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.

The 2005 version of ASOP No. 12 contains zero references to the word "model." That is the gap the Actuarial Standards Board is now closing. A second exposure draft of the risk classification standard, whose comments closed in March 2025, adds sections on data and models, multivariate effects, unintended bias, and protected classes. A final standard could be adopted at the ASB's September or December 2026 meeting, with an effective date six months after adoption.

Key Takeaways

Zero references to "model" appear in the 2005 standard, which was written when a GLM with a handful of variables was the most sophisticated classification tool in common use.
57 comment letters landed on the first exposure draft between January and May 2024, from the NAIC, the American Academy of Actuaries, Allstate, Travelers, USAA, Verisk and others.
Four new sections carry the revision: 3.2.2 (Data and Model), 3.2.4 (Multivariate Effects), 3.4 (Unintended Bias) and 3.5 (Protected Classes). None had a counterpart in 2005.
Nine disclosure categories replace the 2005 standard's general documentation expectation in a restructured Section 4.1, from intended purpose through reliance on other parties.

What the Second Exposure Draft Changes

The revision restructures the standard around models rather than around variables. ASOP No. 12 was adopted in 1989, broadened to "Risk Classification (for All Practice Areas)" in 2005, and touched only lightly in 2011 to regularize deviation language.

The 2005 framework assumed an actuary working through rating variables one at a time: select a characteristic, test whether expected outcomes differ meaningfully across classes, document the rationale. That is a univariate discipline, and it is not how classification work is performed now.

The ASB released the first exposure draft in January 2024, took 57 comment letters by the May 2024 deadline, approved a second draft in October 2024, and reviewed the second round of comments filed through March 2025. The task force is chaired by Brian J. Mullen, with members drawn from across practice areas.

One change is purely terminological and carries the logic of all the others. "Risk Classification System" becomes "Risk Classification Framework," because a trained model is not a mechanism whose internal logic the actuary specifies. It is learned from data, and the word had to move to accommodate that.

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

The Risk Measure row is the least conspicuous of the seven and reaches furthest. Replacing the 2005 emphasis on expected outcomes with a broader measure invites classification aimed along the whole loss distribution, including tail measures and conditional tail expectations, rather than at the mean.

Where the New Sections Land in a Rate Filing

Section 3.2.4 is the one that changes what a pricing actuary files. It requires documenting how risk characteristics interact in a multivariate context: interaction terms and their actuarial justification in a GLM, and in a tree-based or neural model, how the model captures those relationships and whether the classifications that result are reasonable.

Pair that with Section 3.7, which asks that documentation let another actuary qualified in the same practice area assess the reasonableness of the work, and the obligation reaches the whole development pipeline. A 500-variable gradient-boosted model carries preprocessing steps, feature selection criteria, hyperparameter tuning rounds and model selection decisions, and the draft does not say which of them are material. Carrier and consulting commenters asked exactly that: every intermediate model version, or only the deployed one?

Section 3.4 is harder, because the problem it names is structural rather than procedural. Removing a prohibited characteristic from a feature set does not remove its influence when correlated variables remain, and ZIP code, credit score and vehicle type all carry some of it. The Academy's algorithmic accountability work found that machine learning "tends to produce the same results as intentional proxy discrimination" even with the prohibited variable excluded.

Section 3.4 sets no threshold and offers no safe harbor. The duty it creates is to investigate and document, which is roughly what the NAIC Model Bulletin on AI, adopted in December 2023 and now in force in more than half the states, and Colorado's SB 21-169 bias attestation already ask of carriers. What changes is that the duty follows the actuary into states with no AI governance rule at all. It arrives alongside roughly 20 of 52 active ASOPs under simultaneous revision, and it echoes the materiality-tiered, outcomes-analysis structure of the banking sector's SR 26-2 model risk guidance.

The Standard Cannot Settle the Legal Question

The sharpest comment-period disagreement was over the definition of unintended bias: impacts on specific risk subjects that the framework was not intentionally designed to produce. Several commenters read that as capturing any unplanned outcome rather than a problematic one.

The definitional question is not academic, because unfair discrimination is what a state insurance department actually enforces, and it is a legal test with its own elements. An actuary can satisfy the ASOP's assessment obligation on unintended bias and still face challenge under a statutory standard that draws the line somewhere else. A professional standard cannot resolve that, and Section 3.5 on protected classes is where commenters said so most directly, arguing that protected class definitions and disparate impact thresholds are legal questions a professional body should not be answering.

The second unsettled boundary is internal. The new Data and Model section overlaps ASOP No. 56's existing modeling guidance, which already governs model appropriateness, data quality, testing and disclosure of limitations. The CAS asked the task force to align the treatment of proxy discrimination across ASOPs 12, 53 and 56, and the workable line is narrow: No. 56 governs the model as a modeling artifact, No. 12 governs its use as a classification instrument, and No. 53 governs the cost estimate for each resulting class.

That line holds in principle. In a filing it means one gradient-boosted model documented three times against three standards, two of which are themselves in revision.

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

Key Takeaways

What the Second Exposure Draft Changes

Where the New Sections Land in a Rate Filing

The Standard Cannot Settle the Legal Question

Further Reading

Sources