AI Observability

On a Tuesday morning an adjuster in the insurer's Lyon office opens a draft coverage recommendation for a water-damage claim and stops at the second paragraph. The draft declines part of the claim under "exclusion 4(b), gradual seepage", and the policy wording attached to the claim has no clause 4(b). She rejects the draft and files a ticket. The platform team opens its dashboards. The claims API answered with HTTP 200, time to first token was 1.2 seconds against a 2-second target, the model provider reported no errors, and availability for the month stands at 99.97%. Every signal the team watches is green. None of them can say which prompt version wrote that paragraph, which clauses the retriever handed the model, whether the citation check ran, or how many of the roughly 7,000 recommendations drafted that week carry the same invented clause.