For more than a decade, the U.S. Department of Veterans Affairs has been using automation tools to manage one of the most complex benefits systems in the country. Yet as the agency increasingly relies on AI for decision-making, the promise of efficiency is colliding with a growing rate of error, potentially harming veterans, said veterans rights attorney Benjamin Krause.
“The agency is overly relying on AI rollout to accomplish more than it can safely accomplish,” he said.
A System Built for Speed
Within the VA’s digitized claims process and the Veterans Benefits Management System, automated tools review records, extract data and generate decision recommendations.
While the technologies these systems use, including optical character recognition and natural language processing, are generally considered artificial intelligence, the VA labels them “automation” or “bots.”
This ambiguity has led to the uneven application of safeguards, audits and reporting requirements. “When you throw words around like that, your oversight of those systems can be complicated and confusing for Congress and watchdogs,” Krause said.
Errors at Scale
Given the volume and sensitivity of VA claims, the stakes for legal error are high.
Disability compensation cases often involve thousands of pages of medical records, service histories and expert opinions. AI systems are tasked with synthesizing disability information and assisting adjudicators. In some instances, Krause said, automated tools may influence decisions without meaningful human review.
The results are sometimes “nonsensical outcomes,” he said. He pointed to a case in which a veteran seeking benefits for a combat-related traumatic brain injury was denied for reasons tied to toxic exposure (not part of his claim).
Human Oversight and Wrongful Allegations
Although VA automated systems are subject to human oversight, employees tasked with reviewing AI-generated outputs face heavy caseloads and performance quotas.
“They don’t have time to look through a file that’s 10,000 pages,” Krause said. “They often look at what the system says and hit go.”
A recent VA Office of Inspector General review underscored the concern. In examining claims processed under the Honoring our PACT Act, the watchdog identified significant error rates involving tens of thousands of veterans and millions of dollars in benefits.
The VA has also explored large-scale data analysis to identify fraud patterns among veterans. Although intended as a safeguard, the results triggered a high rate of false positives, subjecting veterans to increased scrutiny.
“If you flag someone incorrectly, the downstream effects can be devastating,” Krause said. Wrongful fraud allegations within the VA have resulted in sick or elderly veterans losing medical and other benefits for significant periods of time.
Data Privacy and Oversight
The VA’s reliance on private contractors introduces additional legal and ethical questions.
The agency’s systems integrate platforms developed by private technology vendors and hosted on commercial cloud infrastructure, potentially allowing sensitive veteran data to pass through multiple private entities before reaching a final decision.
Oversight can also be challenging in a hybrid public-private environment. Traditional accountability mechanisms, such as Freedom of Information Act requests, may not fully apply to private contractors operating under government agreements.
Krause noted the VA Office of Inspector General only began focused review of AI systems in the past few years, long after many tools were already in use.
He suggested effective regulation will require greater congressional investment in oversight, as well as clearer definitions and greater transparency.
Where AI Can Help — and Where It Can’t
For Krause, AI is a helpful tool in contexts such as scheduling appointments or organizing data. “Then it is efficient to use it,” he said.
He draws a line, however, at replacing human judgment in complex cases: “You will never truly replace the human experience, especially in systems dealing with disabilities and life-altering benefits.”
Moving forward, he urged veterans to review decisions and seek guidance from licensed representatives. Legal professionals, meanwhile, should remain vigilant in scrutinizing AI-influenced outcomes.
“AI works well when it works well,” Krause said. “But it can also be completely wrong.”