NAIC AI Risk Evaluation Supplement v5.0 Makes the Model Inventory an Examination Exhibit

The NAIC's Big Data and Artificial Intelligence (H) Working Group exposed version 5.0 of its AI Risk Evaluation Supplement on August 31, 2026, opening a 30-day comment period that closes September 29. NAIC staff put the most consequential edit in a single line of the accompanying change summary: "Most notably, regulators now ask for a Model Inventory" (NAIC, August 2026).

The document narrowed its own scope claim and widened what it collects. Renamed from the AI Systems Evaluation Tool, version 5.0 states repeatedly that it supplements the Market Regulation Handbook, the Financial Condition Examiners Handbook and the Financial Analysis Handbook rather than standing alone, and it hands decisions about who receives an inquiry back to those handbooks. Underneath that framing, four exhibits now ask for material a pricing or reserving actuary keeps in internal validation files.

Key Takeaways

Version 5.0, exposed August 31 and open for comment until September 29, 2026, adds an explicit Model Inventory request to Exhibit A that NAIC staff single out as the most notable revision from version 4.0.
Generalized linear models received their own definition and new in-scope guidance in version 5.0, pulling conventional rating plans into an evidence request built for AI systems.
13 operation and program areas make up Exhibit A's grid, four of them financial rather than consumer facing: Reserves/Valuations, Investment/Capital Management, Catastrophe Triage and Reinsurance.
25 states and the District of Columbia had adopted the NAIC Model Bulletin as of July 2026. The bulletin asked for a written program; the supplement asks for the artifacts showing that program runs.
Two new checklist questions, 3o on materiality and 3p on third-party model oversight, turn a company's own scoping threshold and its vendor controls into written examination responses.

What Version 5.0 Changed

This is the first revision written after the field test. The pilot ran from March 2026 across 12 states, among them California, Colorado, Connecticut, Florida, Iowa, Maryland, Pennsylvania, Virginia and Wisconsin (Fenwick), with participating regulators meeting weekly (Monitaur). The change summary attributes several Exhibit B and Exhibit C edits directly to pilot feedback.

The rename tracks a real edit. The intent section now says the exhibits "are designed to supplement existing market conduct, financial analysis, and/or financial examination review procedures," and that inquiries "will be coordinated consistent with the guidance provided by" the three handbooks (NAIC, August 2026). The change summary gives the motive: prior comment periods produced feedback on scope, so decisions on which companies receive supplement-related inquiries now rest on handbook guidance.

Against that narrowing, the evidence requests grew. Exhibit A gained a second part, an AI Systems Request List, naming three items previously left implied: the company's materiality explanation, its risk assessment documentation, and a model inventory carrying each model's name, use case, program area, inherent risk level, consumer impact and financial impact. Exhibit B's checklist gained items 3o and 3p. Exhibit D gained a data dictionary request.

The machine learning guidance is the sharpest edit for actuaries. Version 5.0 adds a definition of generalized linear models and states that "GLMs are not without risk of causing unfair discrimination or other adverse consumer outcomes" and require governance over data quality, development, validation, implementation and ongoing monitoring (NAIC, August 2026). Exhibit A then counts GLMs in their own column, separate from generative and agentic models and from other AI/ML techniques.

That scope is not hypothetical. The NAIC's own private passenger auto survey found 88% of 193 responding insurers use, plan to use, or are exploring AI/ML models, with 70% of 194 homeowners insurers saying the same (Eversheds Sutherland). A conventional GLM rating plan sits inside that population.

Setting the Materiality Threshold

Version 5.0 aligned materiality to the Financial Condition Examiners Handbook, defining it as "the dollar amount above which the examiner's perspective of an insurer's financial position will be influenced" (NAIC, August 2026). Exhibit A's regulator instructions then offer a choice: the regulator specifies the threshold, or the company provides the one it used.

The change summary is blunt about the consequence, describing "a dynamic where a company may be asked to complete Exhibit A responses based on a materiality threshold the company specifies but also that it discloses to regulators" (NAIC, August 2026). A scoping judgment that was an internal convention becomes a written response an examiner reads next to the counts it produced.

For a pricing actuary that creates an asymmetry worth working through before an inquiry lands. Exhibit A splits its counts twice: models with Direct Consumer Impact, and models with Material Financial Impact. Direct Consumer Impact carries no dollar test, covering outputs that determine "underwriting including tier and coverage eligibility determinations, premium determinations, ratemaking, and claims denial and settlement-related decisions" (NAIC, August 2026). Material Financial Impact runs off the disclosed threshold.

A personal auto rating GLM rarely clears a solvency-scale materiality threshold on its own, so it lands in the consumer column and not the financial one. A reserving model at the same carrier does the reverse. Raising the threshold shortens the financial column and leaves the consumer column untouched, and the gap between the two counts sits on one page for the examiner to read. The threshold that minimises one disclosure widens the visible distance to the other.

Set against the Model Bulletin, which 25 states and the District of Columbia had adopted as of July 2026 while California, Colorado, New York and Texas run their own frameworks (Quarles), the shift is from policy to proof. The bulletin asked insurers to maintain a written AI systems program. The supplement asks for the document name and page number where each element of that program is described, a request added to the checklist in version 5.0.

Where the AI Inventory and the Actuarial Model Inventory Diverge

Exhibit A offers what reads like a shortcut. If a company has an existing AI Systems inventory, the instructions say, "it may suggest submission of the inventory in lieu of completing Exhibit A" (NAIC, August 2026). Most insurers built that inventory to the Model Bulletin's frame, which is organised around adverse consumer outcomes.

Exhibit A is organised differently. Its grid runs across 13 operation and program areas, four of which are financial rather than consumer facing: Reserves/Valuations, Investment/Capital Management, Catastrophe Triage and Reinsurance. The appendix widens it again, listing reserving, financial reporting and capital management under other operations, and placing "individual/bulk claim reserving including loss estimation" inside claims.

That is where two lists nobody reconciled meet. The actuarial model inventory maintained for model risk management holds reserving, capital and catastrophe models classified as actuarial rather than AI, usually with a validation history the AI inventory never captured. Submitting the AI inventory in lieu of Exhibit A asserts the two agree. Where a reserving model built on gradient boosting appears in one list and not the other, the discrepancy surfaces during an examination instead of during a validation review.

The vendor questions compound it. Exhibit C field 5 asks whether each model was developed internally, by an "owned" third party or by a true third party, and field 6 asks whether the company can modify it. New checklist item 3p asks how the company governs, monitors, tests and ensures transparency of vendor-developed systems, and the narrative exhibit extends that to third-party providers "including actuarial, claim, MGA, audit, and/or other professional services" (NAIC, August 2026), which reaches a consulting firm's reserving model.

For an AI feature embedded in a policy administration or claims platform, the honest answer to field 6 is no. The insurer cannot modify it, cannot inspect its training data, and in many cases never selected it as a model at all. Exhibit D still asks, for each of 25 data element types running from telematics to facial recognition to geo-demographics, whether the source is internal or a named third-party vendor, with a data dictionary available on request. The handbook deference the rename introduced settles who receives the inquiry. It does not settle what a truthful answer looks like when the model belongs to someone else.

NAIC AI Risk Evaluation Supplement v5.0 Makes the Model Inventory an Examination Exhibit

Key Takeaways

What Version 5.0 Changed

Setting the Materiality Threshold

Where the AI Inventory and the Actuarial Model Inventory Diverge

Further Reading

Sources