NAIC AI Bulletin and SIU Fraud Detection: Required Controls and Audit-Ready Evidence

Does the bulletin reach SIU fraud detection?

The NAIC Model Bulletin on the Use of AI Systems by Insurers was adopted on December 4, 2023. It reminds insurers that decisions or actions affecting consumers that are made or supported by advanced analytical and computational technologies, including AI Systems, must comply with all applicable insurance laws and regulations. It expects each authorized insurer to develop, implement and maintain a written program, the AIS Program, for the responsible use of AI Systems that make, or support decisions related to, regulated insurance practices.

The program is meant to address AI use across the insurance life cycle, and fraud detection is named explicitly alongside case management and claim administration and payment. The program should address governance, risk management controls and internal audit functions.

The phrase to focus on is "make or support." An SIU model rarely denies a claim on its own. It scores a claim, surfaces indicators or ranks a referral queue, and a person acts on the output. That is support in the ordinary sense. The bulletin does not contain an SIU-specific statement, so scoping SIU tools into the AIS Program is an interpretation. It is the conservative one, and a carrier that excludes SIU tools from its inventory should be able to explain why.

Adverse consumer outcomes and unfair discrimination in SIU decisions

The bulletin defines an adverse consumer outcome as a decision by an insurer that is subject to insurance regulatory standards enforced by the Department and that adversely impacts the consumer in a manner that violates those standards. The AIS Program is designed to mitigate the risk of such outcomes. The definition is tied to violations of existing law, not to AI as a category.

For SIU, this means the question is how an AI-influenced step could contribute to a decision that breaches claims-handling or anti-discrimination standards. Two paths are worth examining. A fraud score may route one population of claimants to heightened scrutiny, with delay or denial following, and the pattern could track a protected or otherwise prohibited characteristic through a proxy variable. Or a generative summary may state an unsupported inference as fact, and an investigator may rely on it. In both cases the AI output is one input among several, which is why the bulletin's emphasis on decisions "supported" by AI matters.

Where a use case sits on the spectrum of consequential decisions depends on what the output can trigger. A score that only informs a reviewer's reading order is less consequential than one that triggers a payment hold or a denial recommendation.

SIU AI use cases by compliance exposure

This ranking is a practitioner assessment, not bulletin text. Exposure rises with the degree to which the output can change what happens to a claimant without a documented human decision in between.

• Referral prioritization and claim scoring: higher: Directly shapes which claimants receive scrutiny. Needs documented thresholds, outcome review and a recorded human decision on referral.
• Third-party data and vendor fraud tools: higher: The carrier relies on models and data it did not build. Needs defined permitted use and records of what was sent and received.
• Generative summarization of claim files: moderate: Can introduce unsupported statements into the file. Needs controls on inputs, redaction of sensitive data and reviewer sign-off.
• Indicator surfacing for investigators: moderate: Informs judgment without setting outcomes. Needs a record of which indicators were shown and how they were used.
• Internal triage and case-management support: lower: Administrative effect. Still belongs in the inventory and the program.

Third-party tools and state variation

Many SIU capabilities come from vendors: scoring engines, data enrichment and increasingly generative tools. Carriers should treat these as inside the AIS Program, because the carrier remains the party answerable to the regulator for claim decisions. In practice that means contract terms and due diligence, plus a technical limit on what the tool may receive and what its output may do inside the carrier's own workflow.

State adoption is uneven. The model becomes operative only where a state adopts it as its own bulletin. A NAIC paper with data as of August 2025 counted 24 adopting states and noted that several southern states, including Alabama and Florida, had not yet enacted it. NAIC's implementation map lists adoptions from Alaska (Bulletin B 24-01, February 1, 2024) through Wisconsin (March 18, 2025), and separately lists California Bulletin 2022-5, issued June 30, 2022. As of October 2026, counts differ by source, so a multistate carrier should confirm status state by state and build to the stricter expectation rather than maintain separate SIU programs.

Demonstrable evidence versus asserted evidence

The bulletin advises insurers that regulators may request information and documentation during an investigation or examination. The adopted text includes an item on information related to the techniques, measurements, thresholds and similar details of the AI Systems in use. Policies, committee minutes and model documentation answer part of that request. They show intent.

What they cannot show is behavior: what a specific tool was allowed to access on a specific claim, what it returned, and who approved the resulting action. Records assembled after the fact from policy documents are assertions. Records generated by the control path itself are evidence. An examiner asking about one referral decision will, in practice, test the difference.

The categories below are practitioner inference about what regulators can reasonably expect, not a list drawn from the bulletin. They are a sensible basis for designing SIU evidence.

Runtime enforcement as one way to produce the evidence

Enforcing policy in the execution path means the control acts when the AI is used, not in a later review. Trussed enforces policy at runtime across models, agents, tools and MCP servers: allow, redact, block, rate limit, or escalate for human approval, at the level of individual interactions and tool calls. In an SIU setting, that could mean redacting sensitive claimant fields before a vendor tool receives them, blocking a tool call outside the approved use case, or holding a referral action until an investigator approves it.

That enforcement generates continuous audit-ready evidence, mapped to NAIC among other frameworks. The record is a product of the control path, so it shows what the AI was permitted to do and what it did at the point of use. Trussed runs self-managed in the customer's cloud, with configurable retention and data residency.

The boundaries are plain. Trussed does not provide model lineage, dataset provenance or end-to-end tracing of agentic workflows. It supports the carrier's AIS Program and does not replace it, and this guide is not legal advice. A fair evaluation question for any runtime tool is whether an examiner could pick one SIU interaction and see the policy decision, the output and the approval from the tool's own records.