Atlas Underwriting Group
- Home
- Atlas Underwriting Group
Submission triage from three days to four hours.
Atlas Underwriting Group writes commercial property and casualty risk across four territories. Broker submissions arrived as email attachments — schedules of values in spreadsheets, loss runs as scanned PDFs, cover notes as free text — and an underwriting assistant keyed the relevant fields into the rating system by hand. Median time from submission received to underwriter review was three working days, and roughly a fifth of submissions were never quoted at all because they aged out before anyone reached them.
Two previous vendors had built proof-of-concept extraction models for Atlas. Neither reached production: one could not meet the group's data residency requirements, and the other produced no way for an assistant to see why a field had been filled in the way it had. Atlas came to us asking for the second problem to be solved first.
What we built
A document intelligence pipeline running inside Atlas's own Azure tenancy. Submissions are classified on arrival, split into constituent documents, and passed through targeted extractors for each document type. Every extracted field carries a confidence score and a link back to the exact page and bounding box it came from. Anything below the threshold the underwriting team set is routed to a review queue rather than written to the rating system.
- Document classification on arrival
- Field-level confidence scoring
- Citation back to page and region
- Human review queue for low confidence
- Rating system integration
- Runbooks and a trained internal owner
The result
Median time to underwriter review fell from three working days to four hours. Submissions aging out dropped from 19% to under 3%. Two of the four underwriting assistants moved onto broker relationship work. Atlas's own platform team has since added two new document types without our involvement, which was the point.
