VA is racing into AI with 15 disability-oversight fixes still undone
Summary
The Department of Veterans Affairs reported 40 AI use cases in 2023, 229 in 2024, and 367 by December 2025 -- growth GAO detailed in July 10, 2026 testimony on a disability compensation program that paid over $195 billion to more than 6.9 million veterans and their families in fiscal year 2025. Since 2021, GAO has made 43 recommendations to fix that program; VA has implemented 28, but 15 remain not fully addressed, including gaps in oversight of contracted medical exams and claims-processor training. The oldest thread runs back to 2009: a 2015 review found VA's paperless claims system was built without goals for response times or user satisfaction, and of five recommendations made then, only three are implemented 11 years later -- the other two were closed in 2017 after a user survey GAO says was never statistically validated.
View data as table
| 2023 | 40 |
|---|---|
| 2024 | 229 |
| Dec. 2025 | 367 |
The recommendation backlog behind the AI push
Since 2021, GAO has made 43 recommendations⧉ to improve 's disability compensation program; the department has fully implemented 28 of them, but has taken only partial steps on the remaining 15. Two recent reviews show the pattern up close. After finding breakdowns in VA's oversight of contracted medical examiners⧉ -- including incorrect financial-incentive payments to contractors -- made five recommendations in 2024 and 2025; as of July 2026, had implemented three, leaving open a recalculation of financial incentives paid since 2022 and regular reviews of complex-claim exams. A 2021 review of claims-processor training found similar gaps: of 10 recommendations, has implemented eight, with two still open six years later covering how VBA prioritizes training goals and evaluates results.
View data as table
| Implemented | 28 |
|---|---|
| Not fully implemented | 15 |
A decade-old fix for the system underneath the claims
The AI push sits on top of infrastructure flagged more than a decade ago. In 2009, began building the Veterans Benefits Management System (VBMS), the paperless system that still processes disability claims today. A 2015 GAO review⧉ found 's development of VBMS wasn't driven by robust planning and set no goals for system response times or user satisfaction, prompting five recommendations. As of July 2026 -- 11 years later -- has implemented three of them, including a plan for completing VBMS. The other two were closed in 2017 after surveyed VBMS users without ensuring the survey's statistical validity and without developing the satisfaction goals had asked for.
View data as table
| Implemented | 3 |
|---|---|
| Closed without meeting the goal | 2 |
The AI itself carries its own risk
's own use cases point to why the stakes are real. Its Payment Redirect Fraud model uses AI⧉ to flag suspicious direct-deposit changes to veterans' benefits -- reports that one to two of every 1,000 direct-deposit changes are fraudulent. But has separately reported that generative AI systems largely remain "black boxes" whose own designers may not fully understand how they generate outputs, which can limit 's own ability to detect error or misuse, and can complicate identifying who is accountable for an automated decision. officials told they face their own obstacles to safe adoption: federal cybersecurity, privacy, and IT-acquisition policy that can block cross-agency collaboration; insufficient computational resources and funding; and data-security rules that can prevent generative-AI research in some mission areas.
The takeaway
- AI use cases grew more than 9 times over in two years. reported 40 in 2023, 229 in 2024, and 367 by December 2025, expanding on a program that paid $195 billion to over 6.9 million veterans and their families in fiscal year 2025.
- More than a third of the last five years' recommendations on the program aren't fully in place. 15 of 43 recommendations has made on 's disability compensation program since 2021 remain not fully implemented, including specific gaps in contracted-exam oversight and claims-processor training flagged years ago.
- The oldest gap predates the recommendation count is now tracking. A 2015 review of VBMS, the system disability claims still run on, found no goals for response times or user satisfaction; 11 years later, two of the five recommendations from that review were closed in 2017 without meeting the underlying goal, and it's unclear whether today's AI tools inherit that same gap.
All findings are from -26-109137, " Disability Benefits: Opportunities and Challenges to Modernizing Technology and Adopting AI," testimony of Dr. Sterling Thomas before the House Veterans' Affairs Subcommittee on Technology Modernization, delivered July 13, 2026 and publicly released July 10, 2026 -- read directly in full (22 pages), not just the Highlights summary. Recommendation-implementation tallies reflect 's status as of July 2026 as reported in this testimony; 's underlying reports on each topic (examiner oversight, claims-processor training, VBMS, and generative AI use and management) are cited within it and were not separately re-verified here.
Sources(1) ▾
- U.S. Government Accountability Office, VA Disability Benefits: Opportunities and Challenges to Modernizing Technology and Adopting AI (2026-07-13) — -26-109137, testimony of Dr. Sterling Thomas, Chief Scientist, before the House Veterans' Affairs Subcommittee on Technology Modernization, delivered July 13, 2026 (publicly released July 10, 2026). Read the full 22-page PDF directly (extracted with pdftotext -layout), not just the Highlights page Artemis sealed at capture time -- the full text carries the AI-use-case counts, the recommendation tallies by report, and the VBMS/2015 history that the Highlights summary only gestures at. Direct gao.gov PDF asset access intermittently returns HTTP 403; the Artemis-sealed copy of the product page serves as the one-click capture. gao.gov · original document
Comments
Always open. Logged-in readers can annotate paragraphs in place.
's Veterans Benefits Administration administers one of the largest federal disability benefit programs in the nation⧉, paying more than $195 billion to over 6.9 million veterans and their families in fiscal year 2025. To speed up and improve accuracy in processing those claims, the department is turning to AI fast: GAO testified July 10, 2026⧉ that 's reported AI use cases grew from 40 in 2023 to 229 in 2024 to 367 by December 2025. That expansion is arriving on top of a large, unresolved backlog of oversight recommendations for the program those tools are meant to serve.