A model used in underwriting must be demonstrably fair. What controls are architectural, and what is the hard part?

The architectural controls

• Subgroup evaluation as a required gate, not an optional analysis. Aggregate performance can be excellent while a subgroup is served badly, and that subgroup is frequently the one that matters legally.
• The data needed to measure it, which is the immediate tension: assessing fairness across a protected characteristic requires knowing it, and collecting it may be restricted or undesirable. The resolution is usually a governed, separately-controlled dataset used only for assessment — which must be designed rather than improvised.
• Proxy detection, since a model can reproduce a protected characteristic through correlated features without ever seeing it. Excluding the characteristic does not remove the effect and is the most common misunderstanding.
• Monitoring after deployment, since the population changes and a fair-at-launch model can drift.
• An appeal path that reaches a human with authority to change the outcome, since an appeal returning the same automated answer is not an appeal.

The hard part

Fairness is not one property. Equal outcome rates, equal error rates, equal treatment of similar individuals and calibration across groups are different definitions, and they are mathematically incompatible except in special cases — so satisfying one means failing another.

Which definition applies is a legal and business determination, not an engineering one, and it must be made explicitly before the evaluation is designed. A team that picks one implicitly has made a policy decision without saying so.

The insurance-specific complication

Risk-based pricing is the product, and differentiating on risk is both legitimate and regulated. The line between a legitimate risk factor and a proxy for a protected characteristic is a regulatory judgement that varies by jurisdiction and changes.

That makes the factor set a governed artefact with an effective date, reviewable and defensible — rather than whatever the model found predictive.

The documentation that carries the weight

What was tested, against which definition, on which subgroups, with what result, and what was accepted. That record is the answer to a challenge — and its absence means the answer is a re-analysis conducted under pressure, which is the worst possible circumstance to conduct one.