By Stuart Kerr, Technology Correspondent, LiveAIWire
As of June 2025, the Internal Revenue Service maintained 126 active AI use cases in tax enforcement, up from just 10 in August 2022, according to a March 2026 Government Accountability Office report. The GAO confirmed that the IRS is deploying AI at rapidly expanding scale across audit selection, fraud detection and broader compliance operations. The IRS’s own published tax gap projections put the amount of tax owed but not paid at approximately 696 billion dollars a year for 2022, of which the agency expects only about 90 billion dollars to eventually be recovered through enforcement and late payments.
Between January and May 2025, the IRS cut its workforce by 25 percent, from 103,000 to 77,000 employees. The agency is deploying AI to do more enforcement work with fewer people. A Stanford-led study that the IRS itself later confirmed found that Black taxpayers are audited at rates two point nine to four point seven times higher than other taxpayers, a disparity the researchers traced to the audit selection algorithm rather than to any difference in actual non-compliance. The race to close the tax gap with AI is producing the same equity problems that AI systems have produced in every other domain where they have been trained on historically biased data.
The story of AI in taxation runs in two registers. In one, AI enables governments to detect tax evasion with accuracy and at scale that human enforcement could never achieve, recovering public revenues that would otherwise have funded private wealth. In the other, AI tax enforcement reproduces and amplifies the biases of the data it is trained on, targeting some groups at dramatically higher rates than their actual evasion levels would justify, and operating with a transparency deficit that makes accountability difficult.
What the IRS Is Actually Doing
The IRS’s AI deployment encompasses several distinct functions that should not be conflated. Machine learning models analyse tax returns and financial data to score each filing for non-compliance risk, prioritising the highest-risk returns for audit. AI tools identify specific patterns indicative of fraudulent activity, including identity theft, syndicated conservation easements and abusive micro-captive insurance arrangements. In late 2025, the IRS deployed the Agentforce system across the Office of Chief Counsel, Taxpayer Advocate Services and the Office of Appeals, using AI to search documents and provide case summaries so staff can spend more time on complex casework. Voice bots have handled over 4.8 million calls, while chatbots have resolved more than 450,000 inquiries.
The audit selection application is the most consequential and most contested. AI has changed the distribution of audit targets meaningfully. The IRS has used AI to target dozens of the largest US partnerships, several with assets exceeding 10 billion dollars, recovering unpaid taxes by identifying complex tax avoidance schemes that human auditors working within current resource constraints would not have detected at this scale. On the other end of the income distribution, the same GAO findings point to AI audit selection also increasing scrutiny of lower-income taxpayers claiming the Earned Income Tax Credit, a pattern that reflects the training data the models learned from rather than actual non-compliance rates at that income level.
The Algorithmic Bias in Audit Selection
The finding that Black taxpayers face audit rates roughly three to five times higher than comparable taxpayers is the most serious documented equity failure in IRS AI deployment. When the IRS’s own analysis confirmed the Stanford team’s results in 2023, the agency’s commissioner attributed the disparity to problems in the audit selection algorithm rather than any deliberate targeting, a mechanism consistent with what shows up across AI systems generally: models trained on historical audit data inherit the patterns of past enforcement decisions, which were themselves disproportionate. An AI trained to predict audit productivity based on who was previously audited will learn to target the groups that previous auditors targeted, not the groups with the highest actual non-compliance rates.
The EITC audit problem is specifically documented in the Stanford research: lower-income taxpayers claiming this credit have historically been audited at high rates because their returns are relatively simple to audit and generate quick adjustments, while complex high-income returns with sophisticated avoidance schemes are more expensive to audit and more likely to be contested. The AI training data reflects this historical pattern, and audit selection models built on it replicate and can amplify it. The underlying dynamic is one our coverage of AI-driven insurance pricing has examined in a different context: a model trained to be accurate about a historical pattern will faithfully reproduce whatever inequity was baked into that pattern, whether the domain is audits or premiums.
The Transparency Deficit
The GAO’s March 2026 report identified significant workforce, inventory and governance gaps in IRS AI deployment, flagging insufficient documentation of IRS models and raising concerns about accountability and error correction. The IRS does not publish the rules, model weights or thresholds it uses to score tax returns, and it does not tell taxpayers how or why their return was selected for examination. This opacity is not unique to the IRS, tax authorities in other jurisdictions operate similarly, but it becomes harder to justify as the scale and consequence of AI audit selection grow.
The IRS states that all AI-selected cases are reviewed by a human examiner before an audit proceeds, but that formal requirement does not automatically translate into a genuine check on algorithmic bias. When an algorithm outputs a high-risk score, an examiner’s practical incentive is to conduct the audit rather than to challenge the selection, particularly for lower-income filers who are unlikely to mount a sophisticated legal challenge to the audit itself. The GAO’s recommendations to the IRS, which include workforce planning for AI skills and better documentation of how AI use cases are expected to benefit the agency, are aimed squarely at this gap between the formal safeguard and its practical effect.
