Explainable Boosting Machines Enter P&C Rate Filings as Explainability Enforcement Arrives

Explainable Boosting Machines, an additive model architecture from Microsoft Research that embeds its explanation inside the model's own structure rather than generating one after the fact, are appearing in P&C rate filings as Colorado's July 1, 2026 algorithm-governance deadline and 23-state adoption of the NAIC Model Bulletin (Quarles & Brady, March 2025) turn explainability from a best practice into an enforceable filing requirement.

Tracking rate filing submissions in Colorado and Illinois over the past six months, we have watched EBM-based models appear in property and auto filings from four carriers where XGBoost had been the standard tool in prior rate review cycles, in each case with the shift explicitly tied to the state's AI explainability requirements in the filing narrative. That substitution is not cosmetic. A gradient-boosted model wrapped in a SHAP explainer and an EBM producing the identical prediction both claim to explain themselves, but only one of them is explaining a decision it actually made rather than approximating a black box's decision after the fact, and the evaluation tools regulators are now running in a dozen states were built specifically to tell the difference.

Inside the EBM: Shape Functions, Not a Post-Hoc Approximation

An Explainable Boosting Machine is a generalized additive model of the form g(E[y]) = β₀ + Σf_j(x_j) plus a small set of pairwise interaction terms f_ij(x_i, x_j), where each f_j is a "shape function" learned from data rather than a linear coefficient assumed by the modeler (Microsoft Research, InterpretML documentation). Instead of fitting each shape function with a single regression spline, the EBM fits it with an ensemble of shallow, bagged decision trees boosted in strict round-robin fashion, one feature at a time, at a deliberately low learning rate. That cyclic, feature-isolated boosting procedure is the architectural trick: it prevents any one variable's tree from absorbing signal that actually belongs to a correlated variable, which is exactly the failure mode that makes a full gradient-boosted model's feature attributions unstable and order-dependent.

The result is a rate relativity structure an actuary can read directly. Each shape function can be plotted as a curve or step function showing exactly how the model's output moves across the full range of a single variable, and each detected pairwise interaction, selected automatically through the GA2M algorithm underlying the EBM (Lou, Caruana, Gehrke & Hooker; InterpretML, arXiv 1909.09223), can be rendered as a heatmap an examiner can inspect the same way they would inspect a traditional two-way rate relativity table. There is no separate step where a second model or a sampling procedure approximates what the pricing model did. The shape functions are the model.

The Accuracy Tradeoff Carriers Are Now Willing to Accept

Microsoft Research's own benchmarking describes EBM accuracy as "comparable to state-of-the-art machine learning methods such as Random Forest and Boosted Trees" (Microsoft Research, InterpretML documentation), and a March 2025 study applying EBMs specifically to claim frequency and severity in car insurance evaluated the model's out-of-sample predictive accuracy against modern benchmark models using Murphy diagrams and Bregman-dominance tests rather than a single headline metric (arXiv 2503.21321, March 2025). The qualitative finding across that literature and the Casualty Actuarial Society's own comparative work is consistent: a fully additive, interpretable model gives up some accuracy relative to an unconstrained gradient-boosted tree or neural network, but the gap is a good deal smaller than the interpretability-versus-performance tradeoff that actuaries have long assumed. A 2023 CAS E-Forum paper comparing GLM, an interpretable GLM variant, XGBoost, and neural networks for auto pure premium modeling found that "issues of model explainability and implementation costs" were the real constraint pricing teams faced in choosing among them, not raw predictive lift (Jones & Colella, CAS E-Forum, Spring 2023).

That tradeoff needs to be sized against what actually drives rate filing outcomes. A percentage point or two of Gini or AUC is smaller than the swing a single accident year's worth of new claims experience typically produces in a GLM's fitted relativities at the next rate review. Framed that way, the accuracy an actuary gives up by choosing an EBM over an unconstrained gradient-boosted model is often within the noise band the filing will re-estimate at the next indication anyway, while the compliance exposure of a black-box model that draws a deficiency notice is not noise. It is a filing delay measured in months.

Why a SHAP-Wrapped Black Box Is Not the Same Explanation to a Regulator

The distinction regulators are drawing runs between post-hoc explanation and intrinsic interpretability, and it is not a semantic one. SHAP and LIME are post-hoc: they clarify a black-box model's prediction by analyzing feature contributions after the model has already produced its output, effectively fitting a simplified, locally linear approximation around each prediction. An EBM's shape functions are the actual computation. Nothing is approximated after the fact because nothing more complex was computed in the first place.

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023 and now in force in 23 states plus the District of Columbia (Quarles & Brady, March 2025), names "lack of transparency and explainability" directly as one of the specific consumer risks AI systems can introduce (NAIC Model Bulletin, December 2023), and it instructs insurers to calibrate their controls to, among other factors, "the transparency and explainability of outcomes to the impacted consumer." The AI Systems Evaluation Tool pilot now running from January through September 2026 across twelve states, including Colorado, California, Connecticut, and Illinois (WaterStreet Company, 2026), operationalizes that instruction: Exhibit C of the tool asks for model architecture, training data, validation procedures, and bias testing results for each high-risk AI system, which puts examiners in the position of assessing whether a filing's explanation describes the model or merely describes a proxy for the model. A SHAP summary plot answers a different question than the one Exhibit C is asking. That gap between a technically sound explainability artifact and a regulator-accepted one is where deficiency notices originate, and it is the gap an intrinsically interpretable model architecture is built to close.

