VA’s AI Push Cut Disability Claim Times 43%.
Its Own Inspector General Found a 98% Error Rate in an Automated Benefit.
The Department of Veterans Affairs is running one of the federal government’s most aggressive artificial-intelligence rollouts, and by its own numbers it is working: the agency reported running 229 AI use cases in 2024, up from 40 the year before, and its FY2027 budget submission counts 367 in its most recent inventory. Average disability-claim processing time has fallen 43%, from 141.5 days when the current administration took office to as low as 78.6 days by the end of May 2026. The claims backlog — veterans waiting 125 days or more for a decision — dropped below 70,000 in July 2026, its lowest point in six years.
VA’s own Office of Inspector General has also documented what automation looks like when it goes wrong. A review of 8,100 automated decisions on Dependency and Indemnity Compensation claims — survivor benefits paid to the families of veterans who died from service-connected conditions — found errors in 8,000 of them, a 98% error rate. At least $2.7 million of that involved confirmed overpayments; one case paid out more than $22,000 automatically after the system attributed a veteran’s death to a service-connected condition without verifying the underlying medical evidence.
Congress’s nonpartisan watchdog, meanwhile, says VA still has 27 unimplemented IT recommendations standing between it and safer AI adoption. This page sources every figure to the underlying GAO reports, the OIG audit, and the congressional hearing record, and treats those two kinds of oversight — GAO’s forward-looking findings and the OIG’s case-by-case audit — as distinct rather than interchangeable.
- 229 — AI use cases VA reported operating in 2024, up from 40 in 2023 — VA's FY2027 budget submission counts 367 in its latest inventory · Source: GAO-25-108739; VA FY2027 Budget Submission
- 43% — reduction in average disability-claim processing time, from 141.5 days (Jan. 20, 2025) to as low as 78.6 days (end of May 2026) · Source: Military.com; VA claims data
- Under 70,000 — claims still pending past 125 days as of July 2026 — a 74% drop since the backlog Secretary Doug Collins (R) inherited in January 2025 · Source: Navy Times
- 8,000 of 8,100 — automated survivor-benefit (DIC) decisions a VA OIG audit found contained errors, including omitted legally required favorable findings · Source: VA OIG Report 25-00153-47
- $2.70M+ — in confirmed overpayments from those errors — one case paid out more than $22,000 automatically on faulty rule-based logic · Source: VA OIG Report 25-00153-47
- 27 — VA IT recommendations GAO still counts as unimplemented — 26 on IT resource management, 1 on updating VA's AI inventory · Source: GAO-25-108739
The headline numbers are real, and they are large. VA’s AI use cases grew nearly six-fold between 2023 and 2024 — from 40 to 229 — while its generative-AI applications went from a single pilot to 27, according to GAO Director Carol C. Harris’s September 2025 testimony. VA’s FY2027 budget submission puts the department’s most recent AI inventory at 367 use cases. On July 9, 2026, VA announced its claims backlog — veterans waiting 125 days or longer for a decision — had dropped below 70,000 for the first time since 2020, a 74% decline since January 20, 2025. Just 11.6% of all pending claims are now older than 125 days, compared with 70% of the roughly 600,000 claims pending a decade ago.
VA Secretary Doug Collins (R), a former U.S. Representative from Georgia, took ownership of the numbers directly: “This is what putting veterans first looks like,” he said announcing the backlog milestone. “We are incredibly proud of these historic numbers, which mean faster decisions, better service and more benefits for the men and women who have worn the uniform.” Average processing time has fallen from 141.5 days on Jan. 20, 2025 to 78.6 days by the end of May 2026 — VA itself has described the reduction, measured against an earlier April benchmark of 80.7 days, as a 43% drop.
“This is what putting veterans first looks like. We are incredibly proud of these historic numbers, which mean faster decisions, better service and more benefits for the men and women who have worn the uniform.”
VA Secretary Doug Collins (R), announcing the backlog milestone, July 2026
VA's claims backlog has dropped below 70,000 for the first time in six years — a 74% reduction since January 2025. Faster decisions, delivered with accuracy, remain the top priority for every veteran who has earned these benefits.
Most of the speed gain traces to one system: Automated Decision Support, built under a $485M contract with IBM. ADS sorts through a veteran’s file — medical records, service history, records pulled from the Defense Department — and compiles a summary sheet for the human claims processor who ultimately decides a disability rating. Since IBM came aboard, VA says more than 21.2 million packets have been processed, that claims which once took 27 days to establish now take about 12 hours, and that the program has saved 6.4 million work-hours across 16 benefits offices, covering more than 40 disability conditions.
