Artificial Intelligence for Lending: Underwriting, Fraud, and Portfolio Risk

A consumer lender builds an ML underwriting model. The model outperforms the existing scorecard on default prediction in backtesting. The model risk team and fair lending team review. The questions that determine whether the model can go into production: what features drive the decision and can they be explained in adverse action notices under ECOA Reg B, does the model produce disparate impact against protected classes under fair lending, does the feature set include any variables that are proxies for prohibited bases, and can the firm demonstrate that any disparate impact is justified by legitimate business necessity with no less-discriminatory alternative. Generic ML models fail these questions. The model goes back for redesign — not because it doesn't predict defaults well, but because the features and architecture weren't built with fair lending in mind from the start.

Lending AI that passes fair lending review is designed backward from the regulatory constraints. Features selected for predictive power within the approved feature set — no ZIP codes or surnames or variables that proxy for prohibited bases. Architectures that support adverse action reason code generation (top decline reasons must be explainable to the borrower under Reg B). Disparate impact testing as part of the model development process, not a post-hoc check. Less-discriminatory alternative analysis when disparate impact exists. Model risk documentation that satisfies SR 11-7 for banks and the equivalent expectations at non-bank lenders. With the monitoring that catches performance degradation and disparate impact emergence in production. Done this way, AI delivers measurable credit performance improvement. Done as generic ML, it never gets approved for production.