A denied claim needs a reason that still stands when it is read back.
A claim reaches a person on their worst day, and a regulator soon after. When AI sits in that path the question is not whether the model is good. It is whether you can name the rule that applied and show it was the rule in force.
Building AI for insurers rather than underwriting? Yours is sell into the enterprise.
The decisions that cannot be taken back
| Action | What has to exist afterwards |
|---|---|
| Declining a claim | the rule, the inputs, and an explanation the claimant has a right to |
| Pricing a policy | which factors applied, and evidence they were the approved ones |
| Canceling a renewal | notice, reason, and the version of the rule that produced it |
| Referring for investigation | a human in the loop, recorded, before the referral goes out |
A rule can require a person to approve an action above a size you set, or of a type you name. That person's decision is recorded against the case and audited like the automatic ones.
Who is asking
Two jurisdictions, asking the same thing.
US: state insurance departments and the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 2023 and taken up by more than twenty states.
Europe: the EU AI Act, which makes risk assessment and pricing in life and health insurance high-risk (Annex III 5(c)) and gives the deployer a duty of "clear and meaningful explanations" (Art. 86).
Both ask for the record of one decision.
The machinery is shared with the rest of finance
Thresholds, running totals, scores from an outside vendor with that vendor recorded, approvals and referrals. Financial services shows the rule; business rules shows what state a rule can read.
What is specific to insurance is the duty to explain. That is why the record matters more here than stopping the action does.