As AI becomes embedded in insurance, ensuring fairness shifts from a one-time validation to an ongoing operational requirement. RGA’s recent collaboration with EY shows how insurers can operationalize fairness in a way that aligns with evolving regulatory expectations.
AI in insurance is no longer experimental. Predictive models have long supported underwriting, pricing, and claims decisions, and newer forms of AI, including generative tools and emerging agentic systems, are expanding how decisions are informed and executed.
As adoption has accelerated, so has scrutiny. Regulators, boards, and customers increasingly expect insurers to demonstrate that models perform well and that they do so fairly and consistently.
Much of the early industry conversation focused on ethics frameworks, principles, and governance structures. Those efforts established a shared vocabulary for responsible AI. However, they did not fully answer the practical question insurers now face: How does fairness work once models are deployed, reused, and maintained over time?
RGA recently collaborated with EY to address that gap by developing a repeatable approach to fairness testing that could scale alongside advanced analytics and align with evolving regulatory expectations.
The most useful lessons emerged after the framework was in place, when fairness moved from theory into everyday operations.
“AI” is often treated as a single category, but in practice it now spans fundamentally different types of systems.
This distinction affects how fairness is defined, tested, and monitored. Predictive models lend themselves to established fairness metrics and validation approaches. Generative and agentic systems introduce additional uncertainty, requiring adaptation of those same governance principles rather than replacement.
The framework described in this article was developed in the context of predictive models, but its core principles extend across all three, though with different testing and monitoring approaches.
Fairness discussions often assume a single moment of evaluation. A model is developed, tested, and approved. In reality, that moment is only the beginning.
Once deployed, models behave like infrastructure rather than static assets. They are embedded in workflows, adapted for new products, recalibrated as experience data evolves, and interpreted by users far removed from the original development team.
Over time, the assumptions underlying fairness testing can shift. This challenge becomes even more pronounced with generative and agentic systems, where behavior is less stable and less predictable.
As a result, the challenge shifts from defining fairness metrics to ensuring that fairness holds as models evolve, scale, and interact with operational realities.
Insurers that treat fairness as a one-time validation risk discovering issues long after decisions have been made. Those that treat it as an ongoing responsibility are better positioned to respond with evidence, not explanations.
RGA’s collaboration with EY helped formalize a bias testing framework that can be applied consistently across models. Operational experience clarified four defining traits of effective frameworks after deployment.
These principles apply across predictive, generative, and agentic AI, but the methods differ. Predictive models can be re-tested against defined outcomes, while generative and agentic systems require more dynamic monitoring of outputs and use cases.
A recurring lesson from operationalizing fairness is that friction rarely arises from the testing itself. The analytical work is often manageable.
Challenges emerge when organizations must act on findings:
Without predefined decision rights, responsibility diffuses and decisions stall. Effective governance reduces this friction by making roles and escalation paths explicit in advance.
The regulatory landscape can appear fragmented, but expectations are converging around a consistent set of requirements:
In the U.S., guidance such as the NAIC Model Bulletin establishes expectations for formal governance, documentation, and oversight of third-party models. State-level developments, including those in New York and Colorado, further emphasize testing, documentation of less discriminatory alternatives, and ongoing review.
These developments make one point clear: Organizations cannot outsource accountability for fairness outcomes, even when external partners or vendors are involved.
For insurers, this shifts regulatory conversations toward evidence of how fairness is defined, monitored, and maintained over time, rather than toward intent alone.
RGA chose to collaborate with EY because credibility matters. External partners provide independent perspective and help translate regulatory expectations into actionable practices.
Importantly, collaboration does not replace internal accountability. In the RGA-EY collaboration, fairness testing remained internal, with external guidance strengthening rigor and consistency.
The industry recognizes that fairness must be governed. The harder question is how to sustain it without slowing innovation.
Experience suggests the answer lies in integration, rather than addition. Embedding fairness into existing model risk governance ensures that it is continuously reviewed, challenged, and improved alongside performance and risk.
This approach also reshapes regulatory engagement. Organizations that can demonstrate how fairness is operationalized move discussions from principles to evidence, supporting both compliance and continued innovation.
Ethical AI frameworks established a necessary foundation. Operational fairness builds on that foundation by addressing how models behave in real environments.
The distinction between predictive, generative, and agentic AI highlights where existing approaches apply and where they must evolve. At the same time, regulatory expectations are clarifying what organizations must be able to demonstrate.
The experience from the RGA-EY collaboration suggests insurers do not need to choose between innovation and responsibility. With the right operating model, fairness governance can support both.
Read more about the EY-RGA collaboration from EY’s perspective, and contact RGA to discuss how a partnership with us can help advance your AI initiatives.
