The situation
Kestrel Financial underwrote commercial policies from broker submissions arriving as email with attachments — loss runs, financials, applications, occasionally a photograph of a form. Median time from submission to first underwriter look was 3.5 days, and 19% of submissions were declined for reasons visible on page one.
Separately, their eight-person commercial team routed inbound web enquiries by hand. Median speed-to-lead: 3 hours, worse on Fridays and non-existent at weekends. Brokers who didn’t hear back in a day went elsewhere.
Two problems, one root cause: the first pass on any inbound item was a human doing mechanical work with no queue discipline.
What we found in the teardown
We spent a day with their two most senior underwriters writing down how they actually decide: the factors they look for, the order, the disqualifiers, the weights, the point at which they stop reading. It took four hours and produced a two-page rubric neither had ever seen written down.
That rubric became the specification and the eval. We replayed 400 historical submissions through it manually before building anything, and measured agreement with the recorded decision. The rubric alone agreed 86% of the time — which told us the constraint was reading speed and consistency, not judgement.
What we built
Six weeks, in their VPC with their own model endpoints, in two parts.
Intake and memo. Submissions are parsed, classified and checked for completeness; missing documents trigger an automatic broker request. For complete submissions the system produces a structured memo: factors found, source document and line for each, rubric version applied, a confidence figure, and an explicit list of what it could not verify. Low confidence routes to a senior underwriter by design, not by exception. No policy is bound without a human approving the memo.
Inbound routing. Enquiries are enriched, scored against the same rubric, matched to territory and assigned in minutes, with a drafted first response the rep edits rather than writes.
The compounding part is the override loop. Every underwriter override is captured with a reason code, aggregated weekly, and turned into either a rubric change or a new eval case.
Result
Over six months:
- Underwriting cycle time
3.5 days → 11 hours - Speed-to-lead
3h → 4min - Qualified meetings booked
+34%for the eight-person team in one quarter - First-pass agreement with senior reviewers
86% → 94%after two quarters of override feedback - Page-one declines now surfaced in
under 20 minutes
The artifact their regulator liked most wasn’t the model. It was a documented, versioned rubric and a queryable record of every decision and every override.
“Everything ran in our VPC with our keys. Security cleared it in nine days, which has never happened here before.”
— Marcus Bell, Chief Information Security Officer, Kestrel Financial