Filed rating plans built on generalized linear models (GLMs) are in scope of the NAIC's AI Risk Evaluation Supplement unless a state regulator chooses to exclude them. The Big Data and AI working group adopted the Supplement on November 16 and its parent committee on November 17, without the carve-out insurer trade groups requested, and states may use it from 2027.
The Big Data and Artificial Intelligence (H) Working Group acted at the NAIC's Fall National Meeting in Dallas, and the Innovation, Cybersecurity, and Technology (H) Committee followed a day later. The working group's meeting summary says it adopted the supplement "after reaching a consensus to agree on the two technical edits raised." It gives no vote count, and minutes of the session have not been posted. Item 3 of the parent committee's summary of its November 17 meeting is the adoption. Digital Insurance reported on November 17 that the tool was approved over industry objections and that regulators plan to roll it out in 2027.
Both votes end a drafting cycle in which insurer trade groups asked, in comment letters and at public meetings on October 8 and November 2, for rating models to be taken out. On November 10 the site reported that version 7.0 had widened the regulator's option to exclude them while keeping the models in the inventory. That is the text now adopted.
Key Takeaways
- Trade groups wanted the exclusion made mandatory. The life insurers' group asked for "should exclude"; the adopted text says regulators "may exclude some or all GLMs."
- Exhibit A asks whether AI systems in each of 13 operating areas, ratemaking among them, use a GLM. A request to drop rating GLMs from that exhibit failed.
- A rating model picked for deeper review answers 14 items in Exhibit C, covering drift, accuracy and unfair discrimination testing, plus its last validation date and results.
- States may use the Supplement voluntarily from 2027. Whether it goes into the NAIC's examination handbooks will be discussed next year, the working group's chair said.
What the NAIC adopted and when
The adopted text, posted on November 17, is version 7.0 with two changes from the draft released on November 9, a comparison of the two files shows. Exhibit B's narrative no longer asks a company to "include a detailed list of subject matter experts" for follow-up inquiries. Adverse Consumer Impact is now defined as a decision "subject to insurance law, regulations, and other applicable standards enforced by the Department," where the draft said "insurance regulatory standards."
NAMIC had asked for more on that definition. Its October 30 redline wanted the words "in a manner that violates those standards" restored, so that a lawful cancellation, non-renewal or premium increase would fall outside it. As adopted, the wording still ends at a decision "that adversely impacts the consumer."
Nothing else moved. Wisconsin Commissioner Nathan Houdek, the working group's chair, had said on November 2 that no further substantive changes would be made before adoption, according to the draft minutes of that call. He said the adopted version would be available for voluntary state use next year and that discussion of incorporating it into the Market Regulation Handbook or the Financial Condition Examiners Handbook would continue next year.
The Supplement is a set of four optional questionnaires that regulators can add to market conduct, financial analysis and financial examination reviews. Its intent section says it "does not independently create substantive legal duties, standards of conduct, admissions, or expanded regulatory authority."
What a filed GLM has to answer in Exhibits A to D
A rating plan built on a GLM meets the document first in Exhibit A. For each of 13 operation or program areas, one of them "Ratemaking/Rate Classification/Schedule Rating/Premium Audits," the company gives the number of AI systems in use, the number implemented in a period the regulator sets, and a use case stating whether there is a direct consumer impact or a material financial impact. Column 7 is a checkbox for whether any of those systems uses a GLM.
Pricing sits inside the consumer test under the definitions. Direct Consumer Impact covers outputs with the potential to create final consumer outcomes, "including but not limited to underwriting including tier and coverage eligibility determinations, premium determinations, ratemaking."
Exhibit B asks about the governance program as a whole. It includes whether validation and testing are done by someone independent of development, and what changes trigger approval, revalidation or enhanced monitoring.
Per-model questions are in Exhibit C, which covers high-risk models in production, with the risk criteria set by the company. Its 14 items start with name, version, use and model type, where "GLM" is one of three printed choices. Item 10 asks the company to discuss "testing for model drift, accuracy, unfair trade practices, unfair discrimination, performance degradation, and other potential Adverse Consumer Impacts," both before and after deployment. Item 11 asks for the last testing or validation date with "the results and corrective actions." Item 7 asks how, for a third-party model, the company "obtains sufficient evidence to assess and monitor risk and learns of material changes or incidents."
Exhibit D then lists 25 types of data, from loss experience and telematics to income and education level. For each one used in training or testing a high-risk model, the company states whether it came from inside or from a named vendor.
