Click to Enlarge: Includes disability compensation and pension claims that have been received by VA that requires development and a decision by a VA claims processor. Source: Veterans’ Benefits Administration
WASHINGTON, DC — VA announced recently that its disability claims backlog fell below 70,000 for the first time since February 2020. That success has been partially attributed to technology modernization, which includes the use of AI in early stages of the benefits process.
While the department has moved to embrace AI in general, agency watchdogs warned that the more common forms of AI models might not be ready to accurately understand the nuance of a veteran’s medical and service history.
The COVID-19 pandemic, along with the increase in claims stemming from the PACT Act, swelled the claims backlog to an average of 200,000 between 2022 and 2025. A claim is considered backlogged when it’s taken longer than 125 days to process. According to VA, claims now take an average of 77 days to process, with three million claims processed in fiscal year 2025 alone.
Much of that success has been due to a five-year modernization plan mandated by the PACT Act. The plan, currently in its fourth year of implementation, has included the use of automation support tools for claims processing. VA uses smart search technology to scan and locate information within veterans claims evidence folders. It also uses AI to fill out the Toxic Exposure Risk Activity memos needed for PACT Act eligibility determinations.
Testifying before the House VA Technology Modernization Subcommittee last month, VA Deputy Chief Information Officer Robert Orifici assured lawmakers that AI is not making final claims decisions but acting only as a tool for human processors.
“While we use artificial intelligence to help our employees to make faster and better decisions … every disability claim is decided by a trained VA employee, not by AI or automation,” Orifici said. “These tools support human decision-making. They do not replace it.”
Even as a support tool, AI might prove problematic, however, when being used to parse through veterans’ records.
“AI poses a particular threat, a particular challenge, because it follows statistical patterns from large data sets. But veterans have experiences and needs that differ from those patterns, often dramatically,” explained Government Accountability Office (GAO) Chief Scientist Sterling Thomas, PhD. “As VA develops those systems, we need to ensure they use data sets that represent them. If we don’t, AI can lead to harm like denying compensation to deserving veterans. AI is still in its early stages of development and implementation. Rapid deployment without deliberate focus on governance has already led to unintended consequences.”
VA’s IT modernization practices have been on GAO’s high-risk list for years, and previous reports have found that a lack of proper oversight and testing have caused the department to fumble many high-profile technology rollouts.
“It’s critically important [to have] a pilot program to validate the tool works in the way that the vendor says it actually does. That it makes the adjudicator’s job actually easier,” Thomas told lawmakers. “That’s the challenge with a lot of these tools. [Employees] have to spend a lot of time correcting the mistakes of AIs. it actually makes their job harder.”
The debate over AI comes at the same time VA is preparing to settle a class-action lawsuit alleging the department mistakenly closed thousands of benefits appeals because of computer error. Prior to 2019, VA tracked its appeals using the Veterans Appeals Control and Locator System (VACOLS). After a number of veterans and surviving family members reported filing appeals and hearing nothing back from the VA, it was discovered that the system had been erroneously closing cases and never informing veterans. As many as 90,000 appeals closed between 1995 and 2025 could be reopened.
Under the proposed settlement, VA will automatically audit 28,558 files, while giving nearly 65,000 more veterans and servicemembers the opportunity to request a review.
