NAIC AI Model Bulletin Chief Risk Officer Requirements: From Written Program to Runtime Evidence

What the Bulletin expects of an AIS program

The NAIC adopted the Model Bulletin on the use of artificial intelligence systems by insurers in December 2023. It aligns with the NAIC Principles of Artificial Intelligence from 2020. It is a model for states to adopt, not a law in itself, so the operative obligations come from each adopting state's bulletin.

The core expectation is a written program, the AIS Program, for the responsible use of AI Systems that make or support decisions related to regulated insurance practices. Controls should be commensurate with the insurer's own assessment of the degree and nature of risk to consumers, including the nature of the decisions the system supports.

The program should address governance, risk management controls and internal audit functions. It should document the insurer's risk identification, mitigation and management framework, and its internal controls, for AI Systems generally and at each stage of the life cycle from proposed development to retirement. The model text also calls for monitoring, auditing, escalation and reporting protocols.

The language is expectation language, such as "should" and "where appropriate". The Bulletin does not prescribe a technical mechanism. Control strength is left to the insurer's risk assessment, which means the rationale for that strength becomes something an examiner may ask about.

Where accountability falls to the CRO

The Bulletin sets expectations for the insurer as a whole. How accountability is divided among risk, compliance, IT and business owners is the insurer's decision, and it should be written into the AIS Program rather than left implicit.

A workable split is for risk to own the framework, the risk assessment behind control strength by use case, and effective challenge. Compliance maps each state's requirements and coordinates examination responses. IT owns the enforcement mechanisms, retention and access to records. Business owners own the use case, its approval and its day-to-day operation. This is one reasonable arrangement, not a Bulletin requirement.

Two areas fall naturally to the CRO. The first is the oversight and approval process for developing, adopting or acquiring AI Systems. The second is the identification of constraints and controls on automation, so that an AI system's function is balanced against its risk. The second is the closest text in the Bulletin to enforced limits on what a system may do, and the AIS Program should document where those limits are enforced, their scope and any exceptions.

Multi-state carriers: adoption, New York and Colorado

As of October 2026, the NAIC adoption map lists 20 jurisdictions, 19 states plus the District of Columbia, with adopted bulletins. Listed adoption dates run from February 1, 2024 (Alaska) to March 18, 2025 (Wisconsin), and none of the listed dates falls in the past 12 months. California Bulletin 2022-5, issued June 30, 2022, is listed separately. Because each state's text governs, a multi-state carrier is best served by one baseline program with recorded differences by jurisdiction.

NYDFS Insurance Circular Letter No. 7 (July 11, 2024) is more specific about management roles. It states that senior management is responsible for day-to-day implementation of the insurer's ECDIS and AIS development and management, consistent with the board's strategic vision and risk appetite. That includes establishing policies and procedures, assigning competent staff, overseeing model risk management, ensuring effective challenge and independent risk assessment, reviewing internal audit findings and taking prompt remedial action. Where New York applies, the CRO's role in each of these should be defined. They are governance activities that runtime records would inform but not perform.

Colorado has changed. SB 24-205 should no longer be treated as current law. The Colorado Attorney General states that SB 26-189, signed in May 2026, repeals and reenacts the earlier provisions with new requirements on automated decision-making technology in consequential decisions. Any compliance mapping that cites SB 24-205 should be re-baselined against the new law and any rulemaking under it.

Runtime enforcement and evidence in practice

A policy written in advance states intent. A policy enforced at runtime changes what happens to a given interaction, and the record of that decision is what answers an examiner's question. This is an evidence question as much as a control question. A decision record that captures the interaction or tool call, the policy version applied, the outcome and any human approver can be mapped to a specific AIS Program control.

Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with outcomes of allow, redact, block, rate limit or escalate for human approval. Enforcement is at runtime, not reviewed after the fact. The scope is individual interactions and tool calls. The audit-ready evidence generated by that enforcement is mapped to NAIC and NYDFS requirements, among other regimes. Deployment is self-managed in the customer cloud, with configurable retention and data residency. Trussed has run continuously in a regulated enterprise for more than 18 months, with 99.95% uptime on live production traffic, and serves customers in four regulated verticals including insurance.

The limits matter as much as the capability. This evidence does not cover model inventories, validation, drift analysis, data lineage or model provenance, which stay with the carrier's separate governance processes. For third-party AI, a control applies to calls routed through the enforcement point. Vendor due diligence and contractual oversight remain separate work.