How State Regulators Review an AI Pricing Model in a Rate Filing

State regulators judge a model-based rate under the rating law and expect the filer to tie each variable to expected losses, with NAIC guidance on what to file.

For
Pricing actuaries, rate-filing analysts, and compliance leads at U.S. property and casualty insurers whose rating plans include machine-learning or other complex predictive models.
Read if
Your rating plan now includes a gradient boosting model, a neural network, or a vendor score, and your filing team needs to know what a state reviewer will ask to see, and when.

By Simon Li · Published Nov 12, 2026 · 12 min read

Ask AI
Engraved cover illustration: How State Regulators Review an AI Pricing Model in a Rate Filing
In this article

A gradient boosting machine makes a rate filing harder to support without changing the law it is judged by. The state reviewer still needs to see how each variable relates to expected losses or expenses, and a model built from many trees makes that difficult to show and to check.

The NAIC’s Casualty Actuarial and Statistical Task Force (CASTF) has written guidance on what a reviewer can ask for, and when.1 For training, the NAIC’s review team invented a rate filing built on a GLM, planted mistakes in it and listed them. That list is the case this article follows, and we read it as applying to gradient boosting as well.

The test is the rating law, applied to each variable

CASTF’s white paper, “Regulatory Review of Predictive Models,” adopted in 2020, says nothing in it “is intended to, or could, change” the legal standards for approving rating plans.2 Those standards typically require that rates not be excessive, inadequate, or unfairly discriminatory.3

The white paper asks reviewers to determine whether each input characteristic, and the rating factor it produces, relates to differences in expected loss or expense.4 That puts each rating variable under its own question, and each factor needs its own actuarial justification. The NAIC’s rating model law comes in two versions, prior-approval and file-and-use. Both let a classification measure a difference among risks that “can be demonstrated to have a probable effect upon losses or expenses.”5

Where a state has adopted the NAIC’s AI Model Bulletin, the prior-approval model law’s requirements apply “regardless of the methodology” used to develop rates, machine learning included, the bulletin says.6 It adds no new rating standard. What it adds, a written AI Systems program, belongs to AI governance across the company rather than to any one rate filing.6

Who reads the model, and who decides

The decision stays with the state, whose reviewer at the department of insurance judges whether the filing complies and sends any objections. The NAIC’s help is a service set out in the NAIC Model Review Manual, which CASTF adopted on November 4, 2025.7 A state that signs the NAIC’s free Rate Review Support Services Agreement may ask the NAIC rate model review team to review a model in a filing, and need not.8 The NAIC does not publish which states have signed, though a state may say it has.9

The Manual also limits whose models the team takes on: those of the 20 largest insurance groups and of vendors licensed or operating in 10 or more states.10 Exceptions may arise, such as an innovative modeling approach from a smaller company or a new product.10

The team writes a report for the state’s reviewer, which the NAIC calls working papers for rate filing reviews.11 One analysis the team does not do is a separate check for correlation with unlawful characteristics or for what the Manual calls disparate impact.12 Any disparate-impact test a filing faces comes from the state.

The team has no regulatory authority and sends no objections to the company.12 What it finds reaches the filer only through the state, for example in the state’s objections; if the company’s answers leave questions open, the Manual has the state schedule a follow-up call.13 The filer is not given the report as part of the review: the Manual keeps NAIC reports “confidential and for regulator use only.”9

What a reviewer asks for, and when

The white paper’s information list sets out, in four levels, what a reviewer may need to review a predictive model built as a GLM for personal auto or home rating.14 Level 1 is needed to begin the review and ideally comes with the initial submission. Level 2 is needed for all but the most basic models; insurers concerned with speed to market may file it too. Levels 3 and 4 are typically requested only if the reviewer has concerns, Level 4 when they are serious concerns about excessive, inadequate or unfairly discriminatory rates.14 The white paper calls the list long but not exhaustive, and does not mean every item for every filing.15

For a vendor model inside the rating plan, the white paper’s list asks at Level 1 for the vendor’s name and the sub-model’s name and version.16 Under ASOP No. 56, the profession’s modeling standard, an actuary relying on a vendor model with limited ability to see how it works should disclose the extent of that reliance.17

The Manual adds checklists, one for each model type: GLM, tree-based, penalized regression and neural network.18 The Manual as adopted carries them in version 2025.2.19 Each checklist’s items are either “Essential,” which the NAIC team requests before writing a full-scope initial assessment, or “Sometimes Needed,” which may only matter if something about the modeling looks non-standard.20

Neither document says how the white paper’s levels correspond to the checklists’ two categories. The Manual leaves the connection to each state, whose regulators may share the lists with insurers, build them into the state’s rate filing checklists, or raise items through objections.20 For a filer, the working list is the state’s own filing requirements, and the NAIC’s items show what a reviewer may later ask for.

