NAIC AI Model Bulletin for Marketing and Lead Scoring: Scope, Controls and Evidence (as of October 2026)

What the bulletin text says and what is interpretation

The NAIC adopted its model bulletin on the use of AI by insurers in December 2023. It is framed as a reminder to insurers holding certificates of authority that decisions or actions impacting consumers, made or supported by advanced analytical and computational technologies including AI Systems, must comply with all applicable insurance laws and regulations. The model has effect only where a state issues its own bulletin or rule.

Two features matter for marketing. First, the AIS Program is meant to address AI Systems across the insurance life cycle. The November 2023 draft lists product development and design, marketing, use, underwriting, rating and pricing, case management, claim administration and payment, and fraud detection. Marketing is named, so a pre-underwriting use is not excluded by its position in the sales process. The July 2023 exposure draft went further and listed "marketing, lead generation and use", while later drafts read "marketing, use". That is a drafting difference, so a scope position should not rest on the words "lead generation".

Second, the trigger is "made or supported by". This is interpretation, not a regulator statement about marketing: on its face the phrase is broad enough to reach analytics that support a decision rather than make it, such as a score a sales team acts on. NAIC materials identify targeted online advertising and offers to existing customers as common marketing uses of AI, but they do not say whether those are decisions impacting consumers. The practical test is whether the output changes who is contacted, offered a product or routed to a channel.

Typical marketing uses and the scope question each raises

The table applies the bulletin's consumer-impact framing to four common uses. These are practitioner readings, not regulator positions, and each should be recorded with its rationale.

• Audience segmentation: Ask whether segments decide who receives or is excluded from an offer, and which data, including external consumer data, feeds them.
• Propensity scoring: A score that ranks prospects and informs who is contacted or prioritized plausibly 'supports' a consumer-impacting decision. Document the reasoning either way.
• Lead routing: Routing can change which channel, agent or product a consumer sees. Treat it as in scope unless a documented analysis shows otherwise.
• Generative outreach: Review what consumer data enters prompts, what the model may say about products, and which tools or systems it can call.

Which expectations follow for in-scope uses

Where a use is in scope, the AIS Program expectations apply as they would to underwriting or claims AI, scaled to risk. Governance and risk management mean the use sits in an inventory with a named owner and documented controls. The drafts describe validating, testing and retesting to assess how AI outputs generalize on implementation, including the suitability of the data used to develop, train, validate and audit the model. For lead scoring, that means testing scores and segments, and checking whether inputs could produce unfair discrimination under applicable insurance law. The drafts do not prescribe test metrics or thresholds for marketing, so the carrier defines and records its own.

Third-party oversight is where marketing programs are most exposed, because scoring and personalization often run inside vendor platforms and LLM tools. The November 2023 draft lists due diligence on third parties and their data, models or AI Systems. A workable reading is an inventory of which vendor systems touch consumer data or outreach decisions, with due diligence on file for each.

Documentation expectations in the drafts include inventories and descriptions of algorithms and predictive models, data source, provenance and lineage, information about data used for a specific model, and records of validation, testing and auditing. The bulletin is addressed to insurers holding certificates of authority. How it reaches brokerages depends on the state and entity type, though carrier third-party oversight may carry expectations to distribution partners and marketing vendors.

State adoption, Colorado and what has changed

As of October 2026, NAIC's adoption map (status date August 6, 2026) lists 20 adopting jurisdictions: 19 states plus the District of Columbia. Dates run from Alaska on February 1, 2024 to Wisconsin on March 18, 2025, so none of the listed adoptions falls within the past 12 months. California Bulletin 2022-5, issued June 30, 2022 before the model, is listed separately. The map shows adoption dates only. It does not say whether a state's version contains every model element, so each operating state's own text governs.

Colorado has changed. SB 24-205 is no longer the right planning baseline. Senate Bill 26-189, signed in May 2026, repeals and reenacts those provisions with new requirements for automated decision-making technology in consequential decisions. Duties and dates from SB 24-205 should be treated as superseded, and compliance teams should work from the current statute and rulemaking rather than earlier summaries.

The practical consequence is a state-by-state matrix: for each operating state, whether a bulletin or rule exists, what it says about marketing and distribution, and which marketing uses it reaches.

Evidence: runtime records versus model-level documentation

Examination evidence for marketing AI has two layers, and they should not be conflated. Model-level evidence covers inventories, descriptions of algorithms and predictive models, validation and testing records, data suitability, data source and lineage, and vendor due diligence files. Runtime logs of LLM and tool interactions do not produce these artifacts.

Runtime evidence answers a different question: what was the AI allowed to do on a given interaction, and what did it do? For generative outreach and agent-style tools, that includes which policy applied, whether consumer data was redacted before reaching a model, which tool calls were permitted or blocked, and which actions went to a human for approval. Treat this as supplementary to model-level documentation, not a substitute.

A third record is the scope decision itself. For each marketing AI use, keep the analysis against the "made or supported by" language, the state-by-state view, and the date it was last reviewed, so an examiner can see how and why a use was treated as in or out of scope.

Where runtime enforcement fits

Trussed enforces policy in the execution path on individual interactions and tool calls across models, agents, tools and MCP servers. Policies are enforced at runtime, not reviewed after the fact, with outcomes of allow, redact, block, rate limit or escalate for human approval. That enforcement produces continuous audit-ready evidence mapped to NAIC and other regimes. It can be deployed self-managed in the customer cloud, with configurable retention and data residency.

For marketing, this applies to LLM-based outreach tools and agents that read consumer data or call downstream systems. Trussed does not provide model lineage, dataset provenance or transaction-level governance, so those artifacts remain with model and data documentation.