- TIGTA found two of five reviewed high-impact IRS AI use cases lacked documented impact assessments
- Four of the five IRS AI use cases TIGTA reviewed had no data quality testing documentation
- The IRS agreed to finish impact assessments for deployed high-impact AI use cases by November 2026
The Treasury Inspector General for Tax Administration, or TIGTA, has recommended that the Internal Revenue Service improve how it documents risk and data quality evaluations for high-impact artificial intelligence use cases. TIGTA made the two recommendations to align IRS practices with a federal AI governance memo, and the IRS agreed with both.
In a report issued Thursday, TIGTA said the IRS began AI initiatives for tax return classification and issue selection in 2017. The tax agency had an inventory of 225 AI use cases as of December 2025.
Federal civilian agencies are working to put AI governance and data quality standards into practice. Keep the conversation going at the Potomac Officers Club’s 2026 FedCiv Summit on Oct. 29 in Arlington, Virginia. Amy Ritualo, chief data and artificial intelligence officer at the State Department, will join federal technology leaders on a panel about how agencies can modernize civilian IT through AI for mission impact. Save your seat today and be part of the discussion.
What Did TIGTA Find About IRS AI Impact Assessments?
The watchdog examined five deployed high-impact AI use cases that support IRS workload selection and pull data from the agency’s Compliance Data Warehouse. Two of the five had no documented impact assessment. For the other three, the IRS completed assessments under guidance that predated Memorandum M-25-21 from the Office of Management and Budget. According to TIGTA, inconsistent documentation of AI risks and impacts could cause the agency to overlook risk mitigation measures.
What Does OMB Memorandum M-25-21 Require for High-Impact AI?
OMB replaced Memorandum M-24-10 with the new memo in April 2025 and gave agencies until April 2026 to adopt minimum risk management practices for high-impact AI. Those practices call for documented impact assessments covering an AI system’s intended purpose, data quality and potential impacts. Agencies must safely stop using any AI functionality that falls short of the requirements.
The IRS rolled out its own AI governance policy and risk management processes in February 2026 to meet the memo’s requirements. Management in the IRS Research, Applied Analytics and Statistics, or RAAS, organization said the agency had not received implementation guidance from the Treasury Department as of July 2026.
How Did the IRS Document Data Quality for Its AI Models?
Program managers and data scientists at the IRS told TIGTA that their teams performed data quality checks and deemed the data fit for use, though the checks and records differed across use cases. Testing documentation existed for one of the five use cases, which also detailed change management, oversight and validation, while the remaining four had none.
RAAS management agreed that standardized documentation supports repeatable, consistent risk decisions but had not made baseline data quality procedures a priority before using data in high-impact models.
The report also said the data warehouse’s internal web page advises users not to rely on its data in place of source systems for operational decisions.
What Does NIST Recommend for Managing AI Data Risks?
TIGTA cited the National Institute of Standards and Technology’s AI Risk Management Framework 1.0, which describes the challenges of managing AI systems that depend on large datasets.
Under the framework, organizations should match policies and resources to a system’s assessed risk level and potential impact. Systems trained on sensitive information, such as Social Security numbers, and those whose outputs affect people warrant higher priority.
Without written data quality procedures, TIGTA said, high-impact AI could produce inaccurate, biased or unreliable results that lead to financial and reputational harm.
How Will the IRS Address TIGTA’s AI Recommendations?
TIGTA directed its recommendations to the IRS chief data and analytics officer, calling for impact assessments that meet M-25-21 requirements and standard processes to evaluate data quality and document data fitness. In a Sept. 21 response, Lucia Lykke, acting chief data and analytics officer, said the IRS will complete impact assessments for its deployed high-impact use cases by November 2026.
The agency said it put standard data quality evaluation processes in place on July 13. The report follows a March Government Accountability Office review that found gaps in the IRS AI inventory and AI workforce planning.
