Actuarially-Grounded AI: Bridging Machine Learning and Regulatory Compliance in Insurance

The insurance industry stands at a critical inflection point. Machine learning models promise unprecedented accuracy in pricing, reserving, and claims prediction, yet their opacity creates fundamental tensions with regulatory requirements for explainability and actuarial accountability. This paper introduces the concept of "Actuarially-Grounded AI", a framework that synthesises existing techniques into a coherent, regulatory-ready methodology by embedding actuarial principles directly into machine learning workflows. The framework builds upon the Combined Actuarial Neural Network (CANN) architecture introduced by Wüthrich and Merz (2019), integrating it with explainability methods such as SHAP (Lundberg & Lee, 2017) and LIME (Ribeiro et al., 2016), while mapping to current regulatory requirements including the EU AI Act (Union, 2024), EIOPA's Opinion on AI governance (EIOPA, 2025), and FINMA Guidance 08/2024 (FINMA, 2024). This article introduces practical implementation strategies for insurers and reinsurers navigating this transformation.