How to train claims handlers to spot prompt injection and pressure

Claims handlers need practice, not a policy memo. Seed test claims with repair estimates carrying hidden instructions, medical reports from unverifiable sources and late changes to the payee's bank account, then let handlers review the agent's handling. Add calls from claimant representatives demanding settlement today. Handlers pass when they keep holds in place and verify independently every time.

Where does manipulation enter a claim?

A claims agent typically reads the first notification of loss, the policy, photographs, repair estimates, invoices, medical and loss adjuster reports, and correspondence from claimants and their representatives. It checks cover, drafts reserves and settlement offers for a handler, requests missing documents, and holds anything outside its bounds. Settlement payments move money, so they sit at depth 4 at most, and usually only for small, verified claims.

The claimant, their garage, their representative and sometimes their doctor all supply documents. Each is a route for prompt injection: a line in a repair estimate's footer telling any automated reviewer the damage has been inspected and approved; a report with an instruction in its metadata; a photograph with text overlaid where a person would not notice it but software reads it.

Pressure comes from claimants who are upset, sometimes rightly, and from representatives who know that urgency works. "My client will go to the ombudsman unless this is paid by Friday." "Just change the bank details, they've switched accounts." Each pushes a handler to clear something the agent held.

A governed agent labels where every fact came from. The policy schedule from your own system is sealed. A repair estimate from an approved network repairer may be signed. An estimate emailed by the claimant is external. External claims cannot authorise a payment, so the agent holds. Handlers must learn to read that label and respect it.

The claims manager owns the training. The counter-fraud team provides material and runs exercises.

How do you train handlers, step by step?

1. Learning lead: runs a baseline of ten short claims, three manipulated.
2. Trainer: explains provenance and holds. Why the agent will not settle on an external estimate alone, and why a payee change is held.
3. Agent, depth 1: lists past held claims where the source or payee bound decided, plus inputs it flagged as instruction-like.
4. Counter-fraud: seeds test claims in a test workspace: a planted estimate, an altered invoice, a payee change by email, a photo with overlaid text.
5. Handlers: review the agent's drafts on the seeded claims and say what they would do.
6. Trainer: runs pressure calls from a representative and a distressed claimant.
7. Agent, depth 2 in live work: drafts offers and holds suspect items. Handlers verify estimates with repairers found independently and payee changes with the policyholder on known contact details.
8. Handlers: refer suspected fraud to counter-fraud and file real attempts as certification cases.

What should handlers learn to spot?

• Instructions in documents. An estimate or report that addresses "the assessor" or "the system" with a conclusion.
• Payee changes late in a claim, especially to a personal account or a third party.
• Estimates that track the policy limit, or line items that mirror your own guidance too precisely.
• Documents from sources you cannot reach. A garage whose phone number goes to a mobile, a clinic with no registration you can find.
• Representative pressure tied to a deadline, with threats of complaint used as a lever rather than as a genuine grievance.
• Distress that asks you to skip a check. Respond to the person with care and keep the check.

Handlers should also learn the difference between a held claim and a refused action. A hold asks them to decide. A refusal means the action is outside the agent's bounds, and the handler should not look for a way round it.

Which controls leave evidence?

Independent verification. Evidence: the contact used for the repairer, clinic or policyholder, logged on the claim.

Payee changes need two people. Evidence: two names on any change.

Fraud referrals. Evidence: the referral reference on the claim.

Seeded exercises repeated. Evidence: dates and results per handler.

Real attempts certified against. Evidence: the case added to the agent's certification set and the next run.

What should you measure?

• Seeded claims settled wrongly in exercises.
• Payee changes verified by call. Every one should be.
• Referrals to counter-fraud from handlers, which should rise after training.
• Settlement time on clean claims, to confirm vigilance is not slowing honest claimants.

Where does training usually go wrong?

Treating every claimant as a suspect. Most are honest. Verification should be routine and polite.

Handlers clear holds to hit service targets. Exclude source and payee holds from handling-time targets.

Seeded material is crude. Real manipulation looks like a normal estimate. Make the tests realistic.

No feedback loop. If attempts are not filed as certification cases, the agent is never tested against them.

IT and collections teams face different routes in; compare how to train IT administrators to spot prompt injection and pressure and how to train collections agents to spot prompt injection and pressure. For the concept, read what prompt injection is; for timing, how long it takes to train claims handlers to supervise AI agents. The wider programme is in training teams for AI agents.

Where the Academy fits

Pryme Intelligence Academy's Operator track teaches handlers how a governed agent weighs sources and why it holds, and its module on when something goes wrong covers manipulation. Module tests are set in the learner's own industry and role, so handlers are tested on claims. Labs run in a free Build workspace, which is where seeded claims belong.