What the NAIC team flagged in its own sample filing

In April 2026 the NAIC team showed CASTF a sample filing from Golden Retriever Insurance Company, a fictional insurer, for a personal auto plan built on a GLM of pure premium.21 It was written for training, and its data are, in the team’s words, “completely made up.” The filing adds external consumer data from a fictional vendor’s purchase records, and three variables drawn from them: recent purchases of a Fender guitar, a Casualty Actuarial Society seminar, or Coldplay tickets. The team says the filing meets most of the Manual’s GLM checklist, yet it found 30 issues and ranked five as high priority.22 Three of those five come up below: the unstated count of screened variables, fit shown on training data, and the surcharge on missing ticket data.22

The filing says every available variable from the vendor went into a preliminary model, and only those with low p-values were kept.23 It does not say how many there were, and the team noted that a few variables “may pass p-value threshold by chance.”24 The same scrutiny applies when machine learning only screened the variables for a GLM. One Level 1 item on the white paper’s list asks the filer to comment on any automated feature selection and explain how overfitting was addressed.25

The white paper uses “rational explanation” for a plausible account of how a variable connects to real-world circumstances or behaviors that contribute to the risk of insurance loss. It has to be readily understandable to “a consumer or other educated layperson,” but it does not require strict proof of causality.26 The planted explanations in the filing’s data dictionary are meant to be caught: people who play Fender guitars “have good taste and avoid distasteful activities such as car accidents.”27 The subtler error was the insurance score, which the filing justified by its very low p-value. The team answered that a statistical test is not a rational explanation, which should offer “a plausible connection between the characteristic and insurance claim risk.”24 Where no rational explanation can be provided, the white paper says, “greater scrutiny may be appropriate.”28

The filing’s fit plots compare actual and predicted pure premiums on the model’s own training data. Such plots, the team’s note says, “should be provided based on test data.”24 The white paper’s list makes goodness of fit on validation data a Level 1 item; some regulators also require model validation on state-only data.29

Two of the filing’s selected factors go wrong in different ways. The Coldplay variable has a “Missing” level for policies with no ticket data, and the filing gave that level the same factor as “Yes.”30 The team found it unreasonable to surcharge when the company does not know whether the person bought tickets.24

The other departs from the model’s own indication. For insurance score tiers 7 to 10, the filing proposed factors much higher than the indicated ones and gave “current underwriting appetite” as its reason.30 The team asked for additional justification.24 The white paper’s list has a rule for this case: a selected factor should move toward the indicated one. If it does not, “an explanation is necessary,” especially where the gap is material or falls more on one group of consumers than another.31

On our reading, none of these issues depends on the model being a GLM. Each is a link between model and cost that the filing asserts without showing. A filing built on gradient boosting has the same four things to show: how the variables were screened, fit on unseen data, a reason for each variable and a reason for each departure from indication.

The sample is fictional and its model a GLM, and our search found no public regulator record of a real filing turned down for using machine learning. The search covered public pages and documents, not each state’s filing database, so the absence does not show that machine-learning filings pass easily.

What changes for tree ensembles and neural networks

For a GLM, the white paper’s list puts the rational explanation at Level 3, a follow-up request.28 For random forests and gradient boosting machines, CASTF’s Appendix B-TREES uses the same four levels and moves it to Level 1, needed at the start.32 The tree appendix pairs the explanation with a plot for every variable (partial dependence, accumulated local effects or Shapley) and with variable-importance plots, also at Level 1.32 For fit, the Manual’s tree checklist asks, as Essential, for decile plots on data not used in training, for the state and countrywide.33

The two tree documents differ on reproducing a rate. The tree appendix asks at Level 1 for documentation that reproduces a rate “for any theoretical risk,” such as every tree diagram.34 The Manual’s tree checklist asks for less: as Essential, a diagram of the first tree in each model and rating exhibits for 10 sample risks, with full audit documentation Sometimes Needed.33

On our reading, a tree-based filer sends the first-tree diagrams and rating exhibits, and keeps the full trees or split rules ready, since the tree appendix treats them as needed to begin.

