Use case · Claims document processing

Every claim arrives as documents. Extraction is a small-model job.

A 3–7B model fine-tuned on your claim files reads FNOL emails, loss runs and supporting documents into the structured fields your claims system needs — on your infrastructure, at a fraction of frontier API cost, with low-confidence extractions routed to a human.

Insurance claims document processing runs FNOL intake, loss runs and supporting documents through a 3–7B model fine-tuned on your own claim files, deployed on your infrastructure. The model returns validated JSON mapped to your claims platform, routes low-confidence extractions to a human queue, and keeps claimant PII and medical records off third-party APIs. Published small-model migrations cut inference costs 66–80% (Forethought, on AWS).

Claims intake is a document problem before it's a decisions problem.

FNOL emails, ACORD forms, loss runs, adjuster notes and medical reports arrive in formats your claims system can't read. Manual keying is slow; a frontier API means paying open-ended reasoning prices for a closed, repetitive task — and sending claimants' personal and medical data to a third party. Extraction from known document types is precisely the narrow work where a fine-tuned small model wins.

How it works · The Crit Path
01Scope — one document flowWe start with a single high-volume flow — FNOL intake or loss-run ingestion — and build a frozen eval set from your historical claims: documents in, correct fields out, verified by your team.
02Prove — fine-tune for your documentsWe train a 3–7B model on your document types and field schema, producing validated JSON mapped to your claims platform. A lightweight classifier routes claim type and severity so work lands in the right queue.
03Deploy — inside your perimeterThe models run on-prem or in your private cloud. Claimant PII and medical records never transit a third-party API. Confidence thresholds send uncertain extractions to a human review queue instead of into your system of record.
04Assure — measure, monitor, retrainWe track field-level accuracy in production against the eval set. When your document mix shifts — new form versions, new lines of business — we retrain under the Assurance retainer. Format change is an operations event, not an outage.
The economics

A worked example, our assumptions stated. A carrier or MGA processing 10,000 claims a month, at 15–25K tokens of documents per claim, runs 150–250M tokens a month. The Crit recomputes this on your actual claim mix before you commit to anything.

Frontier APIFine-tuned model, yours
Cost shapePer token, linear with claim volumeFlat GPU infrastructure — marginal cost per claim is a rounding error
10,000 claims a monthA five-figure monthly API bill at public list prices, August 2026$2,000–5,000/mo GPU infrastructure — our cost model
Claimant PII and medical recordsTransit a third-party APIStay on your hardware
Accuracy on your document typesGeneral model guessing at your formsTrained on your back catalogue; measured field by field before go-live

Published benchmarks, not our claims. Forethought reports 66–80% lower inference costs from moving routine workloads to fine-tuned small models (published on AWS). distil labs reports 50–68% cost reduction with accuracy rising from 81% to 93% at Knowunity, and ~50% savings with Cerebrium (vendor-reported). We haven't shipped this exact vertical yet; the published evidence says the economics hold, and we prove it on your claims before you commit.

Where this doesn't work

If your claim volume doesn't clear the threshold where ownership beats the API — or your documents are genuinely novel every time, with no back catalogue to train on — we'll say so in the audit, and you keep the numbers. And extraction is not adjudication: coverage decisions stay with your adjusters, by design.

What you get

The pipeline, measured. Yours.

  • A fine-tuned extraction model for your document types, returning validated JSON mapped to your claims platform's fields
  • A claim-type and severity classifier for queue routing
  • A frozen eval set from your historical claims; field-level accuracy measured before go-live
  • A human-review queue for low-confidence extractions, with thresholds your ops team controls
  • Deployment on your hardware or private cloud — no claimant data in third-party APIs
  • Production monitoring and retraining under an Assurance retainer
Who this is for
MGAs and regional carriers keying by handClaims intake still means manual entry from PDFs and email attachments.
Ops leaders who piloted the APIYou tried a frontier model on extraction and found the per-document cost untenable at real volume.
Insurers whose data can't go outCompliance or security has ruled out sending claimant medical and personal data to external AI providers.
Questions, answered straight

Can AI extract data from insurance claims documents accurately?

On known document types with a defined field schema — yes, and this is where fine-tuned small models specifically beat general-purpose ones: the task is narrow and the training data is your own back catalogue of claims. We don't ask you to take that on faith; we measure accuracy field by field against a frozen eval set of your historical claims before the system touches production, and low-confidence extractions go to a human queue.

Do claimant medical records leave our infrastructure?

No. The extraction and classification models run on your hardware or in your private cloud tenancy. There is no third-party API in the path, which keeps your data-protection assessment short and your subprocessor list unchanged.

What happens when document formats change?

The system tells you before it fails silently: we monitor production accuracy against the eval set, and drift — a new ACORD version, a new loss-run layout, a new line of business — triggers retraining under the Assurance retainer. Format change is an operations event, not an outage.

Security · by architecture

Special-category data stays home.

  • Medical records never leave
  • No external AI vendors in claims
  • Human review lanes
  • Full audit trail
Built for review under
UK/EU GDPR FCA ICOBS NYDFS 500 SOC 2

We’re engineers, not your compliance advisers — the architecture keeps your existing compliance intact instead of adding a vendor to it.

Small models · critical hits

Your claim files already wrote the spec.

Two weeks inside one document flow — FNOL or loss runs — and you'll know field-level accuracy, projected cost per claim, and whether ownership beats the API at your volume. The numbers are yours either way.