Palantir is testing a system for the IRS that picks priority targets for tax audits

The US Internal Revenue Service is testing a Palantir tool that helps select the audits and investigations most worthwhile to the government. The pilot is meant to pull signals from dozens of old IRS systems and find more quickly the cases where underpayment, fraud or grounds for a criminal case may exist.

How the pilot works

According to documents obtained through data access requests, the IRS paid Palantir $1.8 million in 2025 to extend the Selection and Analytic Platform, or SNAP. The tool is meant to improve how priority cases are chosen for audits, collection of overdue taxes and potential criminal investigations. For the agency itself, the goal looks pragmatic: instead of wandering by hand through an old stack, assemble the picture in one place and spend less time on audits that bring almost nothing to the budget.

The problem is that the IRS has accumulated very heavy infrastructure. In the documents, the agency acknowledges that it uses more than 100 business systems and about 700 selection methods that have been built up over decades. SNAP does not replace this entire landscape in one step but sits on top of it: it helps auditors see signals from scattered databases and find important details in unstructured attachments that were harder to cross-reference before. According to public procurement data, Palantir has worked with the IRS for more than ten years, and total contracts and payment obligations have exceeded $200 million.

Which cases are in focus

In the first phase, the IRS asked Palantir to build three separate case-selection methods tied to specific areas of the tax code. This is not an abstract search for "suspicious citizens" but a fairly practical pilot with clear categories, where the agency wants to test whether the new approach helps find more productive cases. The design matters: the agency limits the experiment and can compare the new logic with old procedures on specific types of claims and returns.

• claims for tax relief from residents of disaster zones
• Residential Clean Energy Credit - credits for installing solar panels, wind turbines and other home energy equipment
• Form 709 - gift tax returns, filed when stocks, businesses, artworks and other expensive assets are transferred
• supporting documents from which one can infer an asset's valuation, the relationships between the parties and the logic of the calculation

Unstructured data is of particular interest here. In gift tax cases, this can include appraisal documents, balance sheets, business income records and descriptions of the relationship between donor and recipient. The contract materials also state separately that Palantir must work with existing data inside SNAP rather than pull new external data streams into the system at its own discretion. Such files are often where signals hide that are absent from the standard fields of a return.

Why it is controversial

The IRS already has algorithmic selection mechanisms. For decades the agency has relied on the DIF score, an internal metric that estimates the likelihood that a return merits an audit. How it is calculated is not publicly disclosed, and researchers have long called this approach a black box. SNAP takes the next step: it does not just rank returns but helps link more sources together and surface cases an inspector might have missed.

The risk lies not only in privacy but in the quality of decisions. In May 2025, a TIGTA audit wrote that the IRS already uses AI models to select audits and wants to reduce the burden on honest taxpayers, but has not yet fully built a way to evaluate how effective these models are compared with older methods. In other words, the agency wants to target violations more precisely, but it still has to prove that the new models really work better and do not create extra errors.

"The IRS has essentially never brought a full modernization to success since the 1960s," says

Professor Erica Neumann.

That quote helps explain why the agency turns to an outside contractor at all. IRS modernization has stalled for years, and in 2025 the agency also lost tens of thousands of employees through layoffs, deferred resignations and early retirements. With fewer people and too many disconnected systems, the temptation to hand case selection to a smarter analytics layer becomes almost unavoidable. It also explains the interest in systems that promise to compress weeks of manual analysis into hours.

What this means

For Palantir, this is another chance to embed itself deeper in critical US government infrastructure. For the IRS, it is an attempt to turn scattered tax data into a more precise mechanism for selecting audits. If the pilot delivers results, similar systems will start to shape not only the speed of audits but also who the state treats as a priority target for enforcement. The next debate will no longer be about whether such software is needed, but about the rules for auditing it and holding it accountable.