What is ASOP No. 12 and why is it being revised?

ASOP No. 12 (Risk Classification) governs how actuaries design, review, and modify risk classification systems used in insurance and other financial security programs. First adopted in 1989 and last revised in 2005, the standard predates the widespread use of predictive modeling, machine learning, telematics, and non-traditional data sources in risk classification. The Actuarial Standards Board is revising it to address algorithmic classification methods, proxy discrimination, unintended bias, and the documentation requirements created by modern multivariate modeling techniques.

How does the ASOP No. 12 revision address AI and machine learning in risk classification?

The revision adds new sections on Data and Model (Section 3.2.2) and Multivariate Effects (Section 3.2.4) that the 2005 standard lacked entirely. The 2005 version contains zero references to the word 'model.' The revised standard introduces a broader 'Risk Classification Framework' concept replacing the narrower 'Risk Classification System,' and adds affirmative obligations for actuaries to consider unintended bias (Section 3.4) and protected class impacts (Section 3.5) when using ML-based classification methods.

When will the revised ASOP No. 12 take effect?

The ASB released a first exposure draft in January 2024 (57 comment letters received) and approved a second exposure draft in October 2024 with comments due by March 2025. With second-round comments now under review and ASB meetings scheduled through 2026, a final standard could be adopted in mid-to-late 2026. Once adopted, the standard becomes effective six months after the adoption date.

ASOP No. 12 Gets Its First AI-Era Rewrite After 21 Years

The Actuarial Standards Board is revising ASOP No. 12, the risk classification standard that has governed how actuaries group risks since 1989 and was last updated in 2005. That revision predated predictive modeling, telematics, machine learning, and the explosion of non-traditional data sources that now define modern risk classification. With 57 comment letters filed on the first exposure draft and a final standard potentially arriving in late 2026, the revision addresses algorithmic classification methods, proxy discrimination, and documentation requirements that the pre-AI standard never contemplated.

From advising on rate filings that rely heavily on ML-based risk classification, the gap between what models can do and what the 2005 standard contemplated has become a daily challenge that this revision aims to close. ASOP No. 12 was last revised when GLMs were cutting-edge for most pricing teams, when telematics was a concept rather than a product line, and when "non-traditional data" meant credit scores rather than satellite imagery, IoT sensor feeds, or social media-derived behavioral indicators. The 2005 standard contains zero references to the word "model." That single fact captures the distance between the standard and current practice better than any commentary.

The Actuarial Standards Board released a first exposure draft in January 2024, received 57 comment letters by the May 2024 deadline, approved a second exposure draft in October 2024, and is now reviewing second-round comments submitted through March 2025. The task force, chaired by Brian J. Mullen, includes members from across practice areas: Steve D. Armstrong, Keith A. Passwater, Andrew D. Colton, Tony R. Phipps, Brian A. Fannin, Erica S. Rode, and Gordon K. Hay. A final standard could be adopted at the ASB's September or December 2026 meeting, with an effective date six months after adoption.

This revision lands in the middle of the busiest standard-setting cycle the profession has seen in a generation, with roughly 20 of 52 active ASOPs under simultaneous revision. But ASOP No. 12 is unique among them: it is the standard most directly confronting how AI and ML have transformed the actuarial work product.

21 yrs

Since ASOP No. 12 was last revised, predating ML, telematics, and non-traditional data in risk classification

Comment letters filed on the first exposure draft by organizations including the NAIC, AAA, Allstate, Travelers, and USAA

~20

ASOPs under active revision in the 2025-2026 cycle, the busiest in the profession's history

What the 2005 Standard Got Right, and What Changed

ASOP No. 12 was originally adopted in 1989 under the title "Concerning Risk Classification." The 2005 revision broadened it to "Risk Classification (for All Practice Areas)" and established the conceptual framework that actuaries have used since: a risk classification system should reflect expected outcomes, minimize adverse selection, be practical and cost-effective, comply with applicable law, and respect industry practices. A minor 2011 update regularized deviation language but left the substance untouched.

The 2005 framework was well suited to its era. Risk classification in 2005 primarily involved selecting rating variables based on actuarial judgment and statistical testing, assigning risks to classes, and demonstrating that expected outcomes differed meaningfully across classes. The analysis was largely univariate. An actuary evaluating a risk classification system would examine each variable individually, test for statistical significance, and document the rationale for inclusion.

Three developments since 2005 have made that framework insufficient.

