Every personal-lines insurer claims an AI roadmap for claims. The honest question is which pieces are in production at meaningful volume, and which remain pilots. After a year of observing the leaders and the laggards, a pattern is visible.
What is genuinely in production
- Computer-vision damage assessment for low-severity motor claims, with human review for anything above a confidence threshold. Cycle-time reduction is real; loss-ratio impact is small but positive.
- AI triage and routing at first-notice-of-loss, pushing volume claims to fast-track lanes and surfacing complex or potentially fraudulent claims to senior adjusters earlier. This is where the leaders are pulling away.
- Parametric instant payout for narrow product wedges — travel delay, defined-event weather — where the trigger is unambiguous and the claim is a payment, not a negotiation.
What is not yet production-grade
End-to-end agentic settlement of contested or complex claims. LLM-drafted decision letters without human review. Bodily-injury triage on AI signal alone. These appear in demos. They do not appear in the audit-trail-and-complaints-history of insurers we work with.
What separates the production cases
The shipped use cases share a structure: a narrow, bounded decision; a confidence threshold above which the AI acts and below which a human acts; a continuous monitoring loop on accuracy, complaints and Consumer Duty fair-value signals; and a documented model-risk treatment that the FCA could inspect on a Wednesday afternoon. The unshipped ones have impressive accuracy metrics and no governance scaffolding.
The lesson is consistent across the leaders: AI in claims is an operating-model problem, not a model-accuracy problem. The accuracy is a precondition. The operating model is what determines whether the use case ships at scale.