IRS Pilots Agentic AI to Prepare Potential Fraud Cases for Human Review

The Internal Revenue Service (IRS) is piloting an agentic AI tool to help employees process potential tax fraud cases after existing machine-learning models identify anomalies in tax returns. Speaking at ACT-IAC’s Service Delivery and Experience Summit, Jim Clifford, director of return integrity and compliance services in the IRS Taxpayer Service Division, described the pilot as an effort to address routine work after the agency’s initial fraud-detection process. The current workflow routes anomalous returns into review groups, or “buckets,” where they are compared against other data to assess the risk of identity theft or fraud. Clifford said the IRS is already “pretty good” at detecting fraud at the front end; the proposed AI support is aimed at the case-resolution work that follows.

The tool is intended to perform clerical tasks, not make determinations about taxpayers. Clifford said the system could organize taxpayer correspondence, review responses, summarize documents, assemble case materials, and hand a prepared case to an employee when it is ready for a decision. He characterized this as helping staff manage the “rote steps of an audit or a fraud detection or a fraud remediation case.” Human employees would retain authority over judgments and decisions. For contractors, this boundary is central to the opportunity signal: relevant solutions would need to support efficient case preparation while keeping consequential decisions under human control.

The pilot is being developed on a separate workstream, and the IRS has not provided a deployment timeline. Clifford said expansion to other parts of the agency may follow if the technology is tested and proven successful. That makes the announcement a technology-development signal rather than an active procurement opportunity: the source identifies no solicitation, procurement vehicle, contract value, or schedule. Companies tracking IRS modernization can nevertheless monitor for future requirements involving agentic workflow tools, secure handling of taxpayer records, document review and summarization, audit trails, and interoperability with agency systems.

The potential operational stakes are significant. Clifford cited approximately 1 billion attempts each month by fraudsters to obtain unauthorized information, noting that tax returns are valuable targets. He said the IRS is “99% accurate” at thwarting attempted identity-theft refund fraud based on estimates and studies, but that the remaining fraud still carries substantial costs. The Government Accountability Office estimates the federal government loses $116 billion to $304 billion annually to tax fraud. Clifford also pointed to greater data-sharing with federal, state, and local governments, along with improved interoperability, as possible elements of the broader fraud-fighting effort. These factors suggest that any future procurement activity may extend beyond AI models to the secure data and operational infrastructure needed to use them responsibly.

• The IRS is piloting agentic AI for clerical casework after machine-learning models flag anomalous returns.
• Proposed functions include organizing taxpayer correspondence, reviewing responses, summarizing documents, and preparing cases for employee review.
• Human employees retain responsibility for judgments and decisions; the tool is intended to support, not replace, staff decision-making.
• The pilot has no announced deployment timeline, solicitation, procurement vehicle, or contract value; expansion may follow successful testing.
• The agency’s existing process compares flagged returns with other data to assess identity-theft and fraud risk.
• IRS official Jim Clifford cited approximately 1 billion attempts per month to obtain unauthorized information.
• Clifford said the IRS is “99% accurate” at thwarting attempted identity-theft refund fraud, based on estimates and studies.
• The Government Accountability Office estimates annual federal losses from tax fraud at $116 billion to $304 billion.
• Contractors can track future needs in secure document processing, case-management workflows, data interoperability, and human-oversight controls.
• Clifford identified increased data-sharing with federal, state, and local governments as a potential aid to fraud prevention.