What Businesses Are Doing in Response
On the other side of the AI tax enforcement equation, businesses and their advisors are deploying AI to understand, anticipate, and in some cases minimise the exposure that AI audit selection creates. Tax technology platforms analysing a company’s return against algorithmic patterns associated with audit risk are already in commercial use, providing businesses with a view of their exposure before filing. The result is an arms race that mirrors dynamics our reporting has covered in AI’s broader reshaping of global finance: as AI enforcement becomes more sophisticated, tax planning that anticipates AI detection patterns becomes a competitive service in its own right, and the highest-value tax avoidance strategies are increasingly those built to be invisible to the AI models used to detect them.
The UK’s HMRC and equivalent authorities in EU member states are pursuing parallel AI enforcement programmes. Real-time tax reporting requirements, which several EU member states have introduced, create the structured transaction data that AI analysis requires. Tax administrations that have invested in real-time data infrastructure are generating better AI enforcement outcomes than those relying on annual return data, because the models benefit from the granularity and timeliness that transaction-level data provides.
The Corporate Tax Avoidance Dimension
The IRS AI enforcement expansion is most visible in its effects on individual filers, the audit rate disparities, the algorithmic selection of EITC claimants, but the more economically significant application is the targeting of complex corporate and partnership tax structures. The IRS’s deployment of AI to identify complex tax avoidance among the largest US partnerships represents a genuine capability improvement: these structures are sufficiently complex that human auditors working within the IRS’s resource constraints could not systematically identify the patterns of avoidance that AI analysis can surface across thousands of returns simultaneously.
The corporate share of the tax gap is substantially driven by legal avoidance strategies designed to stay within the letter of the law while violating its spirit, and identifying them systematically requires exactly the kind of pattern recognition across large datasets that AI is well suited to. The firms building the most sophisticated corporate tax planning strategies are already incorporating AI analysis of enforcement patterns into their own planning, identifying which avoidance structures are likely to attract AI audit attention and which are novel enough that training data does not yet flag them.
This adversarial dynamic, sophisticated avoidance strategies migrating away from patterns that AI models have learned to recognise, toward arrangements the models have not yet seen, is a feature of the tax system’s AI transition that will likely define its character for the decade ahead, in much the same way that AI-driven personal money management raises questions our piece on algorithms managing your finances explores from the consumer side.
The International Cooperation Dimension
The tax gap problem is international as well as domestic, and AI enforcement is developing in an international context that raises coordination challenges domestic deployment does not. Transfer pricing, the prices at which related companies within the same corporate group transact with each other across national borders, is the mechanism through which the majority of international corporate tax avoidance occurs, and AI analysis of transfer pricing patterns requires data sharing across tax authorities that remains constrained by both legal frameworks and practical data governance limitations.
The OECD’s BEPS framework and the Global Minimum Tax agreement represent attempts to build the international coordination that effective AI enforcement of international tax would require. Their implementation is incomplete and their enforcement mechanisms are weaker than the domestic mechanisms the IRS is now augmenting with AI. The domestic AI enforcement expansion is making domestic tax avoidance harder. The international dimension remains the larger and less tractable challenge.
The Small Business and Self-Employed Dimension
The IRS AI enforcement expansion has a specific dimension that large corporate tax planning resources can address but that small businesses and self-employed individuals typically cannot. AI audit selection models trained on historical compliance data may flag small business returns for patterns that reflect genuine complexity in self-employment income, home office deductions and vehicle use rather than actual non-compliance. The self-employed consultant whose income varies significantly year to year, whose business expenses include items that can be personal or professional, and whose record-keeping may not meet the documentation standards that large-business accounting systems generate automatically, presents exactly the kind of complex return that AI audit selection is likely to flag.
Yet the resources available to that individual to respond to an AI-selected audit, professional representation, comprehensive documentation, a legal challenge, are a fraction of those available to the large partnerships the IRS is also targeting. This asymmetry echoes a theme our coverage of the AI freelance economy has returned to repeatedly: the same automation that creates genuine efficiency gains at scale tends to leave independent workers with the least capacity to absorb its downside.
The equity dimension of AI tax enforcement is not reducible to the racial audit disparity, significant as that is. It is also a small-versus-large business equity issue, a technical-capacity equity issue between sophisticated taxpayers with AI-enabled tax planning and individuals using standard software, and a representation equity issue between those who can afford professional representation in an AI-selected audit and those who cannot.
Tax administration that improves overall compliance while concentrating the burden of enforcement on those least able to bear it is not achieving the equity that tax systems are supposed to embody. The GAO’s identification of governance gaps in IRS AI deployment is the beginning of an accountability process that the agency’s workforce reductions and budgetary pressures make more difficult to pursue rather than less. The governance of AI tax enforcement requires the same rigour and transparency that tax enforcement itself is supposed to provide.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.