ADS is not new, and neither are the warnings about it. A 2023 VA OIG report already found the system “failed to recognize duplicate evidence, identified false evidence, and missed relevant information,” leading to inaccurate decisions on veterans’ claims — three years before the more recent DIC-specific finding described below. GAO’s April 2026 review of federal AI acquisitions, GAO-26-107859, examined 13 AI acquisitions across four agencies, including VA, and found officials across the board struggled to access AI technical experts — data scientists capable of evaluating whether a contractor’s AI proposal actually does what it claims — when writing evaluation criteria. VA was also one of the agencies GAO found was not yet systematically capturing lessons learned from its AI acquisitions, meaning the same procurement mistakes are not guaranteed to get caught the second time.
VA says a new automated tool will scan roughly a million disability benefits questionnaires dating back to 2010 to flag patterns of fraud — mainly targeting for-profit 'claims sharks,' not individual veterans. VA says no claim will be denied because of the effort alone.
Testifying before the House VA Subcommittee on Technology Modernization on September 15, 2025, GAO’s Carol C. Harris laid out five areas where VA’s AI ambitions are outrunning its ability to manage them: difficulty complying with federal AI policy, insufficient technical staff and budget for generative AI specifically, gaps in hiring an AI-capable workforce, data-security and privacy risk, and difficulty running AI acquisitions well. GAO tied those findings to 27 VA IT recommendations that remain unimplemented — 26 on IT resource management and one directing VA to keep its AI use-case inventory current. Rep. Tom Barrett (R-MI), the subcommittee’s chairman, and Rep. Nikki Budzinski (D-IL), its ranking member, have both pressed VA on the pace of implementation.
A separate GAO review, GAO-26-108789 — published October 29, 2025 and publicly released that November — found VA’s disability compensation program has sat on GAO’s High-Risk List since 2003, and that two decades of reform efforts have not consistently delivered the improvements they promised. A third, GAO-26-108844, based on Elizabeth H. Curda’s January 14, 2026 testimony, found VA decisions are still, in part, based on outdated rating criteria — a structural gap that predates the AI push and that automation, on its own, does not fix.
27 unimplemented IT recommendations — 26 on IT resource management, 1 on keeping VA's AI-use-case inventory current. Source: GAO-25-108739.
High-Risk List since 2003 — VA's disability compensation program has been flagged for two decades running. Source: GAO-26-108789.
Outdated rating criteria — VA decisions still partly rely on criteria GAO says need updating, independent of automation. Source: GAO-26-108844.
AI acquisition gaps — VA was among four agencies GAO found struggled to access AI technical expertise and was not systematically capturing acquisition lessons learned. Source: GAO-26-107859.
If GAO’s reports describe risk, the VA OIG’s April 30, 2026 audit describes what already happened. Investigators reviewed 8,100 automated decisions on Dependency and Indemnity Compensation claims — survivor benefits paid to spouses, dependents, and parents of veterans who died from service-connected conditions — and found errors in 8,000 of them, a 98% error rate. Most of those errors were procedural: award letters that omitted legally required favorable findings, or that failed to properly summarize the evidence VA relied on to reach its decision.
A narrower slice — roughly 2% of the decisions reviewed — involved legal errors with real financial consequences: at least $2.70 million in confirmed overpayments to unverified or unqualified applicants. In one case, the system paid out more than $22,000 automatically after attributing a veteran’s death to a service-connected condition without verifying the underlying medical evidence — the OIG traced the failure to flawed rule-based logic, including cases where the system granted service connection for pulmonary hypertension based solely on a documented history of hypertension, without confirming the more specific diagnosis. The OIG also found VA’s quality-review process for automated death-benefit decisions was less rigorous than the review traditionally processed claims receive, and questioned whether the modernization plan VA submitted to Congress fully complied with Section 701(b) of the PACT Act.
8,100 automated survivor-benefit (DIC) decisions reviewed by the VA OIG.
8,000 (98%) contained errors — most involving award letters that omitted required favorable findings or evidence summaries.
~2% of decisions involved legal errors with financial consequences, totaling at least $2.70 million in confirmed overpayments.
$22,000+ paid automatically in a single case, on faulty rule-based logic the OIG says went unverified.
Report: VA Office of Inspector General No. 25-00153-47, “Review of Automated Decisions for Veterans' Service-Connected Death Claims,” published April 30, 2026.
