The Algorithm That Decides Who Gets Audited: IRS Paid Palantir $1.8M to Build AI Targeting Tool

What the Documents Show

The Internal Revenue Service paid Palantir Technologies $1.8 million last year to improve a custom tool designed to help the agency identify what internal documents describe as "highest-value" cases for audits, collection of unpaid taxes, and potential criminal investigations. The contract and scope documents were obtained by WIRED through Freedom of Information Act requests and published March 30, 2026.

The tool, named the "Selection and Analytic Platform" or SNAP, is currently being used as part of a pilot program, according to the documents. Palantir and the IRS both declined to respond to WIRED's requests for comment.

The contract documents describe an IRS that had grown increasingly overwhelmed by its own complexity. At the time the contract was signed, the agency said it was relying on "more than 100 business systems and 700 methods," built over decades, to select cases in which people may have incorrectly reported their taxes or owe money. The IRS wrote in documents obtained by WIRED that this "fragmented landscape can lead to a number of undesirable outcomes including but not limited to duplication of effort and cost, poor understanding of gaps in the coverage, and suboptimal case selection."

SNAP is designed to sit on top of the IRS's existing fragmented databases and help human auditors identify red flags in tax filings they might otherwise have missed, surfacing "key information about contracts, vehicles and vendors" from "unstructured data from supporting documents," according to the contract.

What SNAP Is Looking For

The IRS asked Palantir to build three specific "case selection methods" within SNAP, each targeting a distinct area of the tax code:

Disaster zone claims — a form of tax relief available to natural disaster victims, which the IRS apparently believes may be a source of incorrect or fraudulent filings.

Residential Clean Energy Credits — a tax credit program introduced under the Inflation Reduction Act that offsets the cost of installing solar panels, wind turbines, and similar equipment. According to WIRED's reporting, the IRS is now treating this category as a priority for scrutiny.

Form 709 Gift Tax Returns — filings required when individuals transfer valuable assets such as artwork, stocks, or corporate entities to another person.

Mitchell Gans, a professor at Hofstra University with expertise in gift and estate taxes, told WIRED that if SNAP is analyzing unstructured data from supporting documents on gift returns, it may be examining "adequate disclosure" forms that require detailed descriptions of how gifted property was valued and the relationship between the giver and recipient.

Erica Neuman, an accounting and finance professor at Youngstown State University, noted that public logs from money transfer applications like Venmo, as well as public storefronts on platforms like Etsy and Depop, could contain unstructured data of interest to the IRS. However, the contract documents specify that Palantir is only authorized to use "existing data in SNAP today" — meaning the agency is not directing Palantir to pull in new external data sources at this stage.

Palantir's $200 Million Footprint at the IRS

SNAP is not the beginning of the IRS's relationship with Palantir. Government contracting records show the IRS has been purchasing Palantir technology since 2014. In total, the company has been awarded more than $200 million in IRS contracts and obligated payments. The WIRED documents indicate the agency is now interested in deepening that relationship further.

Palantir's expanding role in U.S. government agencies has drawn consistent attention from civil liberties advocates. The company already holds a $1 billion Department of Homeland Security contract for AI and data analytics, according to reporting cited by health technology watchdog Gloria Maloney. The American Civil Liberties Union has documented Palantir's role in immigration enforcement operations, including helping power the Trump administration's deportation infrastructure.

The Register reported in March 2026 that "Palantir's track record — spanning US defense, intelligence, and immigration enforcement — has made it a lightning rod for concerns about surveillance and civil liberties, especially when deployed in civilian contexts." The same report noted the company had landed a trial contract with the UK's Financial Conduct Authority worth more than £30,000 a week to analyze that regulator's internal data lake of fraud, money laundering, and consumer complaint files.

The Contradiction: Cutting the Staff Who Run AI

The SNAP revelations arrive alongside a sharply critical watchdog report about the IRS's actual capacity to use AI effectively.

The Government Accountability Office, in a performance audit conducted from April 2024 to March 2026, found that the IRS has been cutting the very staff members who support its AI efforts — creating what the GAO described as "skills gaps" that "could greatly affect" the agency's ability to use emerging technology, according to FedScoop's reporting on the GAO findings.

The IRS's Research, Applied Analytics and Statistics division — one of the agency's most active AI users — spent more than $58 million on AI in fiscal 2024 and planned to spend an additional $32 million in fiscal 2026, per the GAO. But the same division lost 63 staffers who had worked either full- or part-time on AI. Many others were reassigned away from AI-focused work, according to FedScoop.

Treasury Secretary Scott Bessent told lawmakers last May that the agency planned to offset workforce reductions through an "AI boom." The GAO's report pushed back on that framing, finding that IRS officials had not identified the skills needed to support AI or developed a plan to address gaps.

"However, IRS officials said they had not identified skills needed to support AI or developed a plan to address the skills gaps," the GAO noted in its press release. "The recent staff reductions, the intent to pursue additional AI initiatives, and the absence of a plan to address AI skills gaps increase the risk that IRS AI efforts will not succeed."

The IRS's AI use case inventory has nevertheless grown from 10 entries in 2022 to 129 as of early 2026, according to the GAO data.

Who Decides Who Gets Flagged?

For decades, the primary method the IRS used to decide which returns to audit was a numerical score called the Discriminant Information Function (DIF) score. According to the IRS, the higher the score, the greater the statistical likelihood of a tax discrepancy. DIF scoring is a rules-based, formula-driven system developed and refined over decades of audit data.

SNAP represents a different approach — one that attempts to synthesize signals from unstructured documents and multiple legacy databases to surface cases that DIF scoring and existing methods may miss. Whether SNAP produces audit selections that are more accurate, more equitable, or more susceptible to opaque algorithmic bias is not addressed in the documents obtained by WIRED.

Civil liberties organizations have long argued that AI-driven selection tools in government enforcement contexts can embed and amplify existing disparities. The ACLU, in its documentation of Palantir's immigration enforcement role, noted that company CEO Alex Karp has publicly argued that Palantir's tools promote civil liberties by ensuring "that the state and its agents can see only what ought to be seen" — a claim the ACLU said is difficult to assess given limited public disclosure of how the tools function.

As of publication, Palantir's SNAP tool remains in pilot status at the IRS. No public timeline for broader deployment has been disclosed.