Tell the Veteran: The AI Disclosure Rule We Adopted Before It Was Required
On November 12, 2026, updated usage policies from Anthropic — the AI company whose models power Guidon and much of the professional software world — take effect. For AI used in high-stakes domains, which explicitly includes public benefits and medical information, three things become mandatory: a qualified human reviews the output before it affects anyone, someone identifiable is accountable for the decision, and the individual affected is told that AI was involved. If you run a veterans service office and any vendor is bringing AI through your door, those three requirements are now the floor. This post is about what they mean in claims work — and what we changed the week we read them.
Two of the three were already the architecture
Qualified human review and clear accountability aren't new rules for us — they're the product. Every Guidon output is a draft until a certified or accredited service officer reviews, amends, and signs it. No auto-submit exists in the product. The audit log records every sign-in, upload, and report view under the officer's own name, so accountability is never a guess. This has been the Pledge since the first pilot, and the governance model we published in September. When the new policy landed, we reviewed every agent in our pipeline against it. Those two requirements: already met, by design rather than by patch.
The third one is the interesting one
Disclosure to the individual. In claims terms: the veteran should know AI-assisted analysis was used in preparing their claim.
We'll be honest about the industry's instinct here, because it used to be half of ours too: nobody wants to open that conversation. The fear is that "AI" sounds like a corner was cut. So the quiet path is tempting — the tool stays in the back office, the veteran never hears a word, and everyone hopes it never comes up.
We think the quiet path is exactly backwards, and that it's how this industry eventually earns itself a scandal. A veteran who finds out afterward that software read their file has every right to ask what else they weren't told. A veteran who's told upfront — "software read all 1,200 pages so I didn't have to skim; every conclusion you see, I reviewed and signed" — hears the truth, which is that their representative used a better flashlight and stayed in command. One of those conversations costs trust. The other one builds it.
It's their claim. Their records, their conditions, their benefits, their decade of waiting. They're entitled to know how the work was done — and the work, done this way, holds up proudly to the telling.
What we changed
Two lines, shipped the same week we read the policy. The sign-off card at the end of every Guidon report — the screen where the officer takes authority over the draft — now tells the officer directly: if this analysis shapes what you deliver to the veteran, let them know AI-assisted analysis was used in preparing it. And step five of the officer quick-start guide says the same thing, so disclosure is part of the workflow from day one, not a memo nobody reads. Small changes. That's the point — when the architecture already respects the people involved, compliance with a rule like this is two sentences, not a retrofit.
If you're evaluating any AI vendor after November 12
Add three questions to the seven we published in June: Who reviews the output before it reaches a veteran, and is that review required by the product's design or merely suggested? Whose name is attached to each action in the system? And what does the vendor recommend you tell the veteran — because "nothing" is now the wrong answer, both by policy and by principle.