Applying the NAIC AI Model Bulletin to Premium Audit and Loss Control AI

What the bulletin covers and where the ambiguity sits

The NAIC adopted the Model Bulletin on the Use of Artificial Intelligence Systems by Insurers on December 4, 2023. It is principles-based guidance. It reminds insurers that decisions or actions impacting consumers, when made or supported by AI and other advanced analytical and computational technologies, must comply with all applicable insurance laws and regulations. The model text is addressed to insurers holding certificates of authority in the state. It does not address brokerages directly.

Two phrases decide most scope questions. The first is "impacting consumers." The second is "made or supported by." The second matters most for operational tools. An AI system does not have to issue the final decision to be relevant. If its output informs a decision or action that affects a policyholder, the tool is within the reach of the wording.

Neither premium audit nor loss control appears in the bulletin text. An NAIC evaluation supplement defines "Direct Consumer Impact" as an AI system that augments or automates human decision-making and whose output has the potential to create final consumer outcomes. Its examples include underwriting, premium determinations, ratemaking, and claims denial and settlement decisions. The list is not exhaustive. Whether an audit-based premium adjustment counts as a premium determination is an interpretive question, so the scope reasoning needs to be written down. This guide is a textual reading, not legal advice.

Use case scope assessment: a starting position

The table gives a starting position for each common workflow. It is not a conclusion. The deciding question is whether the output makes or supports a decision or action that affects a policyholder, and your counsel and state text govern the final answer.

• Payroll and exposure classification: Likely in scope. Classification can drive the premium basis, which links the output to a premium outcome.
• Audit adjustment estimation: Likely in scope. The estimate may support an additional or return premium, close to the premium determinations example.
• Discrepancy flagging: Depends on use. If a flag routes a policyholder to a dispute or adjustment, it supports an action. If it only prioritizes internal review queues, document that position.
• Survey, image and report analysis: Depends on downstream use. Internal analysis that never reaches a policyholder decision may be operational support.
• Risk recommendation generation: In scope if recommendations become policyholder-facing conditions or inform underwriting or renewal. Otherwise record the rationale for treating them as internal support.
• Vendor-supplied scoring or extraction: Scope follows the use, not the supplier. Third-party oversight expectations apply if the use is in scope.

What the bulletin expects for in-scope use cases

Insurers are expected to develop and maintain a written program for the responsible use of AI, referred to as the AIS Program. It should address governance, risk management controls and internal audit functions. It should document the risk identification, mitigation and management framework and internal controls for AI systems generally and at each stage of the AI system life cycle.

The model text also calls for validating, testing and retesting as necessary to assess the generalization of AI system outputs on implementation. That includes the suitability of the data used to develop, train, validate and audit the model. For premium audit, this means testing classification and estimation outputs against the payroll and exposure data they actually see. For loss control, it means testing whether recommendations hold up across the survey data and image quality encountered in the field.

Third-party expectations include audits or other activities to confirm vendor compliance with contractual and, where applicable, regulatory requirements. They also include documented pre-acquisition or pre-use diligence, and monitoring, oversight and auditing of third-party data or AI systems. Contracts should preserve the audit and compliance-confirmation rights these activities depend on. Where a brokerage operates or supplies AI on a carrier's behalf, the carrier's vendor oversight expectations are the practical route by which those obligations reach it.

State adoption and related regimes

The bulletin is a model. Each state department issues its own version, and bulletin dates vary, with Alaska's the earliest on February 1, 2024. A multi-state carrier cannot assume one analysis fits every licensed state. Check whether each state has issued a bulletin as of October 2026, and compare its scope and defined terms to the model before relying on a single enterprise-wide determination.

Other regimes may also apply to the same workflows, and their scope and obligations are defined separately from the NAIC model. Treat each as its own analysis rather than assuming the NAIC scope determination carries over. A practical approach is to build one baseline AIS Program on the model text, then record state-specific differences as addenda per use case.

Runtime controls and where Trussed fits

Written policy tells an examiner what should happen. Runtime records show what did happen. Enforcing policy in the execution path, rather than reviewing after the fact, lets a carrier set boundaries per use case. A payroll classification tool can be allowed to read audit worksheets, while personal data is redacted, out-of-policy tool calls are blocked, and adjustment outputs above a threshold are escalated for human approval. Each enforcement decision then produces a record that can be retrieved by use case and system version.

Trussed is the control, provenance and evidence infrastructure for consequential AI. It enforces policy in the execution path across models, agents, tools and MCP servers: allow, redact, block, rate limit, and escalate for human approval. Enforcement operates on individual interactions and tool calls. It generates continuous audit-ready evidence mapped to NAIC, NYDFS and other frameworks, and deploys self-managed in the customer cloud with configurable retention and data residency. Trussed supports customers in insurance and three other regulated verticals. It does not replace the scope determination, the AIS Program or legal review, which remain with the carrier.