Skip to content

Atlas Mutual · Insurance · 2025

An underwriting copilot that underwriters actually trust

Cut quote turnaround from three days to under four hours by putting a retrieval-backed assistant inside the existing underwriting desk.

Median quote turnaround
3d → 4hMedian quote turnaround
Citations accepted by underwriters
94%Citations accepted by underwriters
Inference cost after routing work
-41%Inference cost after routing work

Services

  • AI engineering
  • Platform & cloud

Stack

  • Next.js
  • TypeScript
  • Postgres
  • AI SDK
  • Vercel

Duration

14 weeks

Start something similar

The problem

Atlas had a working prototype that impressed executives and terrified underwriters. It answered confidently, cited nothing, and was wrong often enough that no one would sign a quote based on it.

What we did

  1. 01Built an evaluation set from 1,200 historical quotes before touching the product surface.
  2. 02Rebuilt retrieval over the policy corpus with span-level citations rendered inline.
  3. 03Added an abstain path — the assistant is allowed to say it does not know, and is scored on doing so.
  4. 04Introduced model routing so routine classifications stopped hitting the frontier model.

Where it landed

The copilot now drafts the first pass on the majority of submissions, with an underwriter reviewing and signing. Atlas’s own platform team has owned the codebase since month five.

They spent the first three weeks building the thing that told us whether we were wrong. Nobody else pitched that, and it turned out to be the whole project.

Dana Whitfield · VP Engineering, Atlas Mutual

Next step

Tell us what you’re building.

A 30-minute call with an engineer, not a salesperson. You will leave with a straight answer about scope, cost and whether we are the right people for it.

Or email hello@nexa.com · Austin, Texas · United States