The Fair Pricing Playbook

A practical framework for Responsible AI in algorithmic pricing (2026)

Why fair pricing matters

Algorithmic pricing is now widely used in insurance and financial services, and advances in AI and machine learning have become central to how insurers assess risk, detect patterns in large datasets, and set prices. The fairness concerns this creates are not new, but they take a different form under automated, large-scale systems. Even when protected attributes such as race, gender, and religion are removed from a model, pricing systems can still produce discriminatory outcomes through proxy variables and opaque algorithms that are difficult to trace or challenge. The rapid adoption of AI and big data has created a regulatory grey area: direct discrimination is prohibited, but indirect discrimination through proxies and complex algorithms is not clearly specified or assessed in most markets (Frees and Huang 2023; Xin and Huang 2024).

Regulatory attention to these issues has grown substantially. Colorado and New York have each proposed rules requiring insurers to test pricing algorithms for unfair discrimination, and the European Union has enacted broader AI governance frameworks that apply to insurance. The EU AI Act classifies credit scoring and insurance risk assessment as high-risk AI applications, placing fairness obligations alongside broader requirements for transparency, human oversight, and accountability. Firms that cannot demonstrate fairness face legal exposure, reputational risk, and the practical challenge of defending pricing decisions under regulatory examination.

This playbook translates research from economics, statistics, actuarial science, and machine learning into a concrete four-step workflow that covers the full journey from defining what fairness means, to building fair models, to measuring who actually gains and loses, to auditing a deployed system. While the case studies use insurance data and draw on insurance regulation, the four-step framework applies to any sector where algorithmic pricing raises fairness concerns, including credit, housing, energy, and mobility. Three insurance case studies provide the technical depth needed for implementation.

TipTwo ways to read

Reader	Recommended path
Compliance, risk, policy, and management	Read the four steps. The case studies are optional technical supplements.
Data science, actuarial, and research	Read the four steps, then follow the case studies linked in the sidebar for implementation detail.

Watch: Introduction to anti-discrimination insurance pricing

The four steps

Step 1 · Define fairness

This step surveys the different lenses of fairness and the criteria they imply. They cannot all be satisfied simultaneously, and the choice is a legal and policy decision, not a technical default.

Step 2 · Design fair pricing

Once a criterion is chosen, this step maps it to a concrete model design and shows where in the pipeline (inputs, training, or outputs) to enforce fairness, based on what regulation requires.

Step 3 · Assess impact

A fair cost model does not guarantee a fair market outcome. This step traces price effects through demand and competition, quantifying consumer welfare and firm profit by group after prices are set.

Step 4 · Audit the system

An audit tests whether a deployed system actually meets the standard. All protocol choices (criterion, tolerances, sample design) are fixed before examining any data. Results are pass, fail, or insufficient information.

Four case studies provide implementation depth. Each is linked from the step it illustrates. Case study: Fairness Metrics follows Step 1 as a qualitative walkthrough with no code. Case study: Fair Models follows Step 2 with a reproducible model pipeline. Case study: Welfare Implications follows Step 3 with a reproducible welfare simulation. Case study: Fairness Testing follows Step 4 with a reproducible audit pipeline. Case studies are technical supplements. The four steps stand alone for non-technical readers.

The pathway at a glance

The diagram shows how the four steps connect across model design, welfare assessment, and fairness testing. Steps 1 through 3 work together to inform a single decision, which fairness criterion fits a specific application or line of business. Step 2 shows the accuracy-fairness trade-off, how much predictive power a fairness constraint costs at the model level. Step 3 shows the welfare-fairness trade-off, how a fair cost model can still produce unequal market outcomes once price optimisation and demand enter the picture. Neither trade-off has a universal answer, which is exactly why these steps matter. Seeing both is what lets stakeholders choose a criterion in Step 1 that fits their product, market, and regulatory context, rather than defaulting to whichever criterion is easiest to implement.

