Actuarial Justification for AI Rating Variables
Actuarial justification for AI rating variables means demonstrating, with documented evidence, that a variable derived from a machine-learning model has a logical and actuarially sound relationship to risk of loss, not merely a statistical correlation, and that the model producing it is governed by an accountable, auditable oversight process.
What Actuarial Justification Requires
The baseline legal standard applied by state regulators, drawn from the NAIC Unfair Trade Practices Act model law, is that insurance rates must not be excessive, inadequate, or unfairly discriminatory. When a rating variable originates from an AI or machine-learning process, regulators apply this standard through an additional lens: does the variable have a demonstrable relationship to risk of loss, and can the insurer produce documentation showing how that relationship was established and is being maintained.
The NAIC's Regulatory Review of Predictive Models framework directs reviewers to look past model output and ask whether a variable has a logical, actuarially sound connection to loss experience, rather than accepting a statistical correlation identified by the model as sufficient on its own. This shifts the burden onto actuarial and compliance teams to produce a rationale that goes beyond model performance metrics.
The Current Regulatory Framework
The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted and issued by multiple states, requires insurers to maintain a documented AI governance program and risk management framework, and to maintain an inventory of AI systems in use, including those affecting rating, with documented purpose, inputs, and oversight mechanisms for each system.
Colorado's Division of Insurance, acting under authority granted by SB21-169, has gone further for life insurers by requiring quantitative testing of external consumer data and algorithms for unfair discrimination tied to protected classes, and by requiring insurers to maintain documentation of testing methodology, results, and remediation steps that must be made available to the Division upon request.
NCOIL has advanced a model act that parallels the NAIC bulletin's governance and accountability concepts, reinforcing that this is a converging multistate expectation rather than an isolated requirement in one jurisdiction.
Regulatory Anchors for AI Rating Variable Review
These frameworks shape how reviewers assess data sources, variable selection, validation methodology, and accountability for rating outcomes.
NAIC AI Model Bulletin
Requires documented AI governance programs, system inventories, and accountability for outcomes affecting rating and underwriting.
Colorado SB21-169 Regulations
Mandates quantitative unfair discrimination testing of algorithms and predictive models, with documentation available on request.
NAIC Predictive Model Review
Provides a question framework regulators use to assess data sources, variable selection, and validation methodology.
NCOIL Model Act
Parallels NAIC governance and accountability concepts at the state legislative level.
Correlation Versus Actuarial Support
Regulators distinguish between two categories of rating variables produced by AI models. The first has an identifiable, plausible risk-based mechanism, similar to variables long accepted in traditional generalized linear models. The second is identified purely through machine-learning feature importance or predictive lift, without an accompanying explanation of why the variable should relate to loss.
Variables in the second category are treated by regulators as higher risk for challenge. The model type matters here: gradient boosting and other less interpretable techniques increase the burden of justification compared to a GLM, because the mechanism connecting the variable to loss is harder to articulate from the model itself. Insurers should expect to pair statistical validation, such as loss ratio relationships and predictive lift, with a written narrative rationale explaining a plausible causal or risk-based mechanism for each variable, not just its statistical performance.
Sustaining Justification After Approval
Regulatory guidance treats actuarial justification as an ongoing obligation rather than a one-time deliverable. Insurers are expected to conduct periodic re-validation and update documentation when models are retrained or when the composition of rating variables changes. Ongoing monitoring should track model drift, changes in variable importance, and outcome disparities that could undermine the original justification provided at approval.
Governance frameworks should include escalation procedures for situations where monitoring detects disparate impact or a weakening correlation between a rating variable and actual loss experience. Because Colorado's prescriptive testing regime differs from the more general unfairly discriminatory standard applied in other states, multistate insurers need a governance approach designed to meet the most stringent applicable requirement rather than the lowest common denominator.
Runtime monitoring and audit logging of the underlying AI systems can support this ongoing obligation by providing a continuous record of model behavior and variable-level changes that actuarial and compliance teams can reference when re-validating justification, rather than relying solely on point-in-time model documentation.
Documentation Elements Regulators Expect
A defensible package typically includes inventory, lineage, testing, and oversight artifacts that reviewers can request on short notice.
- A centralized AI or model inventory documenting purpose, inputs, and rating variables tied to each model
- Variable definitions, data lineage, and training and validation data splits
- Performance metrics broken out by demographic subgroup where feasible
- Disparate impact or proxy discrimination testing linking rating variables to protected class membership
- Testing methodology, results, and remediation steps retained and available on request
- Documented board and senior management oversight accountability for AI systems affecting policyholders
Frequently Asked Questions
Does actuarial justification apply only to newly introduced rating variables?
No. Regulatory guidance treats justification as an ongoing obligation. When a model is retrained or a rating variable's composition changes, insurers are expected to re-validate and update the justification documentation, not rely on the original approval indefinitely.
How does documentation burden differ between GLM and machine-learning models?
Interpretability affects the burden of proof. GLM variables often have a more directly explainable structure, while machine-learning-derived variables, such as those from gradient boosting, typically require a supplemental narrative rationale connecting the variable to risk of loss.
Is Colorado's testing regulation relevant to insurers outside Colorado?
Colorado's algorithm and predictive model testing requirements are specific to that state's authority under SB21-169, but given multistate regulatory alignment trends, they are commonly referenced as a practical benchmark for documentation rigor even by insurers not directly subject to them.
Strengthen Governance Over AI-Derived Rating Variables
Runtime governance and audit logging can provide the continuous oversight record that supports ongoing actuarial justification as AI models and rating variables evolve.
Request a Demo