Multivariate modeling became the default. Modern pricing teams routinely build GLMs, gradient-boosted trees, random forests, and neural networks that simultaneously evaluate dozens or hundreds of risk characteristics. These models capture interaction effects and nonlinear relationships that univariate analysis cannot detect. The 2005 standard's implicit assumption that actuaries examine risk characteristics one at a time does not describe how classification work is actually performed in 2026.

As the task force chair stated in the exposure draft transmittal: "Each of the practice areas has seen tremendous change over the past 20 years in the ways that risk classification work is done." That understates the shift. In P&C personal lines, nearly every major carrier now uses ML-based territory models, telematics-driven driving behavior scores, and algorithmically derived risk segments. The standard governing this work was written for a different discipline.

Non-traditional data sources proliferated. Telematics data from connected vehicles records driving behavior at sub-second intervals. IoT sensors monitor commercial property conditions in real time. Satellite imagery informs wildfire and flood exposure assessments. External consumer data sources provide behavioral indicators that correlate with, but do not directly measure, insurance risk. Each of these creates classification possibilities that the 2005 framers did not contemplate, along with fairness and privacy challenges they could not have anticipated.

Regulatory expectations around algorithmic fairness sharpened. The NAIC Model Bulletin on AI use by insurers, adopted in December 2023 and now in force across more than half the states, establishes governance, bias testing, and transparency expectations for algorithmic decision-making. Colorado's SB 21-169 requires annual bias attestations for algorithms and predictive models. State DOIs increasingly ask rate filing actuaries to demonstrate that ML-based risk classifications do not produce unfair discrimination against protected classes. The 2005 ASOP provides no guidance on how to meet these expectations.

Key Structural Changes in the Revision

The exposure draft signals a conceptual evolution, not just a technical update. Several changes restructure how actuaries should think about classification work.

Terminology shift: "System" becomes "Framework." The revision replaces "Risk Classification System" with "Risk Classification Framework" throughout. This is more than editorial preference. "System" connotes a defined mechanism with fixed inputs and outputs. "Framework" acknowledges that modern classification approaches involve interconnected models, data pipelines, and decision processes that may adapt over time. The broader term accommodates ML-based classification where the "system" is a trained model whose internal logic is not specified by the actuary but learned from data.

New Section 3.2.2: Data and Model. The 2005 standard had no section addressing models. The revision adds explicit guidance on documenting data sources, model selection rationale, and model limitations. For actuaries using gradient-boosted models, neural networks, or ensemble methods for risk classification, this section creates the first ASOP-level obligation to document the modeling choices that produce their classification outputs. The connection to ASOP No. 56's existing modeling requirements is direct: where ASOP No. 56 governs modeling generally, the new ASOP No. 12 section governs modeling specifically in the risk classification context.

New Section 3.2.4: Multivariate Effects. This section requires actuaries to consider how risk characteristics interact in multivariate contexts. The 2005 standard was structured around evaluating individual risk characteristics and their relationship to expected outcomes. The revision acknowledges that "technological advances enable new forms of analysis and better understanding of correlations" and requires documentation of interaction effects, correlation analysis, and multivariate model behavior. For a pricing actuary who builds a 200-variable GBM, this section transforms what was an implicit best practice into an explicit professional obligation.

New Section 3.4: Potential for Unintended Bias. The revision introduces a new definition of "unintended bias" as impacts or outcomes on specific risk subjects resulting from a risk classification framework that is not intentionally designed to produce such effects. This creates an affirmative obligation: actuaries must consider whether their classification frameworks produce disparate impacts, even when those impacts are not deliberately designed. For ML models that may produce proxy discrimination through correlated variables, this section requires active investigation rather than passive acceptance.

New Section 3.5: Protected Classes. Entirely new to the standard, this section addresses how actuaries should consider protected class impacts in risk classification. This generated substantial debate among commenters, who questioned the extent to which a professional standard should address what are fundamentally legal and regulatory matters. The section reflects the reality that state DOIs increasingly expect actuaries to affirmatively demonstrate that their classification methods do not unfairly discriminate, and that ASOP-level guidance on how to meet these expectations is overdue.

Broadened Risk Measure concept. The revision replaces the 2005 emphasis on "expected outcomes" with a broader "Risk Measure" definition. This encourages actuaries to consider multiple aspects of loss distributions beyond mean values, including tail risk measures, conditional tail expectations, and distributional modeling. The shift acknowledges that modern risk classification often aims to differentiate risks along the entire loss distribution, not just at the mean.