CASTF has no neural network appendix yet, so a filer with one uses only the Manual’s checklist.35 Unlike the Manual’s other three checklists, it does not ask for a rational explanation of each variable.36 Its modeling section asks for Shapley plots by variable and for commentary on what drives the score of the 10 worst-scored records, or images for an image model.36 Its implementation section asks whether a person could reproduce a prediction “using an objective criteria checklist.” Under validation, it asks whether predictions were checked against an independent report, such as a roof inspection for a model that labels roofs from images for homeowners pricing.

What NAIC guidance asks for, by model type
ItemGLM: white paper37Tree-based: tree appendix and Manual checklist38Neural network: Manual checklist only36
What comes firstLevel 1 items, ideally in the initial submissionTree appendix: Level 1 items. Manual: Essential itemsEssential items
Per-variable plots and explanationRational explanation, Level 3Tree appendix: a plot and a rational explanation for every variable, Level 1. Manual: both EssentialShapley plots by variable, Essential. Commentary on what drives the score of the 10 worst-scored records or images, with waterfall plots or Grad-CAM images, Essential. No rational explanation item
Fit and validationGoodness of fit on validation data, Level 1Manual: decile plots on data not used in training, state and countrywide, EssentialComparison of predictions with an independent report, Essential
Reproducing a resultFiled plan should let a regulator validate a premium; rating manual checked against the model, Level 1; premium calculator for complex plans, Level 3Tree appendix: a rate reproducible for any theoretical risk, Level 1. Manual: first-tree diagrams and rating exhibits for 10 risks, Essential; full audit documentation, Sometimes NeededWhether a person could reproduce a prediction with an objective criteria checklist, Essential

When guidance becomes a filing requirement

The white paper leaves each state to “decide how best to tailor” its guidance, and notes that states with prior-approval authority over personal lines often already ask for many of its items.39

Nebraska made an NAIC-style list the price of a faster review. Since January 2, 2024, a model-dependent filing gets expedited review only if the filer states each model’s role in the SERFF filing description and attaches a signed ACAS or FCAS opinion.40 The opinion must say the filing clearly presents all supporting information on the models and that the rating system meets the six requirements of the state’s rating statute.40 Those six are the grounds on which the director must disapprove a filing, such as inadequate premiums or rating on subjective factors.41

Nevada’s Bulletin 24-002, of August 2024, reads existing statute to require prior approval of any mathematical model used in underwriting or rating a personal property or casualty line.42 Calling it an underwriting model rather than a rating model does not change that. We did not survey every state’s rules, so whether these two are outliers or early cases of a trend is not known.

What goes into the first submission

Under the prior-approval model law, a filing that arrives without its supporting information waits. The commissioner may require the information, the waiting period starts only when it is furnished, and until then the filing is not complete.43 Enacted statutes set the periods. Nebraska’s gives the director 30 days, extendable by 30, and lets the director disapprove a filing whose requested information has not arrived within 60 days.44

Deciding what to send also means deciding what may become public. The prior-approval model law opens a filing and its supporting information to public inspection on receipt.45 Each state’s law decides what stays confidential, and the white paper warns that a filing “might become part of the public record.”46 Nebraska, for example, keeps specified filing documents out of the public record only if confidential treatment is granted under Neb. Rev. Stat. §84-712.05(3).47 When the NAIC takes part, the white paper adds, the state’s confidentiality and trade-secret protections still apply.46

Read in reverse, the team’s issue list makes a sound first submission. It says how many variables were screened and how the survivors were chosen. It shows fit on data the model never saw and gives each variable a rational explanation a reviewer can test against claim risk. Each selected factor that departs from its indication comes with a reason. A tree-based filing also carries a plot for every variable, the first-tree diagrams and the rating exhibits, with the full trees ready. A neural network filing carries the plots or images its checklist lists and the commentary on what drives the worst scores. A plan with a vendor model names the vendor and the sub-model’s version. Sending what the state asks for at the start keeps the clock from waiting, and material meant to stay private goes in with a request for confidential treatment. If the state has written its own list, as Nebraska has for expedited review, that list comes before any of this.

