Back to blog

Why does your best AI credit model still fail the regulator?

Thắng Đoàn
Thắng Đoàn

Your most accurate model may be the one you cannot deploy. In credit, the models that predict best are often the hardest to explain, and that gap now decides market access.

Performance and explainability pull apart

Explainable AI means a model that can show, in plain terms, why it reached a decision. Revolut's in-house foundation model, PRAGMA, was pre-trained on 24 billion banking events across 26 million user records, per TechTimes. The company reports large gains in fraud detection and credit scoring. Yet TechTimes notes a structural tension: the architecture behind those gains clashes with rules requiring explainable decisions in AI-assisted credit. Better predictions and defensible decisions are becoming two different products.

Regulators are writing the rulebook now

Kenya's Central Bank requires digital lenders using AI for credit to explain automated outcomes, disclose when AI is involved, and keep human oversight, per TechTrends KE. The rules also cover bias, accuracy, robustness and accountability, giving examiners a framework to inspect models. Explainability is no longer a nice-to-have. It is a condition of market access.

Explanations move money, not just audits

CPayant deployed Seekr's explainable credit scoring for African trade finance. Per the company's own press release, the system shows financing partners why a business is creditworthy and shows applicants why they scored as they did. CPayant reports over 75 percent of applicants approved, a company-claimed figure, not an independent audit. A clear profile changes the size of the approval, not just the odds. Framed that way, transparency becomes a sales asset, not overhead.

General-purpose AI fails the compliance bar

A Fico survey found 77 percent of Asia-Pacific banking leaders say AI tools in banking must be built specifically for financial services, the Bangkok Post reports. Banks say they need AI that is explainable, auditable and designed for compliance from the ground up. Data privacy and regulation top the barrier list. Consumer-grade AI imported into lending carries risk you cannot audit away later.

What this does not solve

The Financial Brand's analysis notes that less than half of consumers trust AI with their financial data. An explained decision from a distrusted institution may still lose the customer. And none of these sources prove explainable models match black-box performance at scale. You may still face a real trade-off between the best score and the defensible one.

If your best model were blocked tomorrow for opacity, what would you ship instead, and how much accuracy would you trade for defensibility?
Which of your AI decisions could you explain to a customer, a regulator and an auditor today without engineering help?

List every AI model touching customer decisions in your bank, and mark each one: can you explain a rejection today, yes or no?

Sources

The Financial Brand: Risk and Compliance With AI

PR Newswire: CPayant Deploys Seekr Explainable Credit Scoring

Bangkok Post: Regional banks want purpose-built AI

TechTrends KE: Kenya's digital lenders face tougher AI rules

TechTimes: Revolut PRAGMA foundation model

Share:

Recommended for you

Enjoyed this article?

Subscribe for new articles. No spam. Unsubscribe anytime.

By subscribing you agree to receive the newsletter. No spam, and you can unsubscribe anytime.