Skip to content

Insurance · 2024 · Client anonymised at their request

Cutting claims triage from six days to under one

Document-driven claims triage with a human decision point retained by design.

A regional insurance group

System diagram: scanned claims of variable quality pass through extraction with per-field confidence scoring, where high-confidence records are automatically classified and everything else escalates to a reviewer console showing the source document side by side. Both branches write to the same immutable decision record, and an evaluation harness gates every model change.

6d → 19h

Median triage time

71%

Claims auto-classified

100%

Decisions with an audit record

01

The constraint

Claims arrived as scanned documents of wildly varying quality, and the group could not accept a system that made final decisions without a reviewable human step. Explainability was a regulatory requirement, not a preference.

02

The architecture decision

Extraction ran through a document pipeline with confidence scoring per field, and only high-confidence extractions were routed to automated classification. Everything else escalated to a reviewer with the source document and extracted fields side by side. Every model version was pinned and every decision recorded with the inputs that produced it, so any outcome could be reconstructed months later.

03

What shipped

  • 01

    Document extraction pipeline with per-field confidence scoring

  • 02

    Evaluation harness gating every model change

  • 03

    Reviewer console with side-by-side source comparison

  • 04

    Immutable decision audit log

Adjusters stopped spending their mornings on data entry. The evaluation harness caught two regressions before they reached production, which is the part the group values most.

Figures verified with the client’s own instrumentation, not ours.

04

What it moved

Practices involved

  • Data & AI systems
  • Product engineering
  • Design systems