Footnotes

  1. White paper, Appendix B introduction, pp. 15 to 16 (information a regulator may need, by level); Manual, Appendix C introduction, p. 142 (documentation the NAIC team needs, by model type). Both documents are cited in full at notes 2 and 7. ↩

  2. NAIC Casualty Actuarial and Statistical (C) Task Force, “Regulatory Review of Predictive Models,” white paper dated September 9, 2020, adopted by the Task Force on September 15, 2020 and by the Property and Casualty Insurance (C) Committee on December 8, 2020, p. 2, §I Introduction: https://content.naic.org/sites/default/files/inline-files/9-15%20CASTF%20-%20Predictive%20Model%20White%20Paper%209-09-2020.pdf ↩

  3. White paper, p. 12, Summary: rate standards “typically included in the state rating laws.” The NAIC’s Property and Casualty Model Rating Law states the same standard in its prior-approval version, #1780, §4.A (July 2009 printing): https://content.naic.org/sites/default/files/GL1780.pdf, and its file-and-use version, #1775, §5.A (January 2010 printing): https://content.naic.org/sites/default/files/GL1775.pdf. A model law has effect only where a state enacts it; Nevada’s enacted wording is NRS 686B.050(1): https://www.leg.state.nv.us/nrs/nrs-686b.html ↩

  4. White paper, pp. 6 to 7, §VII Best Practices, item 1.b. ↩

  5. #1780, §4.C; #1775, §5.A(4)(b). Both also bar classification by race, creed, national origin or religion. A state’s own statute may bar further characteristics. ↩

  6. NAIC, Model Bulletin: Use of Artificial Intelligence Systems by Insurers (December 2023), Section 1, paragraph on the Property and Casualty Model Rating Law (#1780), and Section 3: insurers “are expected to develop, implement, and maintain a written program (an ‘AIS Program’) for the responsible use of AI Systems that make, or support decisions related to regulated insurance practices.” The bulletin has force only where a state has issued it: https://content.naic.org/sites/default/files/cmte-h-big-data-artificial-intelligence-wg-ai-model-bulletin.pdf.pdf ↩ ↩2

  7. NAIC Model Review Manual, adopted by the Task Force November 4, 2025, pp. 1 and 3. It bundles the agreement template (Appendix A), the white paper with its approved appendices (Appendix B), and model checklists, version 2025.2 (Appendix C). “The white paper’s list” in this article is the white paper’s own Appendix B, not the Manual’s. The Manual is “subject to revision by the Task Force” (p. 1): https://content.naic.org/sites/default/files/inline-files/NAIC%20Model%20Review%20Manual__%20adopted%20by%20CASTF%2011.04.25_0.pdf ↩

  8. Manual, pp. 3 to 4 (Background and Rate Review Support Services Agreement). The Agreement includes no charges for a review or for database access, and “A state is not obligated by the Agreement to ask for reviews or use the database.” ↩

  9. Manual, p. 4: “NAIC reports are confidential and for regulator use only”; access to the database and to NAIC or state reports “is restricted to regulators only.” The NAIC “does not distribute the list of states who have signed”; “States are allowed to share they have signed the Agreement.” ↩ ↩2

  10. Manual, pp. 5 to 6, Priorities 2: “Limit NAIC review work to top 20 insurance groups (by market size) and third-party vendors licensed or operating in 10 or more states”; “Exceptions may arise; as in the case of an innovative modeling approach initiated by a smaller company or a new product.” Other priorities cap each state’s monthly review requests. ↩ ↩2

  11. CASTF, 2025 Fall National Meeting summary (December 9, 2025), item 10: the team “does not approve or recommend approval of industry or third-party rate models,” and its product “consists of working papers to assist state insurance regulators in conducting their rate filing reviews”: https://content.naic.org/sites/default/files/national_meeting/2025-fanm-summary-c-castf.pdf ↩

  12. Manual, p. 3, Background, which lists what the NAIC does not do; the last item, in the Manual’s words, is “Separate analysis to determine any correlation with unlawful characteristics or to assess disparate impact.” Introduction, same page: “Compliance with a state’s laws and regulations will be performed by the state insurance regulator.” ↩ ↩2

  13. Manual, p. 6, Priorities 5 to 7: the NAIC sets aside two review slots for expected objection responses, “The aim will be to review objection responses within 30 days of receipt from the regulator,” and after a second assessment, if remaining questions “are not fully answered … the state will schedule a follow-up call with the company.” Appendix C introduction (p. 142): regulators can “communicate with insurers through rate filing objections.” ↩

