Home / Case studies / Underwriting engine

An underwriting engine a regulator could audit

$8B+ alternative asset manager · HIPAA- and SEC-regulated environment

99.9%

Decision accuracy in production, benchmarked against labeled human review

−33%

Underwriting cycle time, with no reduction in review standards

0

Regulatory findings attributable to the system

  1. 01Records inscans · faxes · forms
  2. 02Extractstructured data
  3. 03Validatevs labeled set
  4. 04Human approveexceptions routed
  5. 05Bound decisionfull audit trail

Every decision traceable to source document, model version, and approver

The situation

Underwriting decisions depended on human review of long, unstructured medical and application files. Cycle time was a function of reading speed; throughput was capped by qualified reviewers; and HIPAA plus SEC scrutiny ruled out most of what the vendor market was selling. Prior automation attempts had stalled at the accuracy bar — nobody would put a model in front of a bound decision without evidence it was at least as good as a person.

What we built

  • 01Ingestion and extraction engineered for documents as they actually arrive — scans, faxes, inconsistent forms
  • 02An LLM decision engine benchmarked against a jointly-labeled validation set, not a vendor demo set
  • 03Human-in-the-loop approval with exception routing to wherever the model is least confident
  • 04Full lineage — every output traceable to source document, model version, and approver
  • 05A model-risk package written for examiners and auditors before launch, not after the first finding
Why it survived scrutiny

The accuracy number existed before the launch decision did. Governance was designed as a chapter of the build — which is why the model-risk conversation was a review, not a negotiation.

This problem in your world: Life, Annuity & PRT

Bring the hardest version of the problem.

Thirty minutes, no deck and no pitch. Leave with a sharper picture of the problem than you arrived with — and a straight answer on whether we can help.

If there’s a fit, we’ll say so. If there isn’t, we’ll say that too — and usually who to call instead.