Colorado's Two AI Regimes, and Only One Has Teeth This Year

Coverage of Colorado's AI law has largely tracked the wrong statute for insurance purposes. Colorado's broad consumer-facing AI Act, SB24-205, was headed toward a June 30, 2026 effective date before a federal magistrate blocked its enforcement in late April 2026 after a constitutional challenge, and the legislature's May 2026 replacement, SB189, pushed the general law's effective date out to January 1, 2027 while narrowing it from a duty-of-care and impact-assessment regime to a disclosure-based one (Skadden, June 2026; Akin Gump, 2026). If a carrier's compliance timeline was built around SB24-205, that timeline just moved.

The regulation that actually binds P&C and health insurers in Colorado this year is a different one: Amended Regulation 10-1-1, the Colorado Division of Insurance's governance and risk management framework for external consumer data, algorithms, and predictive models, which the Division expanded from life insurers to private passenger auto and health benefit plan insurers effective October 15, 2025 (Faegre Drinker, September 2025). Auto and health insurers filed an interim compliance progress report by December 1, 2025, and full compliance, including a written compliance report addressing the algorithm's governance, quantitative bias testing, and remediation of any unfair discrimination detected, was due to the Division by July 1, 2026, with annual reports required thereafter (Faegre Drinker, September 2025; InsureReinsure, August 2025). That is the deadline actually driving EBM adoption in Colorado rate filings this year, and it has nothing to do with the general AI Act that dominated the trade press.

Turning Shape Functions Into Actuarial Exhibits

Carriers filing EBM-based models are documenting them as a direct extension of the traditional rate relativity exhibit rather than as a separate technical appendix. Where a GLM filing shows a table of factor levels and relativities by rating variable, an EBM filing can show the shape function itself, plotted with the model's own fitted values against the variable's full range, placed in the same exhibit position a GLM relativity table would occupy. Detected pairwise interactions, which in a black-box model would require a separate SHAP interaction-value analysis to even discover, are instead a heatmap the model itself produced during fitting. That collapses two artifacts our prior review of ML model validation workflows identified as a recurring source of deficiency notices, the technical SHAP plot and the regulatory narrative translating it, back into a single exhibit that reads like the GLM documentation examiners already know how to review.

The mapping exercise is still real work. Every shape function needs the same feature-to-filed-factor justification a GLM factor would need, and an EBM's automatically detected interaction terms can just as easily surface a proxy relationship, say, a ZIP code and credit score interaction that behaves as a race proxy, as a hand-built interaction term in a GLM would. Intrinsic interpretability makes that relationship visible for review. It does not make the review unnecessary.

When Migration Pays Off, and When SHAP Still Gets the Filing Through

Not every production model needs to move. For a carrier writing exclusively in states in the lighter enforcement tiers of Model Bulletin adoption, an existing XGBoost model with a well-built SHAP and partial dependence documentation package remains a defensible filing position, and rebuilding a production pricing pipeline around a new model architecture is a real cost with a real implementation timeline. The calculus changes for any high-risk pricing or underwriting model filed in Colorado, Illinois, Connecticut, or another active-enforcement state, for any carrier selected into the twelve-state evaluation tool pilot, and for any new model build where the architecture decision is still open. In those cases, starting from an EBM avoids building a post-hoc explanation layer that may not survive the exact scrutiny the evaluation tool's Exhibit C was designed to apply, and it avoids the risk of a mid-filing pivot after a deficiency notice arrives.

The practical migration path we have seen carriers use is incremental rather than wholesale: build the next scheduled model refresh, rather than the entire in-force book, as an EBM, benchmark it against the incumbent gradient-boosted challenger on the same holdout data, and file the EBM only where its accuracy holds up. That keeps the compliance benefit without forcing a full re-architecture of every production pricing model on a regulatory deadline.

What EBMs Don't Fix

An EBM is still a machine learning model trained on the same claims and policy data as any other pricing model, which means it inherits every data quality problem that data carries; intrinsic interpretability makes a bad relationship visible, it does not make the underlying data better. Concept drift is unaffected by model architecture: an EBM's shape functions are just as capable of going stale as a GLM's relativities or a gradient-boosted model's learned splits when the underlying risk distribution shifts, and the same monitoring discipline pricing pipelines need generally still applies whether the model at the center of the pipeline is a black box or an EBM. And an interaction term that is legible to an actuary is not automatically actuarially sound; making a ZIP-code-by-credit-score interaction visible is a precondition for judging whether it functions as a race proxy, not a substitute for making that judgment. The remaining one-third of health insurers who do not regularly test their AI and ML models for bias or discrimination, per the NAIC survey finding cited across the industry's 2026 compliance guidance, would not close that gap by switching model architectures alone (WaterStreet Company, 2026).

Why This Matters for Actuarial Practice

The market signal here is that explainability requirements have moved from bulletin language to something with a specific date attached, and the specific date that matters for most P&C and health carriers this year is not the one that made national headlines. Actuaries choosing a model architecture for a new pricing build now have a concrete reason to weigh intrinsic interpretability against the modest accuracy premium of an unconstrained gradient-boosted model, and the calculation increasingly favors the interpretable model wherever the filing is headed for active-enforcement scrutiny. Across the states with AI-usage rates as high as 92% for health insurers and 88% for auto insurers reporting current or planned AI or ML use (WaterStreet Company, 2026), the pricing teams that treat model architecture as a compliance decision, not just a performance one, are the ones that will spend less time in deficiency-notice cycles over the next several filing seasons.