Insurance becomes predictive and personal.

Files where cover is unambiguous, the loss can be established from the evidence already submitted, no third party is involved and the fraud checks stay quiet. The threshold belongs to you, set per product line and per claim size, and the model places a file on one side of it rather than inventing the line itself. Anything beyond it goes to a handler together with what the model already gathered, so the review starts from evidence rather than from scratch.

Your claims history only describes risks you let in, so a model trained on it reproduces your current acceptance policy instead of testing it. Retaining declined applications documents that filter, but it does not close the gap: a policy that never ran has no loss outcome to learn from. Closing it takes reject inference, where the likely result of a refused risk is estimated statistically, or a bounded trial in which borderline applications are accepted deliberately and followed until claims emerge.

The ones your business already reports. On the claims side: cycle time, the share of files closed without a manual touch, reopened files and complaints. On the underwriting side: loss ratio and combined ratio per segment, observed over a period long enough for claims to emerge, since a young book always looks profitable. Read both against a control group still running the old route, otherwise seasonality and a shifting portfolio mix get credited to the model.

Which factors carried weight in that decision and how heavily, phrased so the customer can follow it. The file also has to let the decision be rebuilt afterwards: the data that went in, the model version that judged it, and when. That record serves the customer, your own complaints route, an ombudsman and a regulator asking to see the case. Where the consequences are significant, the GDPR also lets the customer ask for human intervention, so build the logging into the design.

It drifts away from the world it was built on. Fraud patterns change once perpetrators work out what you screen for, repair and care costs move with the market, and the mix of your portfolio shifts. Nothing fails visibly, the model simply becomes a little less right over time. The early signs sit at the edges: handlers overriding the score more often, more reopened files, acceptances an underwriter would have called differently. Schedule retraining and keep comparing the model's judgement with realised outcomes.