Is AI deciding your VA claim? Here is what the public record shows so far.

Federal agencies must publish an annual inventory of their AI use cases. VA's 2025 inventory, released in January 2026 and updated in April 2026, lists 367 individual AI use cases across health care, benefits, and operations. Inventory entries tied to claims work include Smart Claim Check, described as helping Veterans Service Representatives identify and resolve issues in extensive claim documentation before they become errors, and an earlier inventory described Claims Profile, which builds computable profiles from historical disability data so adjudicators can compare a new claim against prior decisions faster. These are decision support tools by VA's own description: they organize and flag, and a person acts on what they surface.

VA's claims modernization materials describe automated decision support that pulls federal records, assembles evidence, and generates summary documents for claims processors, and state plainly that it does not replace the need for claims processors to review and make decisions. VA points to this front-end work when it reports speed gains: the agency completed a record volume of claims in fiscal year 2025 and announced on April 15, 2026 that average completion time had fallen to 80.7 days, a 43 percent drop from 141.5 days, crediting automation, workflow changes, and records-access improvements as contributors. The same front-end automation was part of PACT Act processing: VA said in 2024 that automation tooling helped it approve its first million PACT Act claims.

In April 2026, VBA's Principal Deputy Under Secretary for Benefits told Congress that claims automation, quote, will not make any decisions and will not deny a claim, and that, quote, our people are making the decisions. That is the clearest on-record statement the agency has made, and it matches the written descriptions in the use case inventory and the claims modernization pages. It is a policy position, not a statutory requirement, which is exactly the gap lawmakers have started trying to close.

One boundary is already documented: in April 2026 the VA inspector general reported that a separate rules based automation for certain survivor benefit claims had processed some claims from start to finish with no human involvement. From a statistical sample of 160 automated decisions, projected to about 8,100 eligible decisions, OIG estimated that at least 8,000 decisions or notification letters had at least one legal or procedural deficiency, such as omitted favorable findings or incomplete evidence summaries, and that at least 2 percent contained errors with monetary impact, at least 2.7 million dollars in improper payments; the evidence supported the grant in most cases. OIG's public tracker shows recommendations 2 and 3 closed as implemented on July 8, 2026 and recommendation 1 still open. That system is not one of the AI tools described on this page, but it shows the line between assistance and adjudication is drawn by policy and program design, not by a technical barrier.

The Artificial Intelligence Claims Evaluation System, AICES, appears in a privacy impact assessment VA signed in April 2026. Per the assessment, AICES evaluates whether a disability claim is suitable for an Acceptable Clinical Evidence, or ACE, DBQ, meaning whether the claim can be rated from records already in the file instead of scheduling a new in-person C&P exam, and it uses AI, natural language processing, and OCR to fill out DBQs for qualifying cases and upload them into VA's benefits system. The assessment states the system processes personally identifiable information including name, Social Security number, and date of birth, and retains data in AWS GovCloud for up to one year after processing. One wording tension in the document is worth knowing: the assessment says AICES analyzes claims data, quote, to make benefit decisions, while also stating that the DBQs it produces, quote, enable the government employee to make benefit decisions. VA's broader position that people make final claims decisions comes from its testimony and policy statements, not from this assessment. What the assessment itself establishes is that the tool can materially shape the evidence placed before the human who decides, including whether an in-person exam happens at all.

Separately, VA is building a fraud screening tool for newly submitted Disability Benefits Questionnaires. The two systems should not be conflated: AICES is an AI system by VA's own description, while VA says the DBQ fraud tool is not. Initial reporting on March 9, 2026 described an automated screening tool; VA clarified the following week that the tool does not use artificial intelligence, relies on manual data entry and analysis, is forward-looking, flags new submissions for further human review rather than automatically reducing or denying benefits, and is not revisiting finalized claims. Disabled American Veterans publicly raised concerns about the plan, and private DBQ providers have tracked it closely.

While discussing claims automation at an April 2026 hearing, VBA cited a 93.95 percent 12-month issue-based accuracy rate. VA's published issue-level accuracy metric is a national claims-quality measure covering the individual medical issues inside compensation claims; it is not a separate measure of claims processed with AI or automation, and VA has not published a distinct accuracy rate for automation-assisted claims. Lawmakers pressed a second distinction at the same hearing: issue-level accuracy is not claim-level accuracy. A claim contains many issues, so a small per-issue error rate can still leave a meaningful share of claims with at least one error, and VA's published claim-level accuracy runs well below the issue-level figure.

In May 2026, two House lawmakers proposed amendments to the FY2027 VA funding bill that would add AI guardrails, including a proposal by Rep. Paul Gosar to bar AI from making a final disability claim determination. The proposals show congressional concern about preserving human decision making, though they had not been enacted as of this report's latest update. On the health care side, the VA inspector general's final national review, issued in June 2026, found limited coordination with VA's patient safety office and no AI-specific mechanism for identifying safety events involving generative AI clinical documentation, a caution that shapes how the benefits-side rollout is being watched.

VA's generative AI guidance turns on authorization: tools VA has authorized for sensitive data, such as VA GPT and Microsoft Copilot Chat, may be used with that data, and entering VA-sensitive data, including personally identifiable information, into AI services without the appropriate VA authorization, including public web-based chatbots, is prohibited. Staff remain responsible for verifying AI output before relying on it. That rule cuts both ways: it confirms generative AI is in day-to-day use inside the agency, and it sets a data-handling standard that veterans would be wise to apply to their own records.