Dimension	ASOP No. 12 (2005)	Proposed Revision (2024-2026)
Terminology	Risk Classification System	Risk Classification Framework
Model guidance	No section; zero references to "model"	New Section 3.2.2 (Data and Model) with documentation requirements
Variable analysis	Implicitly univariate	New Section 3.2.4 (Multivariate Effects) requires interaction analysis
Bias consideration	Not addressed	New Section 3.4 (Unintended Bias) with affirmative assessment obligation
Protected classes	Not addressed	New Section 3.5 (Protected Classes) requires consideration of disparate impacts
Risk measurement	Expected outcomes focus	Broader "Risk Measure" including distributional and tail measures
Disclosure requirements	General documentation	Nine specific disclosure categories in restructured Section 4

What the 57 Comment Letters Reveal

The first exposure draft drew 57 comment letters between January and May 2024, an unusually high volume that reflects the standard's broad applicability and the intensity of industry interest in AI-related professional guidance. Commenters ranged from individual actuaries to major organizations: the NAIC Casualty Actuarial and Statistical Task Force, the American Academy of Actuaries (whose Casualty Committee submitted a 20-page letter), the CAS Professionalism Education Working Group, Allstate, Travelers, USAA, Zurich, Verisk Analytics, The Hartford, New York Life, Oliver Wyman, WTW, and the National Association of Mutual Insurance Companies, among others.

Five themes dominated the comment period.

The "unintended bias" definition drew the sharpest debate. The proposed definition (impacts resulting from a framework "not intentionally designed to result in such impacts") generated concern about precision and scope. Several commenters argued the definition is too broad, capturing any classification outcome that was not explicitly planned for, rather than outcomes that are problematic. Others questioned how "unintended bias" relates to the legal concept of "unfair discrimination" that regulators enforce. The interaction between a professional standard's guidance and a state DOI's enforcement authority is genuinely complex: an actuary can follow ASOP-level guidance on assessing unintended bias while still facing regulatory challenge under a different legal definition of unfair discrimination.

Proxy discrimination is the hardest practical problem. The CAS specifically recommended aligning ASOP No. 12's treatment of proxy discrimination with ASOPs 53 and 56 to clarify the boundary between grouping risks into classes (ASOP 12's domain) and estimating future costs for each class (ASOP 53's domain). The challenge is fundamental to ML-based classification: removing a prohibited characteristic from a model's feature set does not remove its influence if correlated variables remain. ZIP code, credit score, vehicle type, and dozens of other permissible variables can serve as proxies for race, income, or other protected characteristics. As the AAA's algorithmic accountability work has documented, machine learning "tends to produce the same results as intentional proxy discrimination" even when the prohibited variable is excluded from the model. The revised standard needs to give actuaries workable guidance on a problem that ML makes structurally difficult.

Documentation burden concerned carriers and consulting firms. Section 3.7's requirement that documentation be sufficient for "another actuary qualified in the same practice area" to "assess the reasonableness of the actuary's work" sets a reproducibility standard that goes beyond current practice at many organizations. For a team building a 500-variable gradient-boosted model with multiple feature engineering steps, hyperparameter tuning rounds, and model selection criteria, the documentation required to meet this standard is substantial. Commenters from carriers and consulting firms raised practical questions about scope: does this standard apply to every intermediate model version or only the final deployed model? Does it require documenting every data preprocessing decision or only material ones?

The boundary between ASOP No. 12 and ASOP No. 56 needs clarification. Multiple commenters noted that the new Data and Model section in ASOP No. 12 overlaps with ASOP No. 56's modeling guidance. An actuary building an ML model for risk classification must comply with both standards. If ASOP No. 56 already requires documenting model appropriateness, data quality, and validation results, what incremental obligations does ASOP No. 12's new section create? The task force will need to articulate a clear boundary: ASOP No. 56 governs the model as a modeling artifact, while ASOP No. 12 governs the model's use as a risk classification instrument.

Protected class guidance generated philosophical disagreement. Some commenters supported Section 3.5 as necessary guidance for a profession increasingly asked to evaluate fairness. Others objected that a professional standard should not address what are fundamentally legal questions about protected class definitions, disparate impact thresholds, and regulatory enforcement standards. The NAIC Casualty Actuarial and Statistical Task Force's comment reflected the regulatory perspective: state DOIs need actuaries who can demonstrate compliance with protected class requirements, and professional guidance on how to do so strengthens the regulatory ecosystem. The American Society of Enrolled Actuaries and some carrier commenters pushed back, arguing that the standard risks creating professional obligations that exceed or conflict with applicable law.

