The Architecture of AI Governance in Banking: Bridging Compliance and Explainability Gaps

Abstract

The adoption of AI in the banking sector has made its way into various functions of the banking business such as credit assessment, fraud detection, risk management, compliance, and customer service, thus introducing fresh challenges in the realms of transparency, accountability, and risk management of banking models. This paper reviews the challenges of governance of AI in banking, focusing on the limitations of explainability, regulations and how AI model risks intersect with traditional banking risks. Explainability is not a single technical property but rather a governance capability that helps to document, reconstruct, challenge and monitor decisions throughout the model lifecycle. The regulatory and standards expectations are discussed, such as lifecycle governance, independent validation, data lineage, documentation, monitoring and senior management accountability. Special focus is placed on the constraints of the post hoc explanation, the performance vs. interpretation dilemma, the need for different explanations for different stakeholders and problems in the cases of complex and emerging AI models. Practical application patterns illustrate the application of interpretable models, post hoc methods, counterfactual explanations, audit trails, human oversight, and ongoing testing. Third party models, data quality, cyber security, model drift and integration with existing credit, market, liquidity, operational, compliance and reputational risk models are also taken into account. By leveraging AI, a coordinated governance strategy integrating technical validation, regulatory alignment, traditional risk management, documentation, and necessary human oversight can enhance accountability while addressing the real-world constraints of complex AI systems in regulated banking settings.