AI Agent Governance for Insurance Underwriting and Claims
Insurance carriers deploying AI agents for underwriting, claims adjudication, and policy administration face regulatory scrutiny from state insurance departments, NAIC model laws, and emerging AI-specific regulations (Colorado SB 21-169, EU AI Act). Every automated decision must be explainable and provable.
What Goes Wrong Without Controls
A mid-market carrier deploys agents to automate small-commercial underwriting, first-notice-of-loss triage, and renewal pricing. Within two quarters:
- An underwriting agent binds a $5M policy without the required second-level approval — the binding authority exceeded the agent's delegation, and no gate prevented it
- A claims agent denies a homeowner claim using data from an unapproved third-party risk model — the state insurance department asks for evidence of the data source and decision logic
- A policyholder disputes a renewal price increase — the carrier cannot prove what factors the pricing agent considered because the execution trace was not signed
Regulatory Mapping
| Regulation | Requirement | Claw EA Control |
|---|---|---|
| NAIC Model Laws | Unfair claims practices, rate justification | Proof bundles capture decision chain; approval gates enforce authority limits |
| Colorado SB 21-169 | AI governance for insurance decisions | Work Policy Contracts declare permitted decision criteria; receipts prove compliance |
| EU AI Act (high-risk) | Transparency, human oversight, logging | Two-person rule, tamper-evident logs, offline-verifiable proof |
| State rate filings | Actuarial justification for pricing | Proof bundles include model inputs and outputs — auditors can verify what data informed the price |
Recommended Control Stack
Approval Gates
Enforce binding authority limits. Underwriting decisions above threshold require human sign-off with signed receipt.
Proof Bundles
Self-contained evidence of the full decision chain. What data was accessed, what model produced the output, who approved.
Budget Controls
Cap exposure per agent, per policy, per day. Prevents runaway automated binding without financial guardrails.
Egress Allowlist
Restrict which data sources and APIs agents can access. Unapproved third-party risk models are blocked.