The ASOP No. 56 Intersection

ASOP No. 56 (Modeling), effective since October 2020, is the closest existing standard to a comprehensive model governance framework for actuaries. It requires evaluating model appropriateness, assessing data quality, performing model testing, ensuring governance and controls, and disclosing material limitations. For actuaries using AI and ML models, ASOP No. 56 already establishes the professional floor for model validation and documentation.

The revised ASOP No. 12 builds on that floor rather than replacing it. Where ASOP No. 56 asks whether the model is sound as a modeling artifact, the revised ASOP No. 12 asks whether the model produces a sound risk classification as a classification instrument. The distinction matters operationally: a gradient-boosted model can pass ASOP No. 56's validation requirements (appropriate data, reasonable assumptions, acceptable predictive performance) while still producing risk classifications that embed unintended bias or fail to satisfy regulatory fairness requirements. ASOP No. 12's new sections on unintended bias and protected classes address this gap.

The AAA's October 2024 guidance paper on generative AI professionalism reinforced ASOP No. 56's relevance by confirming that "GenAI is a model, and ASOP No. 56 provides guidance" for it. Actuaries cannot accept AI-generated outputs without validation. For risk classification work that incorporates generative AI components (for example, using an LLM to extract risk-relevant information from unstructured submission data), both ASOP No. 56 and the revised ASOP No. 12 will apply. The actuary must validate the model under ASOP No. 56 and evaluate the resulting classification under ASOP No. 12.

The 2026 standard-setting cycle amplifies this intersection. ASOP No. 30's profit provision rewrite, ASOP No. 41's third exposure draft on actuarial communications (approved March 2025), and ASOP No. 20's expansion to all P&C cash flows are all moving simultaneously. An actuary filing an ML-based rate indication in late 2026 may need to comply with newly effective or recently revised versions of ASOPs 12, 20, 30, 41, and 56. The cumulative documentation burden is significant, and the interdependencies between standards create compliance complexity that no single standard addresses.

Regulatory Overlay: State Filings and the NAIC Pilot

The ASOP No. 12 revision does not exist in a regulatory vacuum. State DOIs are simultaneously developing their own frameworks for evaluating AI-driven risk classification, and the revised standard will shape how actuaries document compliance with those frameworks.

The NAIC Model Bulletin on AI use by insurers, adopted in December 2023 and now in force in over half the states, establishes governance, bias testing, documentation, and third-party oversight expectations for carriers using AI in any operational context, including risk classification. The 12-state AI evaluation tool pilot, running from March through September 2026, provides regulators with structured examination frameworks covering AI usage quantification, governance risk assessment, high-risk system documentation, and data practices. When a state examiner uses the evaluation tool to review a carrier's ML-based risk classification system, the actuary's compliance with the revised ASOP No. 12 will be part of what the examiner evaluates.

The connection between professional standards and regulatory requirements runs in both directions. A revised ASOP No. 12 that requires actuaries to assess unintended bias and consider protected class impacts strengthens regulators' ability to hold carriers accountable: they can point to the actuarial standard as defining the profession's own expectations for fairness analysis. Conversely, regulatory developments like Colorado's annual bias attestation requirement and the NAIC's evolving AI governance expectations create practical context for how ASOP No. 12's new sections will be applied.

For actuaries preparing ML-augmented rate filings, the revised standard will likely reshape documentation requirements in several specific ways. Feature-to-factor mapping, where the actuary translates ML model features into traditional rating factor equivalents for regulatory review, will need to address multivariate effects under the new Section 3.2.4. Bias testing documentation, already expected under the NAIC Model Bulletin and required by Colorado, will need to satisfy the new Section 3.4's unintended bias assessment obligation. SHAP values and other explainability outputs, increasingly standard in rate filing support, will need to be documented with the rigor that Section 3.7's reproducibility standard demands.

The practical implication is that rate filing actuaries should not wait for the final standard to adopt these practices. The regulatory expectation already exists through the NAIC Model Bulletin and state-level requirements. The revised ASOP No. 12 will formalize what many DOIs already expect, and actuaries who build their documentation practices around the exposure draft's structure will be better positioned when the final standard takes effect.

What Practicing Actuaries Need to Prepare

The revised ASOP No. 12 creates several new documentation and analysis obligations that practicing actuaries should begin preparing for now, even before the final standard is adopted.

Nine required disclosures. The restructured Section 4.1 consolidates all disclosure requirements into nine categories: intended purpose of the risk classification framework, data and models used, risk measures employed, risk characteristics considered, adverse selection impact assessment, external influences (regulatory, industry, business environment), effectiveness evaluation, changes made to existing frameworks, and reliance on other parties. For actuaries accustomed to the 2005 standard's more general disclosure expectations, this specificity represents a meaningful increase in documentation scope.

