Private LLM for insurance: claims, underwriting and the NAIC bulletin
A model trained on your claims and policy data and run in your environment, and what the NAIC bulletin and the EU AI Act ask of insurers.
What regulators expect
| Rule | Scope | What it asks (paraphrased) | Source |
|---|---|---|---|
| NAIC Model Bulletin: Use of AI Systems by Insurers (adopted 4 Dec 2023) | Insurers in states that adopt it; it is guidance, not a model law | A written AI program with governance, risk management and internal controls; oversight of third-party AI vendors; decisions must still meet existing laws on unfair discrimination and unfair trade practices | NAIC |
| State adoption | 29 jurisdictions had adopted it as of 3 November 2026, per NAIC's map; California, Colorado, New York and Texas have their own insurance AI rules | Check the version your state adopted | NAIC adoption map |
| EU AI Act, Annex III 5(c) | Risk assessment and pricing for natural persons in life and health insurance | Treated as high-risk: risk management, data governance, logging, human oversight | Regulation (EU) 2024/1689 |
Note the EU scope: P&C pricing and claims handling are not on the Annex III list; life and health pricing is.
Use cases
| Use case | Input | Output | Approach | Human review |
|---|---|---|---|---|
| First notice of loss triage | FNOL forms, call notes | Line, severity, routing | Fine-tune | Sample QA |
| Claims file summary | Adjuster notes, documents | Summary in your template | Fine-tune | Adjuster reviews |
| Submission intake | Broker emails, schedules, SOVs | Structured fields | Fine-tune for extraction | Underwriter confirms |
| Policy wording Q&A | Wordings, endorsements | Answer with clause reference | Retrieval | Spot checks |
| Fraud referral notes | Claim data, notes | Draft referral rationale | Fine-tune | SIU decides |
We don't build models that set prices or deny claims by themselves. The NAIC bulletin expects decisions affecting consumers to be governed and explainable; we keep a person making them.
See how this would work on your data.
Book a callWorked example: submission intake
Commercial underwriters re-key broker submissions into rating systems. We take past submissions and the fields your team captured, fine-tune a model in your environment to extract the same fields, and score it on recent submissions it hasn't seen, field by field, against a pass mark agreed in week 1. Underwriters confirm the fields before they go into rating. Each model version, its training data and test results are recorded — the kind of documentation the NAIC bulletin's vendor and model controls ask for.
What we'd do first
- 1One workflow with plenty of history and a clear right answer — FNOL triage or submission intake — in an 8-week pilot, deployed in your environment.
- 2If you write life or health business in the EU, we scope Annex III duties with your legal and risk team before any pricing-related work.
Frequently asked questions
Is the NAIC bulletin binding?
It is a model for state regulators. It applies where a state adopts it (NAIC); 29 jurisdictions had adopted it as of 3 November 2026 (NAIC adoption map).
Would IRIS8 count as a third-party AI vendor under the bulletin?
Yes, and you would include us in your vendor oversight. Running in your environment and handing you the model and records makes that easier.
Is insurance AI high-risk under the EU AI Act?
Risk assessment and pricing in life and health insurance is (Annex III 5(c)). Other lines aren't on that list (AI Act).
Does claims data leave our network?
No.
Who owns the model?
You do.