CLIENT ALERT

NAIC releases version 7.0 of its AI Risk Evaluation Supplement Ahead of Fall National Meeting

November 13, 2026

Read time: 10 min

Overview

The National Association of Insurance Commissioners’ (NAIC) Big Data and Artificial Intelligence (H) Working Group (the Working Group) has circulated a draft version 7.0 of its Artificial Intelligence (AI) Risk Evaluation Supplement (the “Supplement”, formerly known as the AI Systems Evaluation Tool). As discussed in our recent client alert, the Supplement gives market conduct and financial examiners a set of optional exhibits (Exhibits A through D) for asking about the use of AI by insurers and their agents.

Version 7.0 was released following a 14-day exposure of version 6.0 of the Supplement that closed on October 30, 2026. During that window, the Working Group received more than 20 comment letters from trade associations (including the American Council of Life Insurers, the American Property Casualty Insurance Association, the National Association of Mutual Insurance Companies, America’s Health Insurance Plans, and the Blue Cross Blue Shield Association), vendors, and AI-governance consultants. The Working Group plans to present version 7.0 for adoption at the NAIC’s Fall National Meeting, which takes place in Texas November 14 – 17.

Industry reaction to these proposed changes has generally been positive, reflecting thoughtful consideration of industry input. While adding targeted governance questions, version 7.0 narrows the Supplement in ways that could ease the burden on respondents.

In depth

Key changes in version 7.0

Perhaps the most notable change involves generalized linear models (GLMs). Under version 6.0, regulators could exclude some or all GLMs only from inquiries pursuant to Exhibits B through D. Version 7.0 permits regulators to exclude some or all GLMs and other predictive models, such as rule-based systems, deterministic algorithms, and fixed mathematical formulas, from the Supplement altogether since rate filings already provide sufficient oversight of those models. This follows an industry suggestion that the exclusion turn on whether a model is static and transparent, not on whether it carries the GLM label. The Supplement also makes clear now that model type is just one factor in deciding whether to dig deeper.

Additionally, Version 7.0 adds helpful guidance on “human in the loop” controls, especially for agentic AI where such practices are more common. Regulators are asked to consider three things:

  • Whether the reviewer has real authority to reverse, modify, or escalate a decision (and whether overrides actually happen and are documented)
  • Whether the reviewer has the experience and qualifications to evaluate the performance of the AI tool being used
  • Whether the design of the human review process supports effective oversight and can be shown to work through measures such as override and escalation rates

This proposed framework draws on comments from several AI-governance commenters.

Finally, the Supplement now addresses how examiners verify what insurers tell them. Throughout the exhibits, regulators pair questions with requests for supporting documentation so they can corroborate management’s assertions (i.e., the company’s own descriptions of how it uses AI and how its governance, testing, and human-review controls operate) rather than taking those responses at face value. The Supplement clarifies that, because it is not designed as a standalone resource, it does not provide guidance on standards of fieldwork; regulators deciding whether inquiry alone is sufficient or corroboration is needed are instead directed to the existing NAIC examination and analysis handbooks. In doing so, the Working Group chose not to adopt the formal evidentiary frameworks that several commenters had proposed.

Revisions to the Supplement’s exhibits in version 7.0 respond to industry calls to reduce duplication and burden. For example, insurers that already maintain an AI inventory by business function are encouraged to first talk with the examiner about whether the inventory satisfies the Exhibit A request before submitting the inventory. The materiality-threshold field and related instruction have also been removed from Exhibit A, leaving materiality to the General Guidance.

The Exhibit B checklist regarding the insurer’s AI governance program no longer asks for a “detailed list of subject matter experts.” Insurers can now identify an appropriate contact, function, or team (although the narrative version of Exhibit B still asks for such a list in its governance questions). The narrative question on inherent risk has also been reworded following industry suggestion to refer to “compliance with existing insurance law” rather than presuming unfair trade practices or discrimination.

Exhibit C, which focuses on high-risk AI Models, dropped the standalone agentic AI question. It also removed the question on testing for “proxy relationships,” which industry groups had opposed, and simplified the testing and validation question.

Exhibit D, which focuses on the data AI Models use, is now limited to data used in high-risk AI Models.

Version 7.0 adds new questions, including whether validation is independent from development and what material changes trigger approval, revalidation, or enhanced monitoring (with an illustrative list of triggers). For third-party models, the examiner may ask how the insurer obtains enough evidence to monitor risk and how it learns of material changes or incidents. “Prior authorizations” has also been added to the examples of Direct Consumer Impact.

The Working Group did not incorporate several industry requests, including:

  • The exclusion of GLMs and other predictive models remains discretionary.
  • The definition of “Adverse Consumer Impact” still omits the qualifier that the impact must violate regulatory standards, leaving unresolved the industry’s concern that lawful actions such as non-renewals or premium increases could be captured.
  • Exhibit B still includes questions on discovering unsanctioned AI use, governing enterprise AI assistants, and deciding when a general-purpose tool becomes high-risk, despite feedback that the questions could reach routine productivity tools.

Next steps for insurers

The Supplement will shape what examiners expect to see on AI governance. We anticipate Exhibit A will serve as the main entry point, with the other exhibits reserved for higher risk uses. Insurers may wish to (i) confirm their AI inventories capture the identified fields, (ii) document the materiality thresholds they use to decide whether the effects of an AI System are meaningful and the reasoning behind those determinations, (iii) gather evidence of ensuring meaningful human review, and (iv) prepare explanations of why their GLMs and other predictive models may warrant exclusion, given that the exclusion is at the regulator’s discretion.

Please reach out to Michael Byrne, Lauren J. Dalton, or Leila Fujiwara with any questions.

Authors

H. Michael Byrne

Partner

New York – One Vanderbilt Avenue

More insights