NAIC AI Model Bulletin for Insurance CIOs: Runtime Controls and Evidence as of October 2026

What the bulletin is and how it reaches your organization

The NAIC adopted the Model Bulletin on the use of artificial intelligence systems by insurers on December 4, 2023. It states 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 also sets out how a department expects insurers to govern the development or acquisition and use of these technologies.

A model bulletin has no force on its own. It takes effect in a state only when that state issues its own bulletin, and the NAIC adoption map does not determine whether a state's action contains all model elements. One journal article counted 24 adopting states as of August 2025, with Alaska first on February 1, 2024. A multi-state carrier or brokerage therefore needs to identify which state bulletins apply to each legal entity and read each state's text.

Two related regimes sit outside the bulletin. New York Insurance Circular Letter No. 7, issued July 11, 2024, addresses AI systems in underwriting and pricing. It says an insurer should not use external consumer data and information sources or AI systems for underwriting or pricing unless it can establish that the data source or model does not use and is not based in any way on a class protected under Insurance Law Article 26. It also borrows the NAIC clause on incorporating certain terms into third-party vendor contracts where appropriate.

Colorado has changed. SB 26-189, signed in May 2026, repeals and reenacts provisions on automated decision-making technology in consequential decisions with new requirements. Earlier SB 24-205 duties should not be treated as current. Carriers operating in Colorado should track SB 26-189 and Colorado Attorney General rulemaking.

Runtime enforcement versus review and documentation

A useful way to scope the work is to sort every expectation into one of three groups. This grouping is an analytic framework, not bulletin language.

The first group is enforced in the execution path. Access control, data handling, policy applied at inference or at a tool call, and human approval for sensitive actions belong here. These controls decide what a system may do before it does it, and they generate a record as a by-product. A written policy that says non-public information must not reach an external model is a statement of intent. A control that redacts or blocks it at the moment of the request is enforcement.

The second group is performed after the fact. The bulletin describes validation, testing and retesting as comparing performance on data available at development with post-implementation performance, or against expert review. Drift review and audits are in the same group. They matter, but they are periodic and cannot prevent a specific interaction from breaching policy.

The third group is documentation: the written AI Systems program, inventories, vendor due diligence files and contract terms. Runtime controls do not produce these. The examination gap many carriers face is not missing policy. It is that the first group does not exist or does not leave records, so nobody can show what an AI system was permitted to do and what it did.

Third-party AI and examination readiness

The bulletin advises insurers of the information and documentation a department may request in an investigation or examination about AI use. NAIC Third-Party Data and Models Task Force materials list due diligence on third parties and their data, models or AI Systems, and contracts with third-party vendors, including terms, among such items.

For a CIO, this separates two questions that are often merged. Contract terms and due diligence files are governance artifacts, usually owned by procurement, legal or risk. Technical visibility into how a vendor tool is used inside your environment is an engineering question. A contract clause does not show what an employee sent to a vendor tool last Tuesday. Where vendor AI is reached through your own network, browsers, coding assistants or tool connections, controls in that path can enforce policy and retain records. Where the vendor runs the model entirely inside its own service, your visibility ends at the boundary, and that limit should be recorded rather than assumed away.

Assigning ownership and sequencing the work

Start with the inventory, since the bulletin refers to inventories and descriptions of models and AI Systems, and every other control depends on knowing what exists. Include vendor-supplied systems and internal ones.

Next, record for each expectation whether the evidence is generated automatically in production or compiled manually later, and name the owner. IT typically owns enforcement and retention, compliance owns state applicability and examination response, and risk owns validation and vendor oversight. Review vendor contracts against the contract-terms item in the task force materials and the New York clause. Finally, have counsel confirm each state's bulletin text, and treat New York's underwriting expectations and Colorado's SB 26-189 as separate obligations that the NAIC bulletin does not resolve.

Where Trussed fits, and where it does not

Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. For individual interactions and tool calls it can allow, redact, block, rate limit or escalate for human approval. Enforcement is at runtime, not reviewed after the fact, and the same enforcement generates continuous audit-ready evidence mapped to NAIC, NYDFS and other frameworks. It can be deployed self-managed in the customer's cloud, with configurable retention and data residency. Trussed has customers in insurance among four regulated verticals and has run continuously in production in a regulated enterprise for more than 18 months.

This addresses the access, data handling, policy enforcement and retained-evidence areas above. It does not replace the written AI Systems program, the model inventory, validation and retesting, vendor due diligence or contract terms. Those remain with your governance, risk and procurement processes, and the evidence from enforcement should sit alongside them in an examination file.