Skip to content
Work / Northwind

Cutting claims triage from nine minutes to forty seconds

A retrieval and classification system that reads the first notice of loss, routes it, and explains why, with an evaluation harness that made the accuracy claim defensible.
Client
Northwind
Sector
Insurance
Year
2026
Services
AI Systems, Automation
All case studies

40s

median triage time, from 9 minutes

96.2%

routing accuracy on held-out claims

3.1x

claims handled per adjuster per day

The problem

Challenge

Every first notice of loss arrived as free text, a PDF, or a voicemail transcript, and a human read all of it to decide which of eleven queues it belonged in. Volume had doubled in two years. The team had tried a keyword classifier that nobody trusted, because when it was wrong there was no way to see why.

What we did

Approach

In order, including the steps that changed the plan.

  1. 01

    Build the eval set before the model

    We started by hand-labelling 2,400 historical claims with the queue they should have gone to and the queue they actually went to. That set was the contract for the rest of the project, and it immediately showed that 7% of human routings had been wrong.

  2. 02

    Retrieve the policy, then decide

    Routing depends on coverage, so the system pulls the relevant policy sections before it classifies. Hybrid search over the policy corpus with a reranker lifted the retrieval hit rate from 71% to 94%, and the classification accuracy followed it.

  3. 03

    Make every decision auditable

    Each routing decision stores the extracted facts, the policy clauses retrieved, and a one-sentence rationale. Adjusters can disagree in one click, and disagreements flow straight back into the eval set.

  4. 04

    Escalate rather than guess

    Below a calibrated confidence threshold the claim goes to a human queue with the model's working shown. That path handles 8% of volume and is the reason the other 92% is trusted.

Where it landed

Outcome

Triage now runs in under a minute for nine claims in ten. The measured accuracy is higher than the human baseline it replaced, and because the eval set is maintained, the number is re-verified on every model change rather than quoted from launch week.

Built with

  • Claude
  • pgvector
  • Temporal
  • Postgres
  • Next.js
The evaluation harness was the part we did not know we needed. It turned an argument about whether the model was good enough into a number we check every week.
Dana WhitfieldVP Claims Operations, Northwind
Next step

Tell us what you are building.

A system you want built, a model that has to survive real traffic, or a process that should have been automated a year ago. The first conversation costs nothing, and occasionally ends with us telling you not to build it.

Reply within one working day

A person, not an autoresponder.

A 30-minute call, no deck

We ask about constraints, not budget.

A written view within a week

Including the case for not proceeding.