An insurer uses Claude to draft the explanations that accompany claim decisions. A trained claims handler reads and approves every one before it goes out, and the letters make no reference to AI. Under Anthropic's Usage Policy, what is still owed?

The principle — Anthropic attaches two distinct obligations to high-risk consumer-facing uses, and meeting one of them thoroughly says nothing about the other.

Why the key is correct — insurance claims processing and coverage decisions are named in the policy as a high-risk use case. Two measures follow. The first is a human in the loop: a qualified professional in that field must review the content or decision prior to dissemination or finalisation. The claims handler satisfies that, and satisfies it well. The second is disclosure — where model outputs are presented directly to individuals or consumers, you must tell them you are using AI to help produce your advice, decisions or recommendations. The letters say nothing, so the second obligation is unmet.

Why the others are wrong — the requirement names one qualified professional rather than two; nothing in the policy asks Anthropic to approve a deployment in advance; and the measures it does impose are review and disclosure rather than retention. Each of the three would be a reasonable governance practice, and none of them is the thing the policy says is missing.

Remember this — a careful review process is easy to point at and easy to mistake for compliance. The reader-facing half is the one that gets forgotten, because it is the only obligation whose absence is invisible from inside the organisation. If Claude's output reaches a customer, ask both questions: who checked it, and were they told.

Sources — Anthropic's Usage Policy, on high-risk use case requirements.