Is AI Deciding Your VA Claim?

The short answer is no. A person still signs your rating decision. But software now reads your file before that person does, and what it reads it for is worth knowing.

Everything below comes from VA's own documents. Where VA has said something plainly, I quote it.

New here? Everything under Decisions is free: one page per condition and one per pairing, each built from what the Board actually did. Nothing to sign up for.

How much AI the VA is using

VA publishes a list of every AI system it runs. The current one names 367 uses across the department, from health care to benefits. It is public, and anyone can read it.

Two of those touch disability claims directly. One is Automated Decision Support. The other is the Artificial Intelligence Claims Evaluation System, which VA shortens to AICES.

What AICES actually does

This is the one veterans hear about, and it is usually described wrongly. It does not write a summary for your rater, and it does not decide anything. Here is what VA says it is for, from the privacy assessment its own benefits staff filed in April 2026.

AICES will evaluate the feasibility and effectiveness of using artificial intelligence (AI) to assist in triaging Veteran Benefits Administration (VBA) disability claims for potential Acceptable Clinical Evidence (ACE) eligibility.

Acceptable Clinical Evidence means rating a claim from the records already in your file, with no new in-person exam. So AICES looks at your file and sorts claims by whether the records are already enough. The same document says it uses character recognition and language models to fill in Disability Benefits Questionnaires for the claims that qualify.

That matters to you in one practical way. It is not deciding whether you are right. It is helping decide whether you get called in for an exam.

The document also says what happens when the software cannot read something confidently.

Data which cannot be accurately extracted and validated by the above methods are off ramped for humans to process.

The tool that reads old questionnaires

In March 2026 VA said it would run a fraud-detection tool over more than a million Disability Benefits Questionnaires going back to 2010. That news travelled badly, so here is what VA said about it, in its press secretary's words.

This tool is forward-looking only. VA will not use the tool to revisit previously finalized and processed DBQs.

No veteran's claim or benefit will be reduced or denied because of this effort.

VA says the target is unaccredited businesses that sell claim help, not veterans. The patterns it looks for are plain enough. The same boilerplate wording across many questionnaires. A missing signature block. An examiner based more than 100 miles from the veteran.

Read that list again, because it is useful. Every one of those is also what makes a questionnaire weak in front of a human rater. A form that could have been written about anybody persuades nobody.

Will you be told?

Not separately. The privacy assessment says AICES does not collect anything from you directly. It takes what is already in VA systems, and your notice came from those systems when you filed. So you will not get a letter saying software read your file.

Who is checking

In March 2026 a member of Congress wrote to VA with three questions. How many claims had been processed with Automated Decision Support or AICES. How many of those were later found to have quality errors. At what point each error was caught. That letter is public.

Those questions are the right ones, and they are not answered yet.

About the accuracy number

You may have seen that VA gets 17 percent of claims wrong. That comes from a real VA figure, but VA publishes two and they measure different things.

Claim-based accuracy counts a whole claim as correct only when every issue in it is correct. That one runs near 84 percent. Issue-based accuracy grades each condition separately, and that one runs near 94 percent, against a VA target of 96. Both come off the same public dashboard.

Neither number is about AI. They were measured the same way before any of this software existed. They are a reason to check your own decision letter closely, which was already true.

What to do about it

The advice does not change much, which is the point. Evidence that reads clearly to a person also reads clearly to software.

1. Have the doctor who actually examined you fill in the questionnaire, and have them complete the whole signature block, including their address and credentials.
2. Ask for specifics about you. Dates, measurements, what you cannot do now that you could before. Wording that could describe anyone is the pattern VA flags and the pattern a rater discounts.
3. Send clean copies. Your file is read by character recognition before anyone reads it. A crooked scan of a handwritten note is hard for the software and hard for the rater.
4. Get the records in the file before you file. Since AICES sorts on whether your records are already enough, a complete file is what keeps you out of an unnecessary exam. Step 1 covers which records to pull.
5. Read your decision letter against your evidence. If something is wrong, the routes are the same as they have always been, and the appeal page walks through them.

Every condition on this site has its own free page under Decisions, built the same way: what the Board did, how often, and the decisions behind it.