Core Banking Won't Save You: Why African Banks Need AI-Native Risk Engines

Mobile money velocity, agent float behaviour, airtime top-up regularity, supplier payment rhythm, and the stability of counterparties are richer behavioural predictors than most payslips. A trader who has settled with the same three wholesalers every week for four years is demonstrating something a salary slip cannot: durability under actual trading conditions.

The work is building a risk engine that ingests these natively, as time series and graph features rather than as a handful of ratios copied into a legacy scorecard. Retrofitting two mobile money fields into an existing application form does not do this. It produces a model that still fundamentally scores formality.

Central banks in the region will rightly ask how a model reached its decision, and so will a declined customer. That rules out opaque ensembles trained on features nobody can articulate. Feature engineering has to produce human-legible drivers: turnover stability, counterparty concentration, seasonal dip depth, float recovery time.

This constraint is a gift disguised as a burden. Legible features are also the ones credit officers can challenge, which is how model error gets caught before it becomes portfolio loss.

The safe deployment path is unexciting. Score every application the existing process scores, in parallel, without influencing the outcome. After six months, compare: which loans the new engine would have approved and how they actually performed, and which it would have declined that the bank booked and lost.

Only then does the model touch an approval, and even then on one portfolio with a hard exposure cap. Banks that skip the shadow phase either overlend into an unproven segment or lose confidence at the first default and abandon a model that was working.

Thin-file lending is not charity; it is the largest underserved credit market on the continent, and the institutions that learn to underwrite it will hold customer relationships for decades. Fintechs are already doing this with worse balance sheets and better data pipelines. The banks' advantage is cost of funds and distribution. Their disadvantage is a risk function that cannot see the borrower.