Claims, straight through. Seven days to two.
A regional property and casualty insurer handled first notice of loss through email, fax, and a portal from 2011. Adjusters spent their first hour of every case transcribing. The claims SVP wanted touchless claims and a better cycle time. We rebuilt intake, applied document AI with proper evaluation, and moved 80% of first notices to straight-through processing.
-
7 to 2 days
Median cycle from first notice to first payment decision.
-
80%
Of first notices now processed without manual transcription.
-
+22%
Additional claims handled per adjuster per week.
What was hard.
How we worked.
-
Evaluation-first design
We built the evaluation harness before we built the extraction pipeline. Held-out set of 500 historical FNOLs with ground truth. Accuracy thresholds agreed with the claims lead before development started.
-
Document understanding pipeline
Intake across email, fax, and portal into one normalized document store. Extraction with an LLM pipeline, validated against the evaluation harness on every model change.
-
Policy admin integration
Idempotent write path into the policy admin system. Audit trail for every claim touch. Human-in-the-loop review for anything that fell below the accuracy threshold.
-
Adjuster workflow redesign
Adjuster screen redesigned around the automated draft. Three clicks to accept, one click to escalate, a clear reason code for each.
-
Model governance
Model registry, versioning, and quarterly review. Drift monitoring on document types and extracted field accuracy. Runbooks for downgrade and rollback.
What we shipped.
- 80% straight-through first notice rate inside the pilot portfolio.
- Claims cycle time cut from 7 days to 2 days, median.
- Adjuster productivity up 22% with no quality regression.
- Quarterly model review signed by the claims lead, not by engineering.
Who shipped it.
- 12 weeks
- 4 people
- ML engineer embedded post-launch
- Quarterly model review