Composite study

This is a product-class write-up, not a named-client story and not a results press release. Counts below describe the pack we actually write (approach points, tools, layers) — not invented bug totals or conversion lifts.

4
Approach points
4
Tools in the pack
4
Layers covered
3
Challenges documented

Overview

Users could open a claim in chat. A second session created a duplicate FNOL for the same event. Status text said “approved” while the core system said “review.”

Photo uploads from claim A appeared in a copilot answer for claim B after a retrieved-chunk mixup.

Industry pattern this is grounded in

Public controversy around automated claims and prior-authorization systems is a reminder: status and denial language must match the system of record, not a model’s summary of similar cases.

Highlights

Typical launch bar: Block until pass rules hold on the money, identity, or safety path.

Challenges

01

The claims core lagged; the copilot cached a previous status and sounded current.

02

Image captions generated by a vision model described a license plate that was not in the photo.

03

Partners wanted a public tracker link that did not require login — that link enumerated claim ids.

How we tested

01

Deduplicate FNOL by policy + event window before the model talks.

02

Status is a tool read. The model may explain codes; it may not invent them.

03

Document tenancy on search and on RAG.

04

Adjuster vs policyholder views: notes and reserve amounts stay internal.

Starting point vs what we required

Starting point versus the controls we required
Starting pointWhat we required
The claims core lagged; the copilot cached a previous status and sounded current.Deduplicate FNOL by policy + event window before the model talks.
Image captions generated by a vision model described a license plate that was not in the photo.Status is a tool read. The model may explain codes; it may not invent them.
Partners wanted a public tracker link that did not require login — that link enumerated claim ids.Document tenancy on search and on RAG.

What we delivered

01

Journeys / UI

  • Playwright — FNOL, duplicate, document upload
02

API / contract

  • Postman — Claim IDOR and public tracker enumeration
03

LLM traces / evals

  • LangSmith — Retrieval mixing across claims
04

RAG quality

  • Ragas — Status explanation grounded in the claim record

Outcomes

Enumeration on the public tracker was Critical. Status copy was bound to the API enum. Vision captions had to be marked generated and not stored as facts.

Duplicate FNOL now returns the existing claim id. We automated that with Playwright plus API tests.

Stack

Testing stack used on this product class
LayerToolWhat we validated
Journeys / UIPlaywrightFNOL, duplicate, document upload
API / contractPostmanClaim IDOR and public tracker enumeration
LLM traces / evalsLangSmithRetrieval mixing across claims
RAG qualityRagasStatus explanation grounded in the claim record

Same method as the rest of our work: severity-ranked findings, pass rules in writing, and a ship / wait / block call. See the evaluation method and tool comparison.