Claims automation, examined earlier in this series, is judged largely by whether a specific decision about a specific claim was fair and well-documented. Underwriting automation faces a different, in some ways more demanding test: not merely whether an individual decision was reasonable, but whether the entire pricing methodology behind it can be actuarially justified, filed with a state regulator, and defended as something other than an unlawful proxy for a protected characteristic. This distinction — fairness in application versus justification of the underlying methodology — is what makes underwriting a genuinely separate problem from claims, deserving its own treatment even within the same industry.
1. The Actuarial Justification Standard
Insurance rate-making has always operated under a principle, embedded in state rate-filing statutes across the country, that rates must be actuarially sound — based on a demonstrated relationship between the rating factor and the underlying risk, not on arbitrary or discriminatory grounds. The Actuarial Standards Board's ASOP No. 12, on risk classification, provides the professional standard actuaries are expected to follow in establishing that relationship, requiring that a rating variable be supported by credible data showing it actually predicts loss, not merely that it happens to correlate with loss in a way that might reflect an unrelated, potentially discriminatory pattern.
An AI underwriting model trained to optimize purely for predictive accuracy, without regard to whether each contributing variable can be independently actuarially justified, is optimizing for the wrong objective from a regulatory standpoint. A variable can improve a model's raw predictive accuracy while being, from a rate-filing perspective, indefensible — either because it has no plausible causal or actuarial relationship to risk, or because it functions as a proxy for a characteristic insurers are prohibited from using directly.
2. Proxy Discrimination and the Colorado Framework
Colorado's SB21-169, discussed elsewhere in this series in the claims context, applies with particular force to underwriting and pricing, requiring insurers to test external consumer data and the algorithms built on it for unfairly discriminatory outcomes specifically in this context, and to report the testing methodology to the state regulator. The statute reflects a broader regulatory concern, shared increasingly across states, that a sophisticated pricing model can reconstruct a prohibited characteristic — race, in particular — through a combination of facially neutral variables like geography, credit-adjacent signals, and consumer behavior patterns, even when the model was never given race as an input directly.
This is a genuinely difficult technical problem, not merely a compliance-paperwork exercise: proxy discrimination can be subtle enough that it survives a superficial review of the model's input variables and only becomes visible when outcomes are tested directly across demographic lines, which is precisely why the emerging regulatory framework increasingly mandates that outcome-testing rather than relying on input-variable review alone.
3. State Rate-Filing Requirements as a Design Constraint
Most states require insurers to file and, in many cases, receive approval for rating methodologies before they can be used, a process that requires the insurer to explain, in terms a state regulator can evaluate, exactly how each variable contributes to the final rate. An AI underwriting model built as an opaque, end-to-end system — raw applicant data in, a price out, with no legible intermediate structure — is very difficult to file under this framework, because the regulator's review depends on being able to examine the individual contribution of each rating factor, not merely evaluate the model's aggregate output.
This creates a genuine engineering constraint that should shape model design from the outset, not be addressed after the fact: an underwriting model built for eventual rate-filing needs to preserve enough structural transparency — which factors it uses, and roughly how each one contributes to the final price — that an actuary can translate its behavior into a filing a regulator can actually evaluate, rather than requiring a fundamentally different, less accurate model to be built afterward purely to satisfy the filing process.
4. Architecture as the Practical Answer
The deployment-model question here interacts with the fairness question in a specific way: underwriting draws on some of the most sensitive data categories in the insurance business — medical history, detailed financial records, behavioral and credit-adjacent signals — which argues for a fully on-premises deployment for the underwriting model itself, keeping this especially sensitive data pipeline entirely within the insurer's own controlled and independently auditable environment, where the fairness-testing exercise discussed above can be conducted with full visibility into every input.
A masked secure-cloud deployment may remain appropriate for adjacent, lower-sensitivity workflows — general application intake processing, document collection and organization — that feed into but are structurally separate from the underwriting model's actual rating logic. The line between these two categories should be drawn deliberately around where the actual rating decision occurs, not around a general sense of which workflows feel more or less sensitive.
5. Testing the Model, Not Just the Filing
A rate filing that was accurate and actuarially sound at the time it was approved can drift out of alignment as the underlying model continues to learn or as the applicant population it is applied to shifts over time — and an insurer that treats the filing as a one-time compliance milestone, rather than the starting point of an ongoing testing obligation, risks discovering a proxy-discrimination problem only when a regulator or a plaintiff's expert finds it first. Recurring, structured outcome testing — checking whether approval rates, premium levels, and rating-factor contributions remain consistent with the filed methodology and free of unjustified demographic disparity — should be built into the underwriting program's ongoing operations, not treated as a one-time pre-launch exercise.
This recurring-testing discipline is, again, the same principle that has recurred throughout this series in different regulatory vocabularies: a compliant outcome at launch is not the same as a compliant system, and only continuous monitoring closes that gap reliably.
6. Conclusion
Underwriting automation is held to a standard claims automation is not: every contributing variable has to be actuarially justified, filed, and defensible as something other than a proxy for a prohibited characteristic, not merely applied fairly to the individual case in front of it. An architecture that keeps the underwriting model's most sensitive data fully local, preserves the structural transparency a rate filing requires, and subjects the model to ongoing, not merely pre-launch, fairness testing, is what allows an insurer to modernize underwriting without quietly building a rating methodology it cannot actually defend to the regulator who approved it.