  14. White paper, Appendix B introduction, pp. 15 to 16. Level 2 is “necessary to continue the review of all but the most basic models”; Level 3 is requested on concerns “that the model may not comply with state laws and/or regulations.” ↩ ↩2

  15. White paper, p. 15: the list is “lengthy but not exhaustive”; “Nor is every item on the list intended to be a requirement for every filing.” It is also “not intended to limit the authority of a regulator to request additional information.” For models other than GLMs, “some listed items might not apply” (p. 16), though the best practices “should be readily transferrable” (p. 5, §V). ↩

  16. White paper, Appendix B, item A.2.c (Level 1): “obtain the vendor name, as well as the name and version of the sub-model.” The item adds: “Sub-model SMEs may need to be brought into the conversation with regulators.” ↩

  17. Actuarial Standards Board, ASOP No. 56, Modeling (Doc. No. 195, December 2019, effective for work performed on or after October 1, 2020), §3.4: an actuary with “limited ability either to obtain information about the model or to understand the underlying workings of the model” should “disclose the extent of such reliance.” The standard applies to actuaries, not to insurers: https://www.actuarialstandardsboard.org/wp-content/uploads/2020/01/asop056_195.pdf ↩

  18. Manual, Appendix C, version 2025.2. The Manual’s own text (p. 5, Priorities 1) still says “The current Appendix C contains only GLM information”; the checklists for the other three model types are in the same PDF. ↩

  19. Version 2025.2 is the Appendix C of the Manual adopted on November 4, 2025 (see note 7). As of November 10, 2026, CASTF’s page links that Manual as “NAIC Model Review Manual_adopted by CASTF 11.4.2025”: https://content.naic.org/committees/c/casualty-actuarial-statistical-tf. The agenda for the CASTF meeting at the 2026 Fall National Meeting, November 16, 2026 (draft dated October 21, 2026), has no item on model review: https://content.naic.org/sites/default/files/national_meeting/111626-agenda-castf-fnm_2.pdf ↩

  20. Manual, Appendix C introduction, p. 142. Regulators who use the service asked the team for “a list of rate filing documentation needed for the NAIC to complete a full-scope rate model review.” Regulators “may evaluate these lists and determine the state’s needs” and “can share this list with insurers, revise the state’s rate filing checklists, or communicate with insurers through rate filing objections, when needed.” ↩ ↩2

  21. NAIC staff, “Golden Retriever Insurance Company, Golden Price Model v2.0,” sample filing posted April 27, 2026: https://content.naic.org/sites/default/files/call_materials/Golden%20Retriever%20Insurance%20Company%20Filing.pdf. Presented with NAIC staff slides, “Can Gen AI Review a Pricing GLM?” (April 28, 2026): https://content.naic.org/sites/default/files/call_materials/Can%20Gen%20AI%20Review%20a%20Pricing%20GLM.pdf ↩

  22. “Can Gen AI Review a Pricing GLM?”, slide 9 (“Example filing fulfilling most of the NAIC GLM Checklist”; “Non-Confidential”; “Includes both modeling errors and minor typos”), the issue list in the appendix, pp. 53 to 55, and slide 25, which lists five “High Priority Issues,” among them ”# of initial variables not discussed,” “Actual vs. Expected plots provided on training data” and “Proposed factors assume everyone bought Cold Play tickets.” ↩ ↩2

  23. Sample filing, p. 3, Introduction. ↩

  24. “Can Gen AI Review a Pricing GLM?”, appendix issue list, issues 1 (variables considered), 8 (fit plots on training data), 9 and 24 (the “Missing” level of the concert-ticket variable), 13 (statistical tests as explanation) and 21 (insurance score tiers 7 to 10). Issues 13 and 21 are not on slide 25’s high-priority list. ↩ ↩2 ↩3 ↩4 ↩5

  25. White paper, Appendix B, item B.1.g (Level 1): “comment on the use of automated feature selection algorithms” and “explain how potential overfitting … was addressed.” ↩

  26. White paper, Glossary of Terms (the white paper’s own Appendix C, not the Manual’s), “Rational Explanation”: “a plausible narrative connecting the variable and/or treatment in question with real-world circumstances or behaviors that contribute to the risk of insurance loss in a manner that is readily understandable to a consumer or other educated layperson. A ‘rational explanation’ does not require strict proof of causality but should establish a sufficient degree of confidence that the variable and/or treatment selected are not obscure, irrelevant, or arbitrary.” ↩

