Until the VA fixes its governance failures, the system will keep failing the veterans it serves and the taxpayers funding it — and no amount of artificial intelligence will change that. AI is being sold as the fix. It is not. It is an accelerant on a fire nobody is willing to put out.
The VA outsourced the medical exams that decide disability claims to a $13 billion contractor market. Its own inspector general found those contractors deliver inaccurate exams — and the VA won't penalize them. Now it is scaling artificial intelligence across the same process it already can't govern.
That is the whole story. And it is not a veterans story — it is the clearest live case study in America of what happens when an organization automates a broken process instead of fixing it. Every company deploying AI is about to repeat it.
You cannot automate willpower. The VA doesn't have a technology problem. It has a governance problem — and it's spending to automate its way around it.
On paper, the VA is having its best year in a decade: backlog down more than 70%, average processing time cut from about 141 days to roughly 81, accuracy reported above 94%. Those numbers measure speed and volume. They say nothing about whether the right veterans were approved, whether denials were correct, or whether any of it is sustainable. That gap — between what the VA counts and what actually governs a benefits system — is exactly where it is now scaling AI.
Under the modernized appeals system, the Board of Veterans' Appeals grants at least some relief in more than a third of its 100,000-plus annual decisions, denies outright fewer than 20%, and remands a large share for further development. In plain terms: a VA denial is very often not the final word — the initial decision is frequently incomplete or wrong.
Then the oversight record. In one 2025 case, the VA Office of Inspector General found a single senior processor authorized roughly 85,300 claims over two years at an average of 4.7 minutes each — against a norm near 21 minutes — with management aware, and quality review touching less than 1% of the work. Of 32 sampled decisions, 27 had errors; in at least one, the processor never opened a document.
4.7 minutes per claim. 85,300 claims. Management knew. Quality review checked less than 1%. This is what "record processing speed" can actually mean.
A separate OIG review of the VA's rating software found staff overriding its decisions — over a third of certain overrides wrongful — and regional employees who didn't know the oversight tools even existed. Fast, high-volume, unaudited. Speed went up. Oversight did not.
This is the structural core, and it's a forensic accountant's kind of story. The VA no longer conducts most of the medical exams that decide claims. As of 2024, about 93% of Compensation & Pension exams are done by third-party contractors, not VA clinicians — up from a point in 2019 when only 25% were outsourced.
Those exams run through a small oligopoly: medical exam contracts extended to a total worth of roughly $13 billion, concentrated among five prime contractors, the largest alone holding about $5.1 billion in task-order spend. The contract structure pays mostly for administration and scheduling — claimant communication, appointment scheduling, records management, IT interfaces — with the medical evaluation itself a fraction of the total cost.
Then the finding that defines the whole problem:
In 2022 the VA's own inspector general found the exam vendors "failed to consistently provide accurate exams required by the contracts" — and faulted VA leaders for not penalizing them. As one attorney put it: they were "not docking contractors' money or assessing a penalty for mistakes."
The VA outsourced the medical judgment at the center of every claim to a $13 billion market, its own watchdog found the work inaccurate, and it declines to enforce the contract. That is not a software gap. It is a will-to-govern gap.
Why not fix it? Because hard decisions have become politically impossible. In 2024 the VA hit an unprecedented benefits shortfall; Congress passed an emergency $3 billion supplemental amid open finger-pointing. Any attempt to restructure ratings or hold contractors accountable hits the same wall: anything that could touch a benefit is cast as betrayal, and reform dies.
So the system drifts and spends. Mandatory benefits funding jumped $34.2 billion in one year to $301.2 billion; the Toxic Exposures Fund alone runs from about $20B in 2024 toward a projected $52.6B by 2026. Meanwhile the 2026 budget proposes cutting roughly 2,900 staff positions, most from the benefits administration.
There is also a structural reason the contractor arrangements go unchallenged. The firms that dominate the VA's multibillion-dollar exam and IT contracts are among the largest federal lobbying and PAC operations in the country, filing quarterly lobbying disclosures and maintaining political action committees. Political-science research on campaign finance is consistent on the underlying dynamic: PAC money reliably flows toward the incumbent committee members who oversee the programs an industry depends on, because contributions buy access to the legislators positioned to shape that oversight. That is the textbook shape of regulatory capture: the entities being overseen are financially entwined with the process meant to oversee them. One does not need to allege a specific quid pro quo to recognize that a system built this way has weak structural incentives to police its own vendors — and the OIG's finding that the VA won't penalize failing contractors is exactly what that weakness looks like in practice.
Spending is growing faster than the veteran population. Staff is being cut. AI is filling the gap. That is the fiscal signature of a system substituting automation for governance — the ticking part of the time bomb.
An ungoverned system invites exploitation. Into the space left by an overwhelmed VA stepped a multibillion-dollar industry of unaccredited "claims consultants." The VA sent 40-plus warning letters over a decade telling these companies to cease charging veterans, which may violate federal law; the companies only grew — protected by a 2006 removal of criminal penalties advocates say let them proliferate "like having no penalty for speeding."
But it cuts both ways, and honest analysis holds both truths. The same system that can't police the profiteers also wrongly denies legitimate veterans at scale — roughly 900,000 claims denied in FY2024 — while fraud investigations remain a rounding error against 6.9 million beneficiaries. The failure isn't mainly fraud by veterans. It's a system that can neither reliably approve the deserving nor penalize the profiteers — because it cannot measure and will not enforce.
Now layer AI on. The VA's 2025 inventory lists 367 AI use cases, 28 of them in benefits processing, most still pre-deployment — with new funding of roughly $130 million for VBA automation and AI plus $47.8 million for "Decision Intelligence and Automation." The flagship tool, Automated Decision Support, assembles claim evidence for processors. VA officials testified it "does not make any decisions and will not deny a claim" — a reassurance House lawmakers openly disputed, questioning whether wider automation is raising errors.
Automating an ungoverned process doesn't fix it. It scales the errors at machine speed — and hands everyone a new excuse: "the system flagged it."
Note the irony buried in the budget line: the new AI funding is justified in part to "sustain governance frameworks for safe, effective, and ethical use." The VA is budgeting for the governance its own inspector general says it isn't delivering. If a rating tool can be overridden by someone who doesn't know the oversight exists; if contractor exams are inaccurate and unpenalized; if most appealed denials get reversed or remanded — then accelerating that pipeline with AI doesn't produce better decisions. It produces the same flawed ones, faster, with thinner human judgment.
The governance questions are not "does the AI work?" They are: Who audits its decisions? Who is accountable when an automated denial is wrong? What is the appeal path against an algorithm? Who governs the contractors feeding it data? The VA is scaling AI while its own watchdog flags governance as a top failure.
Strip away the subject and the VA is a generic AI-governance failure — the pattern every company adopting AI is about to reproduce:
The lesson isn't "don't use AI." It's that governance — auditability, accountability, human oversight, vendor control — has to exist before automation scales, not after the errors are already moving at machine speed. The VA is the cautionary case study. The only question for your organization is whether you're building that layer now, or planning to explain its absence later.
That is the class of problem this practice exists to address — auditability, accountability, and vendor control, built in before AI scales.
For veterans navigating the denial and appeals process itself, practical filing guidance is available at VCAnalytics.ai.