Data Engineering for AI-Powered Compliance: A New Paradigm in Banking Risk Management

Financial services companies have specific explainability needs around AI/ML use cases driven by the unique scope and nature of their businesses and compliance needs. Banks interact with their customers’ money, and regulatory exams assess financial stability, risk management, and compliance with applicable laws in a complex dynamic regulatory landscape. One notable recent regulatory event is the transition to the Calculator tax regime for interest rate problem banks. A $250 monthly fine was imposed for not complying with a multi-year regulatory deadline with complexity similar to that of Building a Large Language Model. Index banks navigating this compliance need using decision trees and linear models circumvented the risk of high-impact black-box models. Some prior financial services explainability needs detailed use cases in the scope of fairness and discrimination in credit decisions. Other needs arise from one’s role as a regulated financial institution, particularly those mentioned in model risk management. Global Systemically Important Banks (G-SIBs) provide a unique scope of explainability needs for AI/ML use cases verifying compliance with a specific AI governance standard specified by the Basel Committee on Banking Supervision. This standard, BCBS 239, is in the public domain, safeguarded by an explanation as to how it further families the competence model risk management. Fully complying with the standard is an arduous and multifaceted task but opening the doors to explainable AI treatment enhancing without enormous upfront or ongoing effort. As all business-related inquiries can be framed as answers to these standard questions, initial runway development tasks in any target area would assess current status toward compliance with the standard and prioritize potential additional treatment. In short, proactively treating coordination and controlled use of AI can concretely ease verification of compliance with an AI governance standard and win over some “if” arguments on human-AI collaboration needed. G-SIBs’ explanations are broadening without specifications to input data context and output reporting while still needing more direct machine learning interpretability.