NAIC AI Model Bulletin Requirements for P&C Insurers

What the bulletin expects

The NAIC adopted the Model Bulletin on the Use of Artificial Intelligence Systems by Insurers 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.

The central expectation is a written AIS Program. Insurers authorized to do business in a state are expected to develop, implement and maintain a program for the responsible use of AI Systems that make, or support decisions related to, regulated insurance practices. The program should be designed to mitigate the risk of Adverse Consumer Outcomes. The bulletin defines an Adverse Consumer Outcome as a decision by an insurer, subject to insurance regulatory standards enforced by the Department, that adversely impacts the consumer in a manner that violates those standards.

Scope follows the insurance life cycle: product development and design, marketing, underwriting, rating and pricing, case management, claim administration and payment, and fraud detection. For a P&C carrier, that reaches most places where AI informs a decision.

The bulletin is a model. It has effect only where a state department of insurance issues its own bulletin, and state adoption varies. California's Bulletin 2022-5, issued June 30, 2022, predates the model. As of October 2026, a carrier should review the text issued by each state where it holds a certificate of authority rather than assume uniformity.

Mapping each expectation to evidence

The bulletin is principles-based. It names categories of documentation and risk considerations rather than prescribing technical controls, so the mapping below is interpretation. It shows what a carrier can reasonably put in front of an examiner for each expectation.

• Governance and accountability: Evidence is mostly documentary: the written AIS Program and a documented owner for it. Confirm detailed wording against the full adopted text and your state's version.
• Risk management and internal controls: Mixed. Policies and risk assessments are documents. Controls such as access rules, data handling restrictions and human approval need records showing they operated.
• Human involvement in final decisions: A named risk consideration. Document where people review or approve AI-supported decisions, and retain records that the step occurred.
• Auditability, traceability and drift: The bulletin refers to these where appropriate. Scope is left to insurer interpretation, so define and record what you will measure.
• Third-party AI systems and data: Documents: due diligence files and contracts. Operationally: records of how third-party AI was used inside your environment.
• Regulator information requests: Inventories, validation, testing and auditing documentation, and retained records of AI system use, produced on request.

Documents versus operational evidence

The bulletin tells insurers what information a Department may request in an investigation or examination. Examples include inventories and descriptions of Predictive Models and AI Systems used to make or support decisions that can result in Adverse Consumer Outcomes, and documentation on validation, testing and auditing. Insurers should expect questions about their AI programs during investigations and market conduct actions, so documentation should be retained.

Most of those categories are documents. A written program, an inventory of underwriting, pricing, claims and fraud models, validation reports and vendor contracts can all be assembled ahead of time. Many carriers can do this well.

The harder cases are expectations that describe behavior. If a program states that higher-risk AI actions require human approval, an examiner can reasonably ask for evidence that the approval happened for specific interactions. If a program restricts what data an AI system may receive, the question becomes what the system was actually permitted to receive and what it did. A policy document cannot answer that. Only records generated at the time of operation can.

This is the practical gap for many P&C carriers: the policy is well described, but there is no durable record connecting the policy to production behavior. The bulletin does not prescribe a human approval step or any specific control, so each carrier decides which controls to adopt. Whichever it chooses, the control needs evidence that it operated.

Third-party AI systems and data

Where data, Predictive Models or AI Systems are collected or developed in whole or in part by third parties, insurers should expect requests for the due diligence performed on those third parties and for the contracts with vendors. Contract terms of interest include representations and warranties, data security and privacy, data sourcing, intellectual property, confidentiality and disclosures, and cooperation with regulators.

The bulletin also lists reliance on third-party data, models and AI Systems as a risk-related consideration. Responsibility does not move with the vendor: insurers remain accountable for compliance with insurance laws, including fairness, accuracy and avoiding unfair discrimination, when they use AI.

In practice, that means two sets of evidence. The first is the contractual and due diligence file, retained so it can be produced in an examination. The second is a record of how the vendor's AI was used in your environment, including which applications, agents and tools reached it, and under what rules. Contracts describe what a vendor agreed to do. They do not show how your own users and systems interacted with the vendor's AI.

Where runtime enforcement and evidence fit

Trussed enforces policy in the execution path of AI interactions and tool calls, across models, agents, MCP servers and coding assistants. For each interaction, policy is enforced at runtime, not reviewed after the fact: the interaction is allowed, redacted, blocked, rate limited, or escalated for human approval. The enforcement decisions generate audit-ready evidence, mapped to NAIC among other frameworks. Deployment is self-managed in the carrier's own cloud, with configurable retention and data residency.

Against the bulletin, this addresses the operational layer. It can produce records of what an AI interaction was permitted to do and what happened, including where a human approval step was triggered. That supports the bulletin's expectations on human involvement, risk controls over data handling, and retained documentation of AI system use.

It has limits that should be stated plainly. Runtime evidence does not replace the written AIS Program, the AI inventory, model validation and testing documentation, or third-party due diligence and contracts. Trussed does not provide model provenance or lineage, and it is not a governance or GRC tool. It is one source of evidence within a broader program that the carrier still owns.