How to keep claims handlers ready for an AI agent’s 90-day recertification

Focus every cycle on money and declines. Each 90 days the claims agent is recertified on your own claim files, and the result decides which settlements it may pay, which it must hold and which fraud signals it escalates. Claims handlers need to hear those changes before their queue shows them, practise on real-looking files, and know the fallback if a run fails.

What changes for claims handlers when the agent recertifies?

A claims agent usually works across four capabilities. It reads the first notice of loss and the policy at depth 1. It drafts coverage letters, reserve suggestions and requests for information at depth 2. It may settle small, clean claims at depth 4, because paying a claim moves money, but only within the insurer's own limit and only if certification supports it. Anything above the limit is refused: the agent may only draft it, and a handler makes the payment.

Recertification re-runs the agent's cases on recent claims and scores them against a floor. Payment authority is the most likely to change, because the cases that test it are the hardest: duplicate invoices from repairers, estimates that arrive by email, and policies with endorsements the agent may misread. A repairer's emailed estimate is external provenance and cannot by itself authorise a payment.

If a run clears no depth, the agent drops to depth 1 the same day. Every claim it would have settled arrives as a draft, and handlers carry the full volume until the next successful certification. More on the general mechanism is in what recertification is.

Who owns readiness?

The claims team manager owns handler readiness and the rota. The builder who certifies the agent owns recertification. The claims technical lead sets the insurer's own limits, such as the settlement value above which the agent may only draft a payment. They should agree the change note together a week before the certificate's expiry.

The cycle, in eight steps

1. Person (claims manager): fix the date clear of peak events such as storm surges.
2. Agent, depth 1: produce the cycle report. Claims read, letters drafted, settlements paid and held, fraud escalations and refusals, with the deciding bound for each.
3. Person (technical lead): audit a sample of paid claims. Check that each payment was within the policy and within the limit.
4. Person (builder): recertify on recent files, including the audited ones.
5. Person (builder and claims manager): write the change note: settlement authority before and after, any new settlement limit, new escalation triggers, any letter that moved depth.
6. Agent, depth 2: draft practice files that show each change, checked by the technical lead.
7. Person (claims manager): brief handlers with the files; each handler states the decision and the reason.
8. Person (technical lead): sample the first two weeks of held settlements cleared under the new certificate.

What should handlers retrain on?

Held settlements. A handler clearing a held payment should read the deciding bound first. "Refused: above customer limit" means a handler makes the payment. "Held: source external" calls for verifying the repairer's estimate against a known contact or approved list.

Declines. The agent never declines a claim. Handlers should be confident about what the agent's evidence summary contains and what it may have missed, such as a later endorsement.

Fraud escalations. A change in fraud indicators often arrives with recertification. Handlers need two or three examples of the new pattern, not a list of rules.

The fallback. Every handler should know the rota if the agent drops to depth 1.

Controls and the evidence they leave

• Paid-claim audit per cycle. Evidence: the sample, results and any recoveries.
• Change note agreed by three owners. Evidence: signed, dated note.
• Handler briefing. Evidence: attendance and answers to the practice files.
• Named approval on every held settlement. Evidence: the approval, approver and certificate status in the append-only record.
• Practicum for new approvers. Evidence: the practicum log template for claims handlers.

Pryme Intelligence Academy's Operator track, which covers human review and when something goes wrong, is one way to give approvers the certificate behind those approvals.

What to measure

Three figures. The number of paid claims in the audit sample that should have been held. The time held settlements wait before a handler clears them, before and after the cycle. And the edit rate on coverage letters. Watch customer complaints that mention speed too: if settlements slow after a drop in depth, customers notice first.

Where it usually goes wrong

Handlers trust a paid claim because the agent paid it. If the audit never runs, an error in settlement logic can repeat across many claims before anyone notices.

The limit moves without the handlers knowing. If the technical lead lowers the settlement limit, handlers suddenly see more claims left for them to pay and assume the agent is broken. Put limit changes in the change note.

Handlers rubber-stamp held settlements during a surge. Volume pressure is when the deciding bound matters most. Accountability is not shifted to the agent, as are claims handlers accountable for an AI agent's mistakes explains. Operations leads face the same rota questions; see keeping operations managers ready for recertification and the guide to AI literacy and certification for people who work with agents.