What the research shows

NoteKey findings from the research

• Fairness through unawareness is not enough. Removing the protected attribute leaves proxy effects intact. Indirect discrimination can survive even when all obvious proxies are excluded. See Step 1 →
• Fair models are feasible. Four model families (MU, MDP, MCDP, MC) implement four distinct criteria with manageable accuracy trade-offs. On benchmark data, proxy removal (MCDP) reduced adverse selection risk more than simple exclusion (MU). See Step 2 →
• Fair models do not guarantee fair outcomes. Price optimisation and competitive dynamics reintroduce group disparities after cost-level fairness is achieved. Demographic parity on premiums can widen markup disparities even as it closes price gaps. When all firms in the market adopt the same fairness rule, welfare losses for female customers are nearly eightfold larger than when only one firm is regulated ($101.67 vs $13.43 per female customer), because consumers can no longer avoid the impact by switching to an unregulated competitor. See Step 3 →
• Standard statistical tests are unreliable for auditing algorithmic pricing. Corrected inference, equivalence testing, and (where race is unobserved) ethnicity-estimation corrections are all required for a valid audit. Insufficient data yields an inconclusive result, not a pass. See Step 4 →

Step	Key references
1 · Define fairness	Frees and Huang (2023); Xin and Huang (2024); Krafcheck et al. (2026)
2 · Design fair pricing	Xin and Huang (2024)
3 · Assess impact	Huang et al. (2026); Huang and Shimao (2026)
4 · Audit the system	Huang and Hooker (2026); Xin et al. (2026); Xin et al. (2025)

End-to-end audit flow

Step 4 follows an integrated Plan, Audit, Decide, and Improve protocol (Huang and Hooker 2026). All design choices are fixed before examining data.

Phase	Actions
Plan	Select criterion (PD or CDP), legitimate factors, tolerance bands, representative sample
Audit	Collect prices and run statistical fairness tests with corrected inference
Decide	Pass, insufficient information, or fail (equivalence testing)
Improve	Remediate and re-test. Collect more data if result is unclear

Full detail is in Step 4: Audit the system.

See the About page for author bio, acknowledgements, citation, and licence.

References

Frees, Edward W, and Fei Huang. 2023. “The Discriminating (Pricing) Actuary.” North American Actuarial Journal 27 (1): 2–24.

Huang, Fei, and Giles Hooker. 2026. Fairness Testing for Algorithmic Pricing. https://arxiv.org/abs/2605.11614.

Huang, Fei, and Hajime Shimao. 2026. “Welfare Implications of Fair and Accountable Insurance Pricing.” Journal of Risk and Insurance, ahead of print. https://doi.org/10.1111/jori.70051.

Huang, Fei, Hajime Shimao, and Warut Khern-am-nuai. 2026. Do Fair Algorithms Improve Welfare? Evidence from the Insurance Market. UNSW Business School Research Paper Forthcoming. https://doi.org/10.2139/ssrn.5112616.

Krafcheck, Eric, Igor Balnozan, and Fei Huang. 2026. Fairness Metrics for Life Insurance. Society of Actuaries Research Institute. https://www.soa.org/resources/research-reports/2026/fairness-metrics-life-insurance/.

Xin, Xi, Giles Hooker, and Fei Huang. 2025. “Pitfalls in Machine Learning Interpretability: Manipulating Partial Dependence Plots to Hide Discrimination.” Insurance: Mathematics and Economics 125: 103135. https://doi.org/10.1016/j.insmatheco.2025.103135.

Xin, Xi, Giles Hooker, and Fei Huang. 2026. How Proxy Race Distorts Regression-Based Fairness Audits. https://arxiv.org/abs/2603.17106.

Xin, Xi, and Fei Huang. 2024. “Antidiscrimination Insurance Pricing: Regulations, Fairness Criteria, and Models.” North American Actuarial Journal 28 (2): 285–319.