If you reported more than $10 million in income on your most recent tax return, the IRS is screening it with tools that barely existed two years ago. The agency now operates 125 active artificial intelligence systems, according to a Government Accountability Office review of IRS artificial intelligence use. That is more than double the 54 AI use cases the agency disclosed in its publicly released 2024 AI use-case inventory. A large and growing share of those systems focus on tax compliance, fraud detection, and deciding which returns deserve a closer look from human examiners.
For wealthy filers and the professionals who advise them, the shift is not theoretical. Examination coverage rates for the highest income brackets are trending upward after a decade of decline, and the algorithms selecting audit candidates are multiplying faster than most taxpayers realize.
How the IRS built its AI arsenal
The buildup traces directly to money. The Inflation Reduction Act of 2022 directed tens of billions of dollars toward IRS enforcement, technology upgrades, and hiring. That funding allowed the agency to move AI out of scattered pilot programs and into production-scale systems woven into daily operations.
The GAO’s inventory sorts IRS AI deployments into three broad buckets: taxpayer services (chatbots, phone routing, document processing), operational efficiency (workforce scheduling, case management), and tax compliance and fraud detection. Audit selection sits squarely in that third category. The public inventory does not break down exactly how many of the 125 systems are dedicated to flagging returns versus handling other tasks, but the GAO review indicates that compliance-related AI represents a substantial and expanding portion of the agency’s overall AI portfolio.
Governance has expanded alongside the technology. The IRS’s internal policy manual, IRM 10.24.1, sets agency-wide rules for AI oversight. It defines what counts as an AI use case, requires inventory maintenance and validation, and establishes reporting obligations. Systems that affect individual rights, safety, or access to government services receive a “high-impact AI” designation that triggers additional layers of review. The Treasury Department publishes a separate AI Use Case Inventory covering all bureaus, including the IRS, updated on federal reporting cycles.
These disclosure requirements grew out of a 2019 executive order on American leadership in artificial intelligence and the National AI Initiative Act, which was enacted as part of a defense authorization package. Both mandated that federal agencies inventory and publicly disclose their AI deployments. The IRS is following that playbook, but the sheer volume of its enforcement-related adoption sets it apart from nearly every other civilian agency.
What the audit numbers actually show
The IRS publishes examination coverage rates in its annual Data Book, broken down by return type and income bracket. The trend line for the wealthiest filers tells a clear story of decline followed by partial recovery.
According to Data Book tables covering examination coverage by income category, taxpayers with total positive income above $10 million faced a 13.6% examination rate for tax year 2012. By tax year 2018, that figure had fallen to 9.2%, a direct consequence of budget cuts and staff losses that gutted the agency’s ability to pursue complex cases. Since then, the trajectory has reversed. The IRS has publicly committed to restoring audit coverage for high-income individuals and large corporations, using IRA-funded resources to rebuild examination teams and deploy technology capable of processing intricate returns more efficiently.
In an official statement on examination coverage, the agency noted that Data Book rates represent a fiscal-year-end snapshot and that coverage for higher-income brackets tends to rise over time as additional exams are opened and closed within the statutory window. That upward drift means published rates at any given moment likely understate the eventual audit exposure for wealthy filers.
Whether those rates will fully return to pre-2010 levels depends on factors no algorithm controls: congressional appropriations, leadership priorities at the agency, and the political durability of enforcement funding that has already survived multiple reduction attempts on Capitol Hill. The IRS has signaled the intent, and the tools are in place, but no official projection ties a specific target percentage to a specific year. The headline framing of rates “climbing back toward pre-2010 levels” reflects the stated direction of agency policy and the upward trend visible in recent Data Book figures, not a guaranteed destination.
What AI changes about audit selection
For decades, the IRS chose which returns to examine using a combination of statistical scoring, primarily the Discriminant Index Function (DIF), and human judgment. AI does not replace that process. It layers on top of it. Machine learning models can ingest far more data points than traditional scoring formulas, cross-referencing a filer’s return against third-party information reports, historical filing patterns, entity relationships, and known fraud indicators at a speed no human analyst can match.
The practical effect is that complex arrangements, such as multi-entity pass-through structures, aggressive loss harvesting strategies, or opaque offshore holdings, are precisely the patterns these systems are trained to detect. A return that might have cleared a manual review because no single line item looked unusual can now be flagged because the combination of line items matches a risk profile built on thousands of prior examinations.
Tax practitioners say the change is already visible in the types of information document requests they receive. Examiners appear to arrive at audits with a more detailed picture of a taxpayer’s overall financial footprint than they did even three or four years ago, suggesting that AI-driven case building is shaping the process before a revenue agent ever picks up the phone.
None of this changes the rules once an audit is underway. The burden of substantiation still falls on the taxpayer. Thorough documentation, contemporaneous records, and clear economic substance remain the strongest defenses, regardless of whether a case was opened by an algorithm or a revenue agent’s instinct.
The transparency gap
The rapid expansion raises questions the IRS has not fully addressed. The agency has not published independent assessments of whether AI-driven targeting produces equitable outcomes across demographic groups. The GAO has flagged the importance of monitoring AI systems for fairness, but specific impact evaluations tied to enforcement equity do not appear in the current public record.
Then there is the question of alignment between AI selection and the agency’s own promises. The IRS has repeatedly pledged not to raise audit rates on households earning less than $400,000, a commitment reinforced by Treasury directives issued alongside the IRA funding. How that policy constraint is encoded into the models, and how consistently it holds across millions of returns, is not visible from outside the agency. The governance rules in IRM 10.24.1 create a baseline of accountability, but they stop well short of revealing how specific models weigh variables or what training data they rely on.
For tax professionals and policymakers, this opacity creates a genuine tension. AI allows the IRS to sift through enormous volumes of returns and concentrate scarce human resources on the cases with the highest estimated compliance risk. That is a clear efficiency gain. But the same opacity that makes the models effective also makes them harder to scrutinize from the outside, an irony familiar to practitioners who spend their careers interpreting the agency’s own rules.
Taxpayer advocates have raised a related concern: if a filer believes an AI system unfairly selected their return for examination, no formal process currently exists to challenge the algorithmic decision itself. The audit can be contested on its merits, but the selection mechanism remains a black box.
What high earners should prepare for now
AI-assisted audit selection is no longer experimental. It is an active, scaled system backed by more than 100 deployed use cases and formal governance structures. Filers in the highest income brackets should assume their returns pass through multiple algorithmic screens before a human examiner ever opens a case file. That does not guarantee an audit, but it means that patterns of income, deductions, and entity structures are being measured against increasingly sophisticated risk profiles every filing season.
For tax advisors working with wealthy clients, a few steps stand out. First, assume the structures drawing the most AI attention involve layered entities, unusually large or irregular deductions, and cross-border capital flows. Second, build documentation that anticipates algorithmic pattern-matching, not just the questions a human reviewer might ask. Third, track the IRS’s evolving public AI inventories and compliance guidance to stay current on which areas are attracting the heaviest technological investment.
Where enforcement AI is headed by the end of the decade
The trajectory is difficult to miss. Verified records show a rapid expansion of algorithmic tools at the IRS, a renewed institutional focus on high-income enforcement, and a governance framework still catching up to the technology it is supposed to oversee. As of June 2026, the agency’s AI footprint in enforcement is larger than that of most federal regulatory bodies, and it is still growing.
For filers at the top of the income scale, the age of the AI-powered audit is not approaching. It is already here, and the systems behind it are getting sharper with every filing season that feeds them new data.