A VA Inspector General review found errors in 8,000 of 8,100 automated decisions on survivor death-benefit claims — a 98% error rate, with more than $2.7 million in confirmed overpayments. VA is processing claims faster than ever, but this is what some of that speed cost.
At an April 16, 2026 House hearing on VA’s Automated Decision Support tool, Margarita Devlin, VBA’s Principal Deputy Under Secretary for Benefits, told lawmakers claims accuracy stood above 94% as of the end of March — “the highest it’s been in two years,” she said, alongside Sandra Flint, VBA’s Deputy Under Secretary for Field Operations. But that 94% figure is what VA calls “issue-level” accuracy — a measure of individual line items within a claim. The figure for full claims packets, checked end to end, was 83.31% over the same period. Rep. Tim Kennedy (D-NY) pressed the distinction directly: “speed does not equal success,” he told the panel, questioning why VA leads with the more favorable of the two numbers.
“Speed does not equal success.”
Rep. Tim Kennedy (D-NY), House hearing on VA's Automated Decision Support tool, April 16, 2026
Rep. Maxine Dexter (D-OR) told the same hearing that veteran service officers in her district reported their impression was that “errors are increasing” even as decisions get faster, and submitted for the hearing record a claim decision that cited a Google search to explain a veteran’s employability, followed by language she said “certainly appears” to be AI-generated. Nearly six weeks earlier, on March 10, 2026, Rep. Steve Cohen (D-TN) sent a formal letter to Secretary Collins citing VA’s own February 28, 2026 claims dashboard, which showed three-month claims-based accuracy at 81.73% — “nearly 1 in 5 claims” containing an error, against a ten-year historical average closer to 1 in 9. Not every member of the committee shared the concern: Rep. Morgan Luttrell (R-TX), who chairs the Subcommittee on Disability Assistance and Memorial Affairs, said the panel’s general sentiment supported “the injection of AI models into the system.”
Accuracy is not the only front where AI has drawn scrutiny. In February 2026, VA told a House subcommittee it was building an automated tool to scan roughly a million Disability Benefits Questionnaires — medical forms doctors submit supporting a veteran’s claim — dating back to 2010, flagging patterns like boilerplate language repeated across many forms or questionnaires from doctors located more than 100 miles from the veteran’s home. The tool is aimed at “claims sharks” — unaccredited, for-profit firms that charge veterans to file paperwork — not individual veterans. Disabled American Veterans, led by National Commander Coleman Nee, pushed back, asking VA how the tool would be validated before touching veterans’ files, what specifically would trigger a flag, and how affected veterans would be notified.
By March 16, 2026, VA had narrowed the plan: the historical DBQ data would train the fraud-detection model, but the tool itself would not revisit previously finalized claims — only flag newly submitted questionnaires going forward. “This initiative will not change how VA evaluates or decides claims,” VA press secretary Peter Kasperowicz said. “No veteran’s claim or benefit will be reduced or denied because of this effort.” That revision — a public walk-back after direct advocacy pressure — is itself a data point: it happened before the OIG’s DIC findings became public, meaning VA had already faced one round of scrutiny over an automated tool’s reach before the 98% error rate ever surfaced.
The money behind all of this keeps growing. VA’s FY2027 budget request asks for $130M for “automation and artificial intelligence investments modernizing veterans claims processing by reducing errors and delivering benefits to veterans faster,” including $47.80M earmarked specifically for “Decision Intelligence and Automation” — a 10.9% increase over FY2026 enacted levels, driven mainly by what VA calls its “AI Infrastructure solution.” That request sits inside a compensation and pension program that Congress funded at $161.85B in FY2024 and $195.85B in FY2025 — meaning the AI line item is small relative to the whole program, but it is the piece both parties in Congress are now asking the most pointed questions about.
VA's AI-driven modernization has produced a real, measurable result: faster decisions and a claims backlog at its lowest point in six years, numbers Secretary Doug Collins (R) has staked his tenure on. It has also produced a real, documented failure: a VA Inspector General audit found errors in 98% of the automated decisions it reviewed on survivor death-benefit claims, with confirmed overpayments running past $2.7 million. GAO says 27 IT recommendations remain unimplemented, and lawmakers from both parties — Rep. Tim Kennedy (D-NY) pressing on accuracy, Rep. Morgan Luttrell (R-TX) backing the technology — are still arguing about which number matters more. Both numbers are real. Neither cancels the other out.