How a regulator can exclude GLMs and what trade groups asked
A regulator can leave a rating model out of every exhibit. The general guidance reads: "Based on existing regulatory oversight provided via rate making, regulators may exclude some or all GLMs and other predictive models such as rule-based systems, deterministic algorithms, and fixed mathematical formulas from the Supplement if appropriate." One sentence earlier, the guidance says GLMs "are not without risk of causing unfair discrimination or other Adverse Consumer Impacts."
Houdek described the expected practice for models that stay in on October 8. Companies using AI would receive Exhibit A, many would receive Exhibit B, and some would be asked to complete Exhibits C and D "for a specific set of AI models or data sets rather than for a large number of models," the minutes of that meeting say.
The 160-page compilation of comments on version 6.0 carries the trade groups' last written requests. Four of the five below wanted the exclusion written into the text and taken out of the individual regulator's hands.
Requests on GLMs in the version 6.0 comment letters and the adopted text
| Commenter, letter date | Request | Adopted text |
|---|---|---|
| InsurTech Coalition, October 22 | "An exclusion throughout the Supplement, including Exhibit A," for GLMs behind filed rates, rating factors, risk classifications and tier placement | Exhibit A keeps the GLM column for all 13 areas |
| NAMIC, October 30 | GLMs and other predictive models used in ratemaking "explicitly out of scope" of the whole Supplement | Exclusion from the whole Supplement is available, at the regulator's option |
| APCIA, October 30 | Exclude GLMs from Exhibits B, C and D | "May exclude some or all GLMs" |
| ACLI, October 30 | "Regulators should exclude GLMs and other predictive models" that produce rule-based systems, deterministic algorithms and fixed formulas | ACLI's list of model types added; the verb stayed "may" |
| Verisk, October 30 | Say the Supplement asks about models "when utilized within an AI System" | Added in those words |
Sources: NAIC, AI Risk Evaluation Supplement version 6.0 comments received; AI Risk Evaluation Supplement version 7.0 as adopted.
The InsurTech Coalition's letter, signed by Jennifer Crutchfield, tied its request to existing rate review. "The NAIC's Regulatory Review of Predictive Models and Model Review Manual support review of the actuarial characteristics and risks of ratemaking models," it said. Leaving rating GLMs in, it argued, "requires actuaries to participate in a parallel governance process."
APCIA wrote that although version 6.0 let regulators exclude GLMs, "that decision remains discretionary." NAMIC said an explicit exclusion would aid uniformity, "as opposed to some departments choosing to include them in the ratemaking context and others not."
Pilot results gave the working group a different reading of company practice. The summary prepared for the meeting by Coral Manning of Wisconsin says more than 100 companies took part; the pilot ran in 12 states from March through September. Most companies that answered a follow-up survey reported they could respond from existing documentation. On model scope the slides say: "Companies had varying practices on their inclusion of GLMs in scope of their AI governance framework."
How each state will apply the exclusion is undecided. So is confidentiality beyond the document's one line that responses "will be held confidential, consistent with the requesting state's relevant authority," and so is the choice of examinations that will carry the exhibits. On November 17 the committee adopted 2027 charges that keep the working group coordinating AI systems evaluation across the NAIC, its meeting materials show. An eight-part NAIC webinar series for regulators on the Supplement concludes in early 2027.
Further Reading
- NAIC Puts AI Supplement v7.0 to a Vote With GLMs Still in Scope: what changed between versions 6.0 and 7.0 before the vote.
- Six Trade Groups Ask the NAIC to Drop GLMs From Its AI Supplement: the first round of letters and the July actuarial panel.
- NAIC Draft AI Supplement Asks Insurers for a Model Inventory: the version 5.0 inventory request and materiality threshold.
- Model Risk Management for Actuarial Models, From GLM to LLM: validation practice for the model types Exhibit C names.
Sources
- NAIC Big Data and Artificial Intelligence (H) Working Group, Meeting Summary Report, 16 November 2026
- NAIC Innovation, Cybersecurity, and Technology (H) Committee, Meeting Summary Report, 17 November 2026
- NAIC, AI Risk Evaluation Supplement version 7.0, text for adoption, posted 17 November 2026
- NAIC, AI Risk Evaluation Supplement v7.0 (clean), 9 November 2026
- NAIC Big Data and Artificial Intelligence (H) Working Group, Fall National Meeting agenda and materials, 16 November 2026 (draft minutes of 2 November; pilot summary)
- NAIC Big Data and Artificial Intelligence (H) Working Group, 2 November 2026 materials (draft minutes of 8 October)
- NAIC, AI Risk Evaluation Supplement version 6.0, comments received
- NAIC Innovation, Cybersecurity, and Technology (H) Committee, Fall National Meeting agenda and materials, 17 November 2026
- Digital Insurance, "Pricing models to stay in AI review: NAIC rejects industry bid," 17 November 2026