Adverse Action Explanations in Credit Decisions (Compliance)
Under the Fair Credit Reporting Act (FCRA), lenders must provide an "Adverse Action Notice" containing specific, accurate reasons when denying credit. If a machine learning model (e.g., XGBoost) outputs feature importances or SHAP values indicating why a loan was denied, and you use a Generative AI agent to draft a personalized letter explaining this decision to the customer, how do you prevent the LLM from hallucinating reasons or violating compliance?