NAIC AI Model Bulletin vs NYDFS Circular Letter 7

Legal nature and applicability

The two documents differ in origin before they differ in content. The NAIC adopted its model bulletin in December 2023. It is a model: each state issues its own bulletin through its Department of Insurance, so the obligations a carrier faces depend on which states have issued a bulletin and whether they changed the model text. The NAIC describes the bulletin as setting guidelines and expectations for responsible AI use that align with the NAIC Principles of Artificial Intelligence, adopted by the full membership in 2020.

NYDFS Circular Letter 7 is a single regulator's guidance, titled for use of AI systems and external consumer data and information sources in insurance underwriting and pricing. The final version was issued July 11, 2024, after a proposed version, and as of October 2026 no amendment or replacement has been reported. DFS describes it as guidance outlining its expectations. Its addressees are all insurers authorized to write insurance in New York, Article 43 corporations, health maintenance organizations, licensed fraternal benefit societies and the New York State Insurance Fund.

This page is a compliance comparison, not legal advice. Questions about enforceability, penalties or equivalence between the regimes belong with counsel.

Side-by-side comparison by compliance dimension

The table is organized by dimension rather than by regime. NAIC entries reflect the model bulletin's adoption and the published drafts that preceded it, so check each state's issued bulletin before building a clause-level crosswalk.

• Nature and applicability: NAIC: model bulletin adopted December 2023, issued by each state Department to insurers holding certificates of authority. NYDFS: one circular letter, issued July 11, 2024, addressed to insurers and related entities authorized in New York.
• Scope and definitions: NAIC: the November 2023 draft covers decisions or actions affecting consumers that are made or supported by advanced analytical and computational technologies, including AI systems. NYDFS: defined by function. An AIS is covered if used to supplement or act as a proxy for underwriting or pricing, or to identify lifestyle indicators. ECDIS is defined on the same functional basis. AIS use is in scope whether or not ECDIS is used.
• Governance and risk management: NAIC: an AIS Program designed to mitigate the risk of Adverse Consumer Outcomes, with protection of non-public information, including unauthorized access to predictive models, as a consideration. NYDFS: expectations are set for insurers as a whole, and accountability for third-party AI and data stays with the insurer.
• Discrimination testing: NAIC: decisions must comply with all applicable insurance laws, and adverse consumer outcomes are the stated risk. NYDFS: quantitative proxy assessments and disproportionate-adverse-effect analyses, limited to protected classes with available or reasonably imputable data. Insurers are not expected to collect additional data about individuals.
• Third-party oversight: NAIC: draft topics include due diligence and vendor contracts covering data security and privacy, data sourcing and use, confidentiality and cooperation with regulators. NYDFS: appropriate vendor oversight, plus a clause drawn from the NAIC bulletin that contract terms be included where appropriate.
• Documentation and examination: NAIC: drafts anticipate examiner questions on governance framework, risk management, internal controls, and specific models, including data source, provenance, lineage, techniques, measurements and thresholds. NYDFS: the analyses and vendor oversight above are the artifacts to be ready to produce.

Where the regimes overlap and where they diverge

The overlap is substantial. Both place accountability on the insurer, including for AI and data supplied by third parties. Both expect a governance structure around AI use, attention to discrimination risk, and oversight of vendors. NYDFS even borrowed contract-term language from the NAIC bulletin when it finalized its letter, so vendor contract reviews can be shared across the two.

The divergence is mainly in how scope is drawn and how specific the testing expectation is. NYDFS defines coverage by function within underwriting and pricing and states that AIS use is covered regardless of whether external data is involved. A model inventory that excludes systems because they use no external data would miss in-scope systems in New York. The NAIC draft language reaches further, to consumer-impacting decisions made or supported by AI systems, which can include activity beyond underwriting and pricing.

NYDFS is also more explicit about analysis. The final letter narrowed its proxy and disproportionate-adverse-effect provisions to protected classes for which data are available or can reasonably be imputed with statistical methods. Teams that rely on imputation should document the method, because the analysis only holds up if the approach can be explained. The NAIC material frames testing through the AIS Program and adverse consumer outcomes rather than prescribing this specific analysis.

One caution applies to multi-state carriers. There is no single uniform NAIC obligation set, because states issue and may adapt their own bulletins.

Which obligations depend on runtime control and durable evidence

Most of what both regimes ask for is program-level: governance structures, risk assessments, statistical testing, contract terms and documentation of methods. Runtime enforcement does not replace any of these. Proxy analysis, disproportionate-adverse-effect testing and vendor due diligence are performed and documented outside the execution path.

A narrower set of expectations depends on what AI was actually allowed to do and what it did. Protecting non-public consumer information, restricting access to predictive models, and being able to show per use which controls applied all turn on behavior at the point of use. Examiner questions about specific models and data sources are easier to answer when records tie each AI interaction to the policy that governed it. Whether a regulator requires records generated at the point of use is not stated in either document, so this is a design choice rather than a mandate.

Trussed addresses this narrower layer. It enforces policy in the execution path across models, agents, tools and MCP servers, with actions to allow, redact, block, rate limit or escalate an interaction for human approval. That enforcement produces audit-ready evidence, mapped to NAIC and NYDFS, of what AI was allowed to do and what it did. The scope is individual interactions and tool calls. Trussed does not perform bias testing, does not establish a governance program, and does not provide model lineage or dataset provenance, so those items remain with the carrier's own program.

Scoping a multi-jurisdiction compliance approach

Build one control set mapped to both regimes, then add jurisdiction-specific work where the documents diverge. The sequence below keeps the shared work shared.

• Inventory by function and by impact: Tag systems that meet the NYDFS AIS definition, including those using no external data. Separately flag consumer-impacting uses that fall under NAIC-style coverage.
• Confirm the state-level texts: Identify which of your states have issued the NAIC-based bulletin and whether any changed the model language before mapping clauses.
• Plan the NYDFS analyses: Define protected classes with available or imputable data, document the statistical imputation approach, and avoid new data collection.
• Review vendor contracts once: Check for data security, data sourcing and use, confidentiality and regulator cooperation terms, which serve both regimes.
• Separate evidence types: Assign testing and governance artifacts to program owners, and assign per-interaction control records to runtime enforcement.
• Rehearse a regulator request: Test whether you can produce, per model or use, data source, thresholds applied and the controls in force.