Board Case Study · Financial (Lending)
The AI credit model approved more borrowers and cut defaults, then a fair-lending exam found it declined a protected class at twice the rate, using proxies for race it was never given.
The situation
A bank deployed a machine-learning underwriting model that expanded approvals and lowered default rates, a clear commercial win. A regulatory fair-lending examination found the model denies a protected class at roughly twice the rate of comparable applicants, a potential disparate-impact violation of ECOA and the Fair Housing Act.
The model was never given race. But it used ZIP code, education, and alternative data that function as proxies, reproducing historical lending patterns, algorithmic redlining. Adverse-action notices generated by the model may not meet the law's requirement to give specific, accurate reasons for denial.
The board must address remediation for affected applicants, consent-order risk, model explainability, and the tension between a model that is both more profitable and potentially discriminatory.
“We never gave it race. It found race anyway, in the ZIP code. Is a more profitable model worth a fair-lending violation?”
Chair, Risk Committee
The decision on the table
The full case, scenario architecture, board materials, and facilitation notes, is shared with boards and partners on request.