How can I balance accuracy and explainability in a regulated credit risk model?

[Aria_R]

I'm a data scientist working on a credit risk model, and while our new ensemble method has excellent performance, the compliance team is pushing back because we cannot adequately explain individual predictions, which is a regulatory requirement. We need to implement some form of Explainable AI to satisfy both internal stakeholders and auditors, but I'm unsure whether to use model-agnostic methods like SHAP or LIME or to switch to an inherently interpretable model, potentially sacrificing some accuracy. For others in regulated industries, what practical frameworks or tools have you successfully used to bridge the gap between complex model performance and the need for clear, actionable explanations of why a specific application was approved or denied?