IRS data documentation missing for 80% of sampled AI use cases
TIGTA finds informal data governance that can't survive an audit, the inventory exists, but the paper trail doesn't, and that's the exposure OMB's AI mandate was supposed to close.
TL;DR
A Treasury Inspector General for Tax Administration report found that four of five sampled IRS AI use cases (drawn from the agency's OMB-required inventory) had no testing documentation for their data. RAAS, the IRS unit that manages the AI inventory, told TIGTA it conducted informal data quality checks but hadn't standardized or documented the procedures. IRS management agreed with TIGTA's recommendation to adopt formal data-quality assessment processes and committed to completing AI impact assessments for deployed high-impact use cases by November 2026, per the report.
The finding lands in the gap OMB's 2023 AI executive order implementation was supposed to close: agencies built the inventories, but the governance infrastructure underneath remains patchy.
TIGTA's sample was small (five use cases) and the 80% figure shouldn't be extrapolated across all 126 entries in the IRS inventory. But the pattern matters. RAAS officials told the watchdog they'd done the work: data fit checks, appropriateness assessments, the kind of informal quality review that competent teams perform as a matter of course. What they hadn't done was write any of it down in a way that would survive an OMB audit, a FOIA request, or a challenge to a model's output.
That's the tension TIGTA named directly. RAAS acknowledged that standardized documentation "would likely improve risk management decisions" but didn't prioritize it. The report's warning is unusually blunt for an IG: without formal procedures for assessing data quality, high-impact AI models risk "inaccurate, biased, or unreliable AI outputs that can cause significant financial losses and reputational damage."
The IRS agreed with TIGTA's recommendation and said corrective steps are underway. It also committed to completing AI impact assessments for currently deployed high-impact use cases by November 2026, a date that now functions as a de facto compliance milestone, though it remains unclear whether OMB will treat it as a hard enforcement trigger or continue the agreement-based remediation pattern TIGTA used here.
The broader GAO work on IRS AI, published in March 2026, reinforces the picture. GAO found the IRS inventory as of June 2025 was incomplete, it omitted some active AI applications and contained entries with quality issues. Meanwhile, staffing reductions hit the AI workforce directly: RAAS alone lost 63 employees working full- or part-time on AI, and other IRS units reported similar cuts.
The documentation gap TIGTA identified isn't a surprise against that backdrop. It's what happens when inventory compliance becomes a checkbox exercise (the list gets built, the underlying governance doesn't) and the people who'd close the gap are departing.
Published ·Deep Fathom