  27. Sample filing, data dictionary, whose rational explanation column the filing defines as a plausible account of how the variable indicates a difference in insurance risk. The team’s list (issues 14 to 16) calls the explanations for the guitar, seminar and concert variables “not intuitive” and each variable “questionable.” ↩

  28. White paper, Appendix B, item B.3.d (Level 3): “proving causation is neither practical nor expected.” The tree appendix’s item B.3.d uses the same words. Actuarial Standard of Practice No. 12, Risk Classification (Doc. No. 132, December 2005), §3.2.2, likewise does not require a cause-and-effect relationship: https://www.actuarialstandardsboard.org/wp-content/uploads/2014/02/asop012_132.pdf. The Actuarial Standards Board exposed a proposed revision in September 2023; we found no record of its adoption: https://www.actuarialstandardsboard.org/?p=25333 ↩ ↩2

  29. White paper, Appendix B, item B.4.a (Level 1). The white paper also says a complete filing “should include validation evidence for the rating plan, not just the statistical model” (pp. 4 to 5). ↩

  30. Sample filing, selections: the proposed factor for the “Missing” level “is set equal to the same rating factor for ‘Yes’”; “Insurance Score Tier – Selections were made based on current underwriting appetite.” ↩ ↩2

  31. White paper, Appendix B, items C.3.a and C.3.b (both Level 1). For tree-based models, the Manual’s checklist asks for documentation of deviations between indicated and selected factors “if applicable.” ↩

  32. CASTF, Appendix B-TREES, p. 18, which “should not be referenced in the review of other model types,” and items B.3.d and B.3.f (both Level 1): https://content.naic.org/sites/default/files/inline-files/CASTF%20Tree-based%20Model%20Appendix%20%28B-Trees%29.pdf. B.3.d asks for “at least one plot for every variable used in the model” and notes that partial dependence calculations “assume independence with other variables in the model.” NAIC staff slides date the appendix’s adoption to summer 2022: “Non-GLM Model Documentation” (April 23, 2024), https://content.naic.org/sites/default/files/call_materials/Non-GLM%20Documentation%20Presentation%202024%20Book%20Club.pdf ↩ ↩2

  33. Manual, Appendix C, version 2025.2, tree-based model checklist. Validation: “Decile plots (quantile plots with at least 10 buckets) for both state specific data and countrywide data, built on data not used for model training.” Implementation: tree diagrams for the first tree in each model and sample rating exhibits for 10 risks (Essential); “Complete documentation that would allow future audits of model predictions” (Sometimes Needed). ↩ ↩2

  34. Tree appendix, item C.9.d (Level 1), whose examples include every tree diagram and a table of every combination of risk characteristics; it also accepts the trees written as if-else logic statements. ↩

  35. The appendices CASTF lists on its page (see note 19) are those for GLMs (the white paper’s Appendix B), generalized additive models, regularized GLMs and trees. ↩

  36. Manual, Appendix C, version 2025.2, neural network checklist, Essential items: Modeling (“For Artificial Neural Networks (ANN), provide Shapley plots by variable”; “For Artificial Neural Networks (ANN), provide waterfall plots for” sets of records, starting with “10 records with the worst score and commentary on what is driving the score”; “For Convolutional Neural Networks (CNN), provide Grad-CAM images for” sets of images, starting with “10 images with the worst score and commentary on what is driving the score”); Validation (“A description whether the model predictions were compared to an independent report (example: roof image classification versus findings from an actual roof inspection)”); Implementation (whether the prediction “can be reproduced by a person using an objective criteria checklist”). The GLM, tree-based and penalized regression checklists each ask for “the rational explanation for each modeled variable that discusses why it would plausibly impact insurance risk as discussed in the CASTF white paper”; the neural network checklist has no such item. ↩ ↩2 ↩3

  37. White paper, Appendix B introduction and items B.3.d, B.4.a, C.7.g and C.8.a. C.7.g (Level 3): “Obtain a means to calculate the rate charged a consumer.” Its comment opens, “The filed rating plan should contain enough information for a regulator to be able to validate policy premium,” and adds that for a complex model or rating plan “a score or premium calculator via Excel or similar means would be ideal.” C.8.a (Level 1) asks for enough information “to verify that the rating plan’s manual, in fact, reflects the model output and any adjustments made to the model output.” ↩

