Submission review cut from four hours to twenty-five minutes

The Challenge

Every submission arrived as an email with somewhere between four and forty attachments: broker schedules in inconsistent spreadsheet layouts, scanned loss runs, endorsement wording, occasionally a photograph of a building. An underwriter spent most of a morning turning that into the fifteen fields needed to make a decision, and the work was identical every time.

The cost was not only time. Because review was expensive, submissions were triaged by broker relationship rather than by quality, and marginal risks got the same attention as good ones. The team knew which submissions they wanted to see first; they just had no way to know before reading everything.

The Solution

We built an intake system that reads the whole submission and produces a structured record with a citation for every extracted value — page, document, and the text it came from. Nothing is inferred silently: where the documents conflict or a required field is missing, the system says so instead of producing a confident guess.

On top of extraction sit the underwriter’s own appetite and completeness rules, so the summary an underwriter opens leads with the exceptions. A clean submission is confirmed in a couple of minutes. A problematic one arrives with the problem already identified.

The record writes into the existing policy administration system through its API, which meant no change to downstream reporting and no parallel data store to reconcile. Extraction quality is measured continuously against a held-back evaluation set, and reviewed with the underwriting team monthly.