ISO/IEC 42005 vs NAIC Bulletin Consumer Impact Classification: One Assessment or Two?

What each framework asks a carrier to assess

ISO/IEC 42005:2025 is titled "AI system impact assessment". It is an international standard published through ISO/IEC JTC 1/SC 42, not a law. Its title fixes the unit of analysis as the AI system, which for an insurer means an underwriting triage tool, a claims routing workflow or a customer-facing assistant, together with the models inside it.

The NAIC Model Bulletin on the Use of Artificial Intelligence by Insurance Companies was adopted by the NAIC in December 2023. It sets guidelines and expectations for responsible AI use by insurers, aligned with the NAIC Principles of Artificial Intelligence. Because it is a model bulletin, the version a carrier must follow depends on what each state in which it operates has adopted.

The NAIC AI Risk Evaluation Supplement, version 5.0, is where classification becomes concrete. As of November 2026, version 5.0 has been circulated for stakeholder comment, and its final form should be confirmed before the mapping is fixed. Commenters describe the Bulletin as emphasizing governance, accountability, proportional controls and risk-based flexibility, and the Supplement as requiring detailed inventories, documentation and evidence. In the Supplement, risk criteria are set by the insurance company, and materiality should be a consideration in determining the level of risk. It requests a model risk classification (high, moderate, low, etc.) and a description of the potential for adverse consumer impact.

Side-by-side: where they align and diverge

The two frameworks overlap in purpose. Both ask a carrier to reason about what an AI system could do to people and to record the result. They differ in the unit they assess, who defines the criteria, and what artifact is expected at the end.

• Nature: ISO/IEC 42005: international standard for AI system impact assessment. NAIC: model bulletin adopted December 2023, with a Supplement that asks for inventories, documentation and evidence.
• Unit of analysis: ISO/IEC 42005: the AI system. NAIC Supplement as described: classification at the model level, with at least one commenter recommending the system as the primary unit.
• Criteria: NAIC Supplement: set by the insurer and informed by materiality. A regulator may specify a materiality threshold or rely on the company's determination.
• Classification output: NAIC Supplement: a risk label (high, moderate, low, etc.) plus a description of potential adverse consumer impact.
• Model type: NAIC Supplement Exhibit A counts generative and agentic models separately. A shared record should carry model type as an attribute.
• Evidence emphasis: NAIC: documentation and evidence that an examiner can review. A shared record should point to retained evidence, not only to the assessment narrative.

From assessed tier to runtime enforcement and examiner evidence

An assessment describes expected risk. An examiner will also want to know what the system was permitted to do and what it did. That gap is where programs tend to fail. A high tier that exists only in a spreadsheet does not change what a claims agent, a model or a tool call can do in production.

The structured tier field is the link. Each tier can map to a runtime policy set that states what interactions and tool calls are allowed, which data must be redacted, which actions are blocked, and which require human approval. The retained evidence should then show that the policy for the assessed tier was applied, with enough detail to be reviewed later.

This is the point at which a runtime enforcement layer supports the requirements. Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants: allow, redact, block, rate limit, and escalate for human approval, at the level of individual interactions and tool calls. Enforcement is at runtime, not reviewed after the fact, and it generates continuous audit-ready evidence mapped to NAIC, among other regimes. Trussed supports 100+ models under one policy plane and is deployed self-managed in the customer cloud, with configurable retention and data residency.

Trussed does not perform the impact assessment or assign the tier. It also does not provide model lineage or transaction-level governance. The carrier owns the assessment. The enforcement layer applies the resulting policy and produces the evidence that it operated.

Limits of the one-assessment approach

A single core record reduces duplicated work and inconsistent tiers, but it does not remove every difference. Three limits are worth planning for.

First, unit of analysis. If NAIC remains model-oriented while the carrier assesses at the system level, some outputs will need a model-level view derived from the system record. If NAIC moves to a system-level approach, as one commenter recommends, the mapping simplifies, but only after the final Supplement text is known.

Second, criteria ownership. Because the Supplement lets the insurer set criteria, internal tier definitions can drift from what a regulator expects. A regulator may specify a materiality threshold, so the record must accommodate more than one threshold.

Third, state variation. Carriers should track which states have adopted the Bulletin and whether any state expectations differ, since each jurisdiction-specific output should reflect the state in question. Where a state adds requirements the core record cannot satisfy, a supplementary process is justified. Even then, the core assessment should remain the single source of truth, so the tier an examiner sees matches the tier enforced in production.