Status
As of 30 July 2026, this is a completed feasibility study — a proof of concept, not a production system. It shows the approach can rank risk better than the current rules, but the part of the model that estimates loss size is not yet stable enough to deploy, and the study rests on assumptions that need confirming with the business. The next step is a deeper study with the data owners and richer claims data, to validate the value and harden the model for production.
Faster property underwriting decisions
Summary
A model that ranks commercial-property underwriting requests by their likely loss — learned from the insurer's own claims history — puts the riskiest cases in front of reviewers first. In a feasibility study by OXFIDA, the 20% of requests the model ranked riskiest accounted for 33% of all claims: a sharper ranking than Aioi Nissay Dowa Insurance's current hand-maintained rules, achieved using just 14 details about each property.
The goal is to approve low-risk changes faster and free underwriters to concentrate on the cases that genuinely need their judgement.
The problem
When a client adds a building or structure to a commercial-property policy, a sales agent uses a large, complex ruleset to decide whether to approve it or send it to the Underwriting Department. That ruleset is hard to maintain, and it refers many low-risk requests for review anyway. The result is specialist time spent on cases that were never likely to be a problem — slower turnaround for customers, and underwriting capacity tied up on low-risk changes instead of the ones that carry real exposure.
Solution
The model estimates two things for each request: how often a property is likely to suffer a loss, and how large that loss would be. It combines them into a full picture of expected loss, so requests below a risk threshold can be approved automatically while the rest are queued for review, riskiest first.
Two choices make the approach practical for insurance. It is built from generalised linear models — the established, interpretable standard actuaries and regulators already trust — so each decision can be explained rather than taken on trust. And it learns risk at the level of individual buildings even though the insurer records claims only per policy, letting it use the data that already exists rather than data the business would have to start collecting.