Claims are documents. Documents are what small models do best.
Extraction, triage, and classification for claims files, policy documents, and broker submissions — running on infrastructure you control, because claims data is some of the most sensitive data there is.
Insurance AI starts with documents: FNOL forms, policy schedules, broker submissions, loss runs. A 3–7B model fine-tuned on your document types extracts and triages them on infrastructure you control, so claims data — health details, financial records, personal history — never transits a third-party API. Published migrations report 66–80% lower inference costs on routine workloads (Forethought, on AWS).
Insurance is a document business with a data-sensitivity problem. Claims files mix health details, financial records, and personal history — precisely the material that shouldn't transit a third-party API as a matter of course. Yet the daily work done on those files — extracting fields from FNOL forms, triaging claims by complexity, matching submissions against appetite, normalizing loss runs — is narrow, repetitive, and high-volume. That's the exact profile where a fine-tuned small model beats a rented frontier model: on cost, on privacy, and usually on accuracy for the specific task.
The architecture is the same pattern we deploy everywhere: a 3–7B model trained on your document types, running in your VPC or on-prem, gated behind a frozen eval set measured per form type. Claims handlers and underwriters get structured output in their existing systems; ambiguous cases escalate to a human or, stripped of identifiers, to a frontier model — under rules you set, not a vendor's.
We haven't shipped inside a carrier or MGA yet. The published benchmarks below are why we think the economics hold; The Crit is how we'd prove it on a sample of your own submissions before you commit.
The claims file, structured.
Claims data routinely includes special-category personal data under UK and EU GDPR, and insurance is a regulated sector on both sides of the Atlantic — FCA conduct rules in the UK, state-level regulation in the US. The practical consequence for AI is vendor scrutiny: every third party in the claims data flow is something to assess, contract for, and answer for. A model deployed inside your perimeter removes that third party from the flow entirely. We're engineers, not your compliance advisers — we just make the data-flow diagram short.
We haven't shipped this vertical yet. Here is the published evidence the economics hold — and here is how we'd prove it on a sample of your own submissions, against a frozen eval set, before you commit.
| Published result | Source |
|---|---|
| Inference costs fell 66–80% after moving routine workloads to fine-tuned small models | Forethought, published on AWS |
| Task accuracy rose from 81% to 93% while inference costs fell 50–68% | distil labs × Knowunity, vendor-reported |
| Serving costs fell roughly 50% on dedicated small-model infrastructure | distil labs on Cerebrium, vendor-reported |
| Task-specific small models to see 3× the adoption of general-purpose LLMs by 2027 | Gartner forecast |
Can a small model handle messy claims documents?
Yes, for the routine bulk — because we train it on your document types and measure it per form type against a frozen eval set before anything goes live. FNOL forms, loss runs, and broker submissions are narrow, repetitive extraction work: exactly the profile where a fine-tuned small model beats a general model on the specific task.
What happens to ambiguous claims?
They escalate — to a human, or, stripped of identifiers, to a frontier model — under rules you set, not a vendor's. The local model handles the routine bulk; the escalation policy is part of the deliverable, and claims handlers see structured output in the systems they already use.
Do we need to run AI on claims already?
No. If you don't, we build the extraction pipeline from scratch and prove it on a sample of your own submissions first. If you do, most of that spend is routine work — published migrations report 66–80% lower inference costs after moving routine workloads to fine-tuned small models (Forethought, on AWS).
Special-category data stays home.
- Medical records never leave
- No external AI vendors in claims
- Human review lanes
- Full audit trail
We’re engineers, not your compliance advisers — the architecture keeps your existing compliance intact instead of adding a vendor to it.
See if the numbers hold on your claims volume.
The Crit is a teardown of your AI spend and workflows: where a small model wins, where an API is fine, and where AI shouldn't be used at all. You get the numbers either way.
If your workload doesn't clear roughly 50M tokens a month and you have no privacy constraint, a frontier API is probably fine — and we'll tell you so.