Version 7.0 of the NAIC AI supplement drops the proxy question

Version 7.0 of the NAIC’s AI Risk Evaluation Supplement, the exhibits regulators would use to question insurers about their AI systems, is now posted as a clean copy and an edited copy with margin notes. Every page is still marked DRAFT. Item 4 on the November 16 agenda of the Big Data and Artificial Intelligence (H) Working Group, meeting in Dallas, is “Consider Adoption of the AI Risk Evaluation Supplement v7.0.”

This week’s story: what 7.0 changed and what it kept

Version 6.0 allowed GLM exclusions only for Exhibit B through D inquiries; 7.0 lets regulators exclude GLMs and other predictive models “such as rule-based systems, deterministic algorithms, and fixed mathematical formulas” from the whole supplement. Of the industry requests on 6.0, that is closest to ACLI’s, though regulators still only “may” exclude.

Exhibit C’s item 10a, on “potential proxy relationships or effects,” carries the margin note “Deleted to address APCIA, BCBSA, and NAMIC’s input.” Item 3a, the agentic AI question 6.0 added, goes too; a margin note beside it credits BCBSA. Exhibit A keeps its agentic AI checkbox.

A new section on human in the loop lists what regulators “should consider” when judging human oversight controls, including whether overrides “actually occur and are documented with reasons.”

The Adverse Consumer Impact definition is unchanged from 6.0. Neither the qualifier NAMIC asked to restore nor BCBSA’s alternative wording made it in.

What to do this week

Two suggestions of ours, neither a requirement in 7.0:

• Find out, for each human review of a decision an AI system helps make, whether you could produce override and escalation rates and a log of overrides with reasons, the evidence 7.0 names. Then check each review against the three points 7.0 asks regulators to weigh, which it heads Authority, Qualifications, and Capacity and Evidence.

• Add to your GLM list from Issue 15 the rule-based, deterministic and fixed-formula models 7.0 names beside GLMs. Under 7.0 as drafted, a regulator may exclude such models, with GLMs, from the supplement where appropriate, and the list shows how much of your inventory that choice could reach.

On the Docket

• November 16, 2026: The Third-Party Data and Models (H) Working Group meets at 8:00 a.m. CT to “Discuss and Expose the Property/Casualty (P/C) Pricing and Underwriting Framework and Confidentiality Pathway,” and its meeting materials include a framework draft dated November 4.

• November 16, 2026: From noon to 1:00 p.m. CT, the working group behind the supplement is to hear a summary report on the pilot before item 4, consideration of 7.0 for adoption. On November 17, from 12:15 to 1:30 p.m. CT, the Innovation, Cybersecurity, and Technology (H) Committee is to consider adopting the reports of its task force and working groups, this one’s included, and hear Deloitte’s Michelle Donahue on how the audit profession is adapting for AI. Committee agenda

• November 21, 2026: At the annual meeting NCOIL (the National Council of Insurance Legislators) holds in Sanibel, Florida, item 6 of the NCOIL-NAIC dialogue is an “Update on Work of NAIC’s Big Data and Artificial Intelligence (H) Working Group.” NCOIL dates the schedule October 27 and marks it subject to change.

This week in brief

Governance

NAIC draft minutes of the October 8 meeting, filed with the November 2 meeting materials, say that a possible move of parts of the supplement into the financial or market conduct examination handbooks is to be discussed next year and that no decision has been made. Our November 4 report covers the rest of the minutes.

State adoption figures on our state tracker map date from November 3. That day, 28 of 51 jurisdictions had adopted the NAIC model bulletin, 4 had their own framework and 19 had issued nothing yet.

Business Lines

The Barrows plaintiffs, who allege Humana used naviHealth’s nH Predict model in coverage decisions, filed a motion on November 5 to certify a class of enrollees in certain Humana Medicare Advantage plans. The motion, which has long redacted passages and was filed alongside a motion to seal, defines the class by the plans’ coverage terms and their delegation of post-acute-care claims review to naviHealth. Georgetown’s case tracker, updated November 6, shows no ruling on the plaintiffs’ bid to name naviHealth as a defendant.

Pew Research Center asked 3,488 U.S. adults, June 22 to 28, about AI in health care and found that, on deciding an insurance claim, 5% were open to AI doing it alone and 38% to AI with a human doing a final review, while 46% were not open to it. Insurance Journal covered Pew’s October 29 report under the headline Americans Reject Use of AI for Medical and Coverage Decisions.

Whether a human stays in charge when AI helps decide is what Pew’s figures and 7.0’s human-in-the-loop section have in common. Melissa Hill, North American head of claims at Allianz, answers it in Insurance Business’s AI wants to please you, and that’s a problem for claims, Allianz exec warns: “The human is always accountable. The AI can’t make the decision.”

Our Elsewhere page collects what other outlets are reporting.

From the Guides

For compliance teams. Our November 4 guide separates the consumer notices an AI-driven underwriting or pricing decision already triggers, under laws based on NAIC Model 670 and under the FCRA, from the few rules that make AI use itself the thing to tell consumers: Do Insurers Have to Tell Consumers They Use AI?

Reply with the 7.0 change your team plans to check first.

– The Editor, InsureAI Wire