Meridian Financial

Underwriting Copilot

An evaluated AI assistant that reads submission packs and drafts the first underwriting opinion.

Client: Meridian Financial
Industry: Financial services
Discipline: AI
Duration: 5 months, 2026

The problem

Meridian'sunderwriterswerespendingthebetterpartofadaypercommercialsubmissionreadingunstructuredpacks:brokeremails,lossruns,valuations,andsitereportsarrivinginelevendifferentformats.

The bottleneck was not judgement, it was extraction. Experienced underwriters were doing clerical work, and turnaround times were costing the firm business to faster competitors.

What we did

We built the evaluation harness before the product. Three hundred historical submissions with known outcomes became the scoring set, which meant every subsequent change could be measured rather than argued about.

The copilot ingests a submission pack, extracts the risk factors into a structured schema, retrieves comparable historical cases, and drafts an opinion with a citation against every claim. Where confidence falls below threshold it says so and routes the case to a senior underwriter untouched.

Underwriters review and correct in the same interface, and every correction is captured as evaluation signal. Nothing is auto-issued: the model drafts, a human decides.

Services applied

Technology

Claude
Anthropic API
Python
TypeScript
pgvector
PostgreSQL
AWS Bedrock
Next.js

Gallery

Inside the build.

Results

What changed, measured after launch.

Average time per submission: 6.5h → 48mAverage time per submission
Extraction accuracy on hold-out set: 0%Extraction accuracy on hold-out set
Auto-issued decisions; humans decide: 0Auto-issued decisions; humans decide
Submissions handled per underwriter: 3.1×Submissions handled per underwriter