What Loan Officers in Lagos Know That Your Credit Model Doesn't

Credit models are remarkable instruments. They take dozens of data points — revenue history, repayment behaviour, collateral, sector exposure, business registration, transaction records — and produce a score. The score tells you whether a borrower qualifies. It does this reliably, at scale, without the inconsistency of human judgment, and with an audit trail that satisfies every compliance requirement a lender has.

And yet loan officers who have worked in Lagos for eight years often know something the model cannot see.

The model isn't wrong. It was built from data — and some of what determines whether credit actually reaches the people it was designed for was never in the data. It was in the room.

The score says this borrower qualifies. The loan officer knows this borrower will not apply. Not because they're ineligible. This branch's reputation, this quarter's relationships, this season's dynamics have already made the calculation for them, in ways no credit model was designed to capture and no application will ever record.

This is not an argument against credit models. It's an observation about what they were built to do and what they weren't. Models score what is in the data. They do not score what never made it into the data: the seasonal patterns that fall outside the measurement window, the informal histories that were resolved without a bureau entry, the market dynamics the sector classification does not reach.

The model scores the form. The officer scores the conversation.

Credit models and loan officers are not competing ways of doing the same thing. They are instruments for different kinds of knowledge — and the gap between them is not a design failure. It's a design boundary.

A model is built from historical data. It learns what patterns predicted outcomes in the past and applies that learning to new applications. It is stable, consistent, and auditable. What it cannot do is update on information that never entered the historical record, because information that was never recorded cannot become a pattern.

A loan officer learns from proximity. Eight years of sitting across from borrowers in one geography produces knowledge that can't be systematised: the relevant variables change faster than any model can be retrained, and the most important signals are often the ones