VA Claims Automation and ADS: What AI Can and Can't Do
VBA’s claims-automation stack is real, it’s expanding, and the VA credits it — alongside aggressive hiring and mandatory overtime — with pushing the disability backlog below 100,000 in February 2026, the first time since 2020. But two things are equally true and the honest version of this page holds both: Automated Decision Support (ADS) is decision support, not decision automation — the human Rating Veterans Service Representative still decides every claim — and disability adjudication is textbook “high-impact AI” under OMB M-25-21, which means the guardrails aren’t optional. For a small SDVOSB AI shop, the realistic opportunity isn’t building the core engine (that flows to large primes); it’s the bounded, lower-risk slices around it, won with responsible-AI discipline as the differentiator.
This is the third and most consequential capability spoke under the VA AI modernization pillar. It sits downstream of document automation (inbound intake) and beside the RAG policy assistant (rater Q&A); this one is the adjudication workflow itself.
What is VBA’s claims-automation stack, exactly?
Automated Decision Support is a rules-based, robotic-process-automation system augmented with machine learning, OCR/NLP, and generative AI. It automatically retrieves a veteran’s records — including VHA medical evidence that raters used to pull by hand — and assembles them into an Automated Review Summary Document (ARSD), then either drafts an examination request or routes the claim “Ready for Decision.” It now spans more than 170 diagnostic codes and all of the PACT Act presumptive conditions, and VBA deploys it under a “Verify, Validate, Graduate” discipline first piloted in Boise. But the boundary is the whole point: the ARSD is a support tool, and a claims processor “must still take manual action to validate the claim, review the eFolder, and make a final decision.” VA’s Principal Deputy Under Secretary for Benefits told Congress in April 2026 that ADS “does not make any decisions and will not deny a claim.”
That line is true — and incomplete. A biased summary, a missing record, or auto-generated boilerplate can still shape what the rater sees, which is exactly why the M-25-21 guardrails below are mandatory rather than nice-to-have.
Is it actually working? The honest answer is “yes, and it’s disputed”
Both halves matter, and a page that only tells one is the kind of vendor blog a VA buyer has learned to discount.
What the VA credits to the full mix — automation, hiring, and overtime together: the backlog fell below 100,000 in February 2026, down 63% from January 2025; VBA processed a record 3 million-plus claims in FY2025; and average days-to-complete dropped from 141.5 to 80.7, a 43% decrease. The agency reported 12-month issue-level accuracy “over 94%,” its best in two years. Crucially, VBA also grew its claims workforce more than 50% since FY2021 — roughly 14,000 new hires — and raised mandatory overtime; the gains are not AI alone.
What the dispute looks like. At an April 2026 House VA Committee hearing, Rep. Tim Kennedy warned that “speed does not equal success,” citing 94% issue-level accuracy against an 83.31% claim-based rate; Rep. Maxine Dexter said service officers’ “impression is that errors are increasing” and entered into the record a claim containing apparently AI-generated content; Ranking Member Mark Takano said “one without the other is not success.” Two VA OIG reports give the concern teeth: a September 2023 review of ADS hypertension automation found that 27% of sampled claims “contained inaccurate and inconsistent determinations,” and a September 2025 review found a single senior VSR authorized about 85,300 claims — 19 times the national average — at 4.7 minutes each. And a planned “Smart Ratings Recommendation” tool that would have proposed ratings, not just summaries, has been “paused indefinitely,” per a VA spokesperson. VBA’s own accuracy target is 98%; the current rate sits below it.
The honest synthesis is the credibility move: the speed-versus-accuracy tension exists with or without AI, and the responsible path is to measure it relentlessly — which is itself a service an SDVOSB can sell.
Why is this “high-impact AI,” and why does that matter?
Because OMB M-25-21 defines high-impact AI as any system whose output is a principal basis for a decision with legal or material effect on access to a government benefit — and disability adjudication is the textbook case. VA carries more of this than any other agency: it accounted for roughly 145 of 227 governmentwide rights-impacting AI use cases in the most recent count. For claims work, the M-25-21 minimum practices attach in full: a pre-deployment impact assessment, independent and real-world testing, ongoing monitoring with the ability to pause, adequate human training, human oversight, and consistent remedies. That last one already exists for veterans as the Appeals Modernization Act review lanes — Higher-Level Review, Supplemental Claim, and Board appeal — with the Board of Veterans’ Appeals and the Court of Appeals for Veterans Claims as the human-accountability backstop behind any AI-shaped decision. The April 3, 2026 compliance deadline for operational high-impact AI is what makes this urgent for VBA buyers right now.