The reproducibility standard. Section 3.7's requirement that documentation enable another qualified actuary to assess the reasonableness of the work is the single most operationally significant change. For traditional rating plans, this was largely achievable through existing documentation practices. For an ML-based classification built through iterative feature engineering, hyperparameter optimization, and model ensembling, reproducibility requires documenting the full model development pipeline: data preprocessing steps, feature selection criteria, model architecture decisions, training procedures, validation results, and the rationale connecting each choice to the classification's intended purpose.

Unintended bias assessment becomes mandatory. Under Section 3.4, actuaries must affirmatively consider whether their classification frameworks produce unintended impacts on specific risk subjects. This is not a safe harbor provision; it does not tell actuaries what outcome is acceptable. It creates an obligation to investigate, assess, and document. For carriers subject to the NAIC Model Bulletin or Colorado's bias attestation requirements, this aligns with existing regulatory expectations. For carriers in states without specific AI governance requirements, it creates a new professional obligation independent of regulatory mandate.

Protected class analysis enters the professional standard. Section 3.5 requires actuaries to consider protected class impacts in their classification work. The scope and specificity of this obligation will be clearer in the final standard, but the direction is unmistakable: the profession is moving toward making fairness analysis an explicit component of actuarial practice, not just a regulatory compliance exercise. Actuaries who have been performing disparate impact analysis for rate filings in regulated states will find this familiar. Those who have not will need to develop new capabilities.

Multivariate documentation is newly required. Section 3.2.4 creates specific obligations around documenting how risk characteristics interact in multivariate models. For pricing actuaries building GLMs, this means documenting interaction terms and their actuarial justification. For those building tree-based or neural network models, it means documenting how the model captures multivariate relationships and how the actuary has evaluated whether those relationships produce reasonable classification outcomes.

Patterns from the SR 26-2 model risk framework that the banking sector adopted in April 2026 offer a useful parallel. The banking guidance's emphasis on conceptual soundness, outcomes analysis, and ongoing monitoring maps directly onto ASOP No. 12's classification evaluation framework. Actuaries building model governance programs for risk classification can draw on SR 26-2's materiality-based tiering approach and continuous monitoring philosophy, even though that guidance was written for banking rather than insurance.

Why This Matters

ASOP No. 12's revision addresses a gap that has been widening for two decades. The standard governing risk classification, the core actuarial function that determines how premiums are distributed across policyholders, was last updated when the most sophisticated classification tool in common use was a GLM with a handful of variables. The profession now classifies risks using models with hundreds of features, trained on data streams that did not exist in 2005, deployed in regulatory environments that are actively developing AI-specific oversight frameworks.

The revision matters for three distinct audiences.

For pricing and underwriting actuaries, it creates the first ASOP-level framework for documenting ML-based classification work. The new sections on data and model, multivariate effects, unintended bias, and protected classes translate current best practices into professional obligations. This raises the floor for all practitioners, which strengthens the profession's credibility with regulators and the public while creating incremental compliance work for actuaries whose documentation practices lag current expectations.

For actuarial candidates preparing for exams, the revised ASOP No. 12 is directly relevant to syllabus coverage of risk classification principles, professional standards, and ethical obligations. The SOA and CAS exam pathways both cover ASOPs, and the revision's treatment of algorithmic fairness and model documentation reflects the profession's direction. Candidates who understand the revised standard's framework will be better prepared for both exam questions and early-career practice.

For regulators and consumer advocates, the revision signals that the actuarial profession is incorporating AI governance into its own professional standards rather than waiting for external mandates. A standard that requires actuaries to assess unintended bias and consider protected class impacts provides regulators with a professional-standards anchor for their own oversight expectations. When the 68% of insurers outsourcing AI submit rate filings, the filing actuary's compliance with the revised ASOP No. 12 becomes part of the regulatory evaluation, strengthening the connection between professional accountability and consumer protection.

The timeline creates urgency. A final standard adopted in late 2026 would take effect in mid-2027. Rate filings submitted after that date would need to comply with the new requirements. Actuaries building ML-based classification models in 2026 should structure their documentation and analysis workflows around the exposure draft's requirements now, rather than retrofitting compliance after the standard takes effect. The regulatory environment, with the NAIC Model Bulletin in force, the 12-state evaluation tool pilot running, and Colorado's bias attestation requirements expanding, already expects much of what the revised ASOP No. 12 will formalize. The standard is catching up to where practice and regulation have already moved.