IRS pilots agentic AI to assemble fraud cases, not decide them

The IRS is piloting an agentic AI tool designed to take over the routine steps of building fraud and identity-theft cases, with a federal watchdog estimating that tax fraud already costs the government between $116 billion and $304 billion every year.

Jim Clifford, director of return integrity and compliance services in the IRS's Taxpayer Service Division, announced the pilot Friday during a panel at ACT-IAC's Service Delivery and Experience Summit. The tool would act after anomalies in tax returns are flagged, not before.

How the IRS catches fraud today

The agency already runs machine-learning models that spot anomalies in incoming returns. Flagged returns land in "buckets, for lack of a better technical word," Clifford said, where the system compares them against "a whole bunch of other data" to gauge the risk of identity theft or fraud.

Clifford rates the agency's performance at "detecting fraud at the front door" as strong. The problem sits downstream. The "back end" — where employees work through audit and remediation cases — moves slowly, and that is where agentic AI would step in.

The takeaway: detection is largely solved by existing models; the bottleneck is the manual case work that follows a flag.

What would the AI actually do?

In a sideline interview after the panel, Clifford told FedScoop that the "clerical, rote work" includes organizing correspondence with taxpayers and reading through the pages they send in.

"What we want the agentic AI to do is review the response, summarize it, build a case, and hand it off to an employee at the point where it's really ready to make a decision," he said.

The distinction matters. Clifford framed the technology as a way to relieve the "constraints" on how fast agency workers can operate — not as a decision-maker.

• The AI reviews and summarizes taxpayer responses.
• It assembles the case file.
• It hands the case to a human employee.
• "Making judgments," Clifford said, will always remain under a person's purview.

In plain terms: the software prepares the paperwork; a person signs off on the outcome.

What is the scale of the threat?

The stakes behind the pilot are large. Clifford said the IRS faces roughly 1 billion attempts each month from fraudsters trying to obtain unauthorized information — about 33 million attempts a day — largely because "your tax returns are gold" to criminals.

The agency is "pretty damn good at protecting" taxpayer data, he added, though there are "very few times" when information is "released inappropriately."

On identity-theft refund fraud specifically, Clifford said the IRS is "99% accurate" in blocking attempts "based on all estimates and studies." Even so, he said the residual cost of fraud is too high — "substantial" doesn't go far enough as a description.

Put differently: stopping 99 out of 100 attempts still leaves a multibillion-dollar hole when the exposure runs to a billion attempts a month.

When will it deploy?

Clifford gave no timeline. He said the pilot is being developed "on another workstream" and will be scaled to other areas of the IRS "once it's tested and proven to be successful."

What else could change?

Beyond the agentic AI tool, Clifford pointed to two levers: more data-sharing across federal, state and local governments, and greater "interoperability" with partners.

"There are legislative and regulatory constraints about us sharing data across agencies, but there are really good things we could do if we could do that," he said on stage.

"We have the technology now with APIs, with the use of" AI, he continued. "We can share information in a very safe way across agencies to serve the American taxpayer better and to protect the American taxpayer better. And I'd love to see us get that breakthrough."

The bottom line: the IRS is automating case preparation, not judgment, and wants the legal room to share data that its own technology can already handle securely.