Banking & Finance

"The model is opaque" is not a defense. The regulator said so in writing.

Every institution deploying AI for credit, claims, or eligibility eventually reaches the same comfortable excuse: the model is too complex to explain. In the United States, the regulator has already answered — in writing — and the answer travels well beyond lending.

In 2022, the CFPB stated plainly that creditors using complex algorithms must still provide the specific reasons for an adverse action — and that a model’s opacity is no excuse under ECOA and Regulation B. If your system denies a consumer credit, you owe them the actual reasons. “The algorithm decided” is a violation, not an explanation.

The logic is portable, and regulators everywhere are porting it. The EU AI Act names credit scoring high-risk by name and demands record-keeping and effective oversight. GDPR Article 22 gives individuals rights against purely automated decisions — including meaningful information about the logic involved. Supervisors in the GCC require demonstrable control over AI decisions. Different laws, one converging principle: if you deploy it, you must be able to explain it — case by case, on the record.

an explanation you construct after the complaint is advocacy. An explanation that existed at the moment of the decision — the rule applied, the factors weighed, the human who signed — is evidence. The difference decides examinations. Post-hoc rationalization of a model’s output is exactly what the CFPB warned against; a decision record produced in the live path is exactly what satisfies it.

every AI-influenced decision carries its reasons at birth — checked against the named rule, grounded in reference data, signed, and stored immutably. When the examiner asks why was this customer denied, the answer isn’t a data-science project. It’s a record. That is the infrastructure we build.

― basis: CFPB Circular 2022-03 · ECOA / Regulation B §1002.9 · EU AI Act Art. 6, Annex III 5(b) · GDPR Art. 22