  38. Tree appendix, items B.3.d, B.3.f and C.9.d; Manual, Appendix C, version 2025.2, tree-based checklist, Data (rational explanation), Validation (decile plots; interpretability plots) and Implementation (see note 33). ↩

  39. White paper, p. 5, §IV (“Best practices are not intended to create standards for filings that include predictive models”) and pp. 12 to 13, §IX. The Manual’s Appendix C introduction (p. 142) likewise says regulators “may evaluate these lists and determine the state’s needs.” ↩

  40. Nebraska Department of Insurance, Filing Requirements: Model-Dependent P&C SERFF Rate Filings, effective January 2, 2024, p. 1 (expedited-review conditions) and the item list, pp. 1 to 7: https://doi.nebraska.gov/sites/default/files/doc/Support%20document%20for%20Complex%20Model%20Review%20%28Final%29.pdf ↩ ↩2

  41. Neb. Rev. Stat. §44-7508(9), summarized: the director “shall disapprove a filing if” it would produce inadequate premiums; where the insurer has more than one rating system, fails to specify objective differences between risks to decide which applies; provides for optional commission differences for agents; is based on the race, creed, national origin, or religion of the insured; would violate the Unfair Discrimination Against Subjects of Abuse in Insurance Act; or discriminates between risks based on subjective factors: https://nebraskalegislature.gov/laws/statutes.php?statute=44-7508 ↩

  42. Nevada Division of Insurance, Bulletin 24-002 (August 27, 2024), under NRS 686B.110; it replaces Bulletin 17-001: https://doi.nv.gov/uploadedFiles/doinvgov/Content/News_Notices/Bulletins/24-002.pdf ↩

  43. #1780, §5.A(3); #1775, §6.A(3) has the same rule. ↩

  44. Neb. Rev. Stat. §44-7508(4) and (10) (see note 41). A filing not disapproved within the review period is deemed to meet the section’s requirements. ↩

  45. #1780, §5.A(2). ↩

  46. White paper, p. 6, §VI: “Each state determines the confidentiality of a rate filing and the supplemental material to the filing”; “When the NAIC or a third party enters the review process, the confidential, proprietary, and trade secret protections of the state on behalf of which a review is being performed will continue to apply.” ↩ ↩2

  47. Nebraska filing requirements (see note 40): specified documents “can be excluded from public record only if confidential treatment is granted consistent with Neb. Rev. Stat. § 84-712.05(3).” ↩

The Bottom Line

  • A state judges a model-based rate under its rating law, which NAIC guidance says it cannot change, and asks whether each variable relates to expected loss or expense.
  • In states that sign up and ask, the NAIC's rate model review team writes confidential reports for the state's reviewer; the filer meets its findings only through the state.
  • In the NAIC's fictional sample filing for a GLM, the team flagged an unstated count of screened variables, fit shown on training data, p-values offered as explanations, and factors with thin justification.
  • For tree-based models, NAIC guidance asks at the start for an interpretability plot and a rational explanation for every variable; none of it binds an insurer unless its state adopts it.
Engraved portrait of Simon Li

Written by

Simon Li · Founding Editor

I write InsureAI Wire and maintain its 51-jurisdiction tracker. Most of the work is reading: NAIC working group papers, state bulletins, bills, court filings, and public comment letters. Every claim on the site carries the document it came from, so you never have to take my word for it.

Contact or report a correction

Information aggregation and analysis, not legal advice. See our disclaimer.

Path complete

You have reached the evidence stage.

Return to the reading map to choose another route or business line.

Choose another reading path

Related reading

This week's briefing

What Courts Have Ruled So Far in Lawsuits Over Insurers' Claim Algorithms

Courts let contract and fair-housing claims over insurers' claim algorithms proceed, while Medicare preemption knocked out most state-law claims.

Keep tracking

  • The State Tracker

    Which states adopted the NAIC AI model bulletin, which wrote their own rules, and the primary source behind every row.

    28 + DC adopted 3 own rules 19 none

    Open the tracker
  • The Vendor Directory

    Which AI vendors sell to insurers, who owns each of them now, and the source behind every line.

    Browse insurance AI vendors
  • The Wire

    Short items on what just changed, with the filing itself one click away.

    Read the Wire