Underwriting Intelligence Model — Property and Casualty Insurer, Toronto, Canada
The Situation
A mid-size Canadian property and casualty insurer had been using a generic LLM to assist underwriters with risk summaries for about 8 months. The model produced summaries that looked structured and confident. The underwriters ignored most of them.
The reasons were specific. The model had no knowledge of Canadian provincial regulatory requirements, which vary enough between Ontario, Alberta, and British Columbia to matter on commercial property risks. It did not know the firm's internal risk appetite guidelines, which had been refined over years and lived in a combination of underwriting manuals and institutional memory. It had never seen the firm's regional loss history, which shaped how senior underwriters read certain exposure types.
Override rates were high. On complex commercial risks, underwriters were overriding the model's risk scoring on roughly 7 out of every 10 submissions. The tool was adding a review step without removing any work.
What Amorisoft Did
Amorisoft built the training dataset from 8 years of the firm's own policy, claims, and underwriter decision records. This covered 34,000 submissions across personal lines, SME commercial, and large commercial segments. Underwriter override decisions were included as training signal, not filtered out, because the pattern of overrides was itself information about where the generic model had been wrong and why.
Provincial regulatory requirements for Ontario, Alberta, and British Columbia were structured into the model as hard constraints on risk scoring outputs. Regional loss history was encoded at the postal code segment level for the firm's primary exposure geographies.
Amorisoft ran a 6-week evaluation process where the model's outputs on 2,400 held-out historical submissions were compared against actual underwriter decisions. Scoring alignment came in at 83% on the held-out set before any production deployment.
The firm's actuarial team was involved from week 3 onward. Their input shaped how regional loss history was weighted in the model relative to submission-level risk factors, which turned out to be the decision with the most impact on final output quality.
Results
Underwriting decision time dropped from 3 days to 4 hours on standard commercial submissions within 10 weeks of go-live. The model's risk summaries were approved without revision on 83% of submissions, against an override rate that had previously run above 70%. 22% of underwriter capacity freed up on routine submissions was redirected to large commercial risks, where the model was not deployed. The legacy data reconciliation added 2 weeks to delivery but the final training dataset covered the full 8-year window as originally scoped.