Credit products differ by an order of magnitude in latency envelope, loss severity, and regulatory treatment, and almost never share a single model cleanly. A BNPL checkout scoring in 500ms has nothing architecturally in common with a mortgage model that runs in 30 seconds against 1,500 features — and the CFPB, OCC, and EU AI Act treat them differently under adverse-action and high-risk classification rules. The most common failure pattern is one-sizing a card-origination model onto BNPL or auto after product launch and inheriting the wrong feature surface and the wrong governance posture.

The 1-5 second real-time envelope is not a soft target — it is set by merchant checkout flow, dealer F&I desk workflow, and consumer abandonment curves. Upstart publicly reports 91% fully-automated decisioning in under a few seconds; the economic advantage of AI underwriting largely evaporates if the model cannot return a decision inside the channel's natural decision window. Explainability adds 30-50% overhead (SHAP KernelExplainer runs 500 model evaluations per prediction), so the latency budget must plan for score + explanation as a single unit, not score alone.

Adverse selection is the signature failure mode of credit models — the model's accepted pool systematically concentrates the borrowers who exploit the features it weights most heavily. Upstart reports 53% lower default rate at equal approval volume, and Zest AI reports ~28% charge-off reduction, but both are measured against vintage-expected curves rather than point-in-time approval rates. Without a vintage-anchored tolerance, model teams optimize approval lift that looks great for 60 days and produces a charge-off wave 12-18 months later.

ECOA / Regulation B (12 CFR 1002) requires a specific, accurate, human-understandable adverse action notice with up to four principal reasons within 30 days of the decision. The CFPB's 2022 Circular and 2024 enforcement posture made it unambiguous that "the model said so" is not compliant — the top factors must map to real features the applicant can understand and, in principle, remediate. An AI model that can't produce reason codes isn't a credit model; it's a research artifact. Designing reason-code generation into the scoring path (rather than bolting it on) is the difference between a 2-second decision and a 15-second decision.

The CFPB in 2024 made clear that examiners are now actively searching for LDAs and that failing to search is itself evidence of non-compliance — "the decision to use an algorithmic tool can itself constitute a discriminatory policy." Any AI credit model in production should have an auditable record of alternative models considered, the disparate-impact results of each, and the business justification for the selection. This is a documentation artifact, not a modeling exercise — and it must exist before the regulator asks.

EU AI Act Article 14 (effective August 2026 for high-risk systems, which explicitly includes credit scoring per Annex III) requires that human oversight be meaningful, not ceremonial — the reviewer must be able to understand, question, and override the automated decision. SR 11-7 reaches the same conclusion from a different angle: the model's use must be governed by defined change, monitoring, and escalation procedures. A "human in the loop" who cannot actually interpret the model's output is non-compliant under both regimes.