Where does a small SDVOSB actually fit?
Not in the core. The engine room is occupied: Booz Allen built the intelligent-document layer and a “Smart Search” tool that has ingested more than a billion veteran documents, and ADS itself is reportedly part of a large IBM contract. A solo-founder SDVOSB is not displacing that. The fit is the bounded, observable, lower-blast-radius slices around it — and several of them directly answer the accuracy concerns Congress and the OIG raised.
| Slice | Why it’s defensible | Fit |
|---|---|---|
| QA / accuracy-checking harness (“AI to check the AI”) | Directly answers the Kennedy/Dexter/OIG accuracy concerns | Highest political defensibility |
| Contention-to-condition classification microservice | Bounded, eval-able, feeds ADS without replacing it | Upstream link to document automation |
| RAG over rating policy (M21-1, 38 CFR) | Public corpus, lower data sensitivity | Sideways to RAG policy assistant |
| Eval harness / regression detection / test datasets | Maps 1:1 to M-25-21 “independent and real-world testing” | Easiest to scope in 90 days |
| Decision-support UI (ARSD usability) | Small, observable, low risk | Subcontract under SPRUCE |
| TERA-memo completeness checking (PACT Act) | VA has acknowledged the memo is error-prone | Narrow and fundable |
The natural lead is the QA harness — a tool that scores ADS outputs for accuracy, flags drift, and produces the monitoring evidence M-25-21 demands. It turns the political sensitivity of this domain into the reason to hire you.
How has this work been contracted?
Through the same lanes as the rest of the cluster. The marquee SDVOSB example is Aquia Nava II (an SDVOSB joint venture), which won a ~$42.75M SPRUCE task order to support and improve the digitized 21-526EZ disability application. On the sole-source side, PingWind (~$4.36M) and Huntridge Labs (~$3.8M) both won Benefits Intake Optimization work under the $5M Veterans First sole-source authority in September 2025. And the money is growing: the FY27 budget request proposes $130M for VBA claims-processing automation and AI, plus a department-wide decision-intelligence line. The best-fit NAICS is 541512 (computer systems design, $34M size standard), with 541519 — the code Aquia Nava II won under — and 541511 as alternates.
What about compliance — this is PHI-heavy work?
It is, and that raises the bar honestly. Claims evidence is medical records, so the production data lives in the VA Enterprise Cloud (VAEC) at FedRAMP High (AWS GovCloud or Azure Government), under the VA ATO process and VA Handbook 6500/6517, with the VA’s 60-day accelerated AI ATO as the on-ramp. This is where the standing Cloudflare caveat matters most: Cloudflare for Government is FedRAMP Moderate today, with FedRAMP High targeted for its AI tooling in 2026 — so production PHI ingest runs inside VAEC’s FedRAMP High boundary, not on Cloudflare’s Moderate one. A Cloudflare-native stack is appropriate for the lower-sensitivity edges — a public-policy RAG, an eval harness on synthetic data, a decision-support UI — not for production medical evidence. Truvisory builds to FedRAMP control families and VA Handbook 6500/6517 — FedRAMP-aware, and not CMMC-certified, because CMMC is a DoD program that doesn’t apply to VA work (more here).
Frequently asked
Does AI decide my claim?
Is this high-impact AI?
Can a small SDVOSB really win claims-automation work?
Why is responsible-AI framing the selling point?
Why is FedRAMP-aware enough, and where does the PHI go?
Working with Truvisory
Truvisory is an SBA-verified SDVOSB founded by a combat veteran, building working AI on a Cloudflare-native, FedRAMP-aware architecture, fixed-scope, in 90 days, human-in-the-loop by design.
If you’re a VBA program owner or contracting officer facing the M-25-21 high-impact-AI deadline, we can stand up a single bounded pilot — our lead offer is an AI QA/accuracy-checking harness for ADS outputs (“AI to check the AI”), with the monitoring and testing evidence M-25-21 requires as deliverables, on synthetic or VAEC-hosted data. Book a scoping call. For the procurement path, see the $5M sole-source guide and the VA AI modernization pillar; for the upstream and sideways capabilities, see document automation and the RAG policy assistant.