforcement operation, one that catches sophisticated tax avoidance without burdening compliant filers. If it does not, a larger model inventory may generate more audit activity without proportionally better outcomes for the Treasury or for the taxpayers swept into the process.
That is the tension sitting underneath every return filed this season: a technologically upgraded IRS that has more tools than ever, and is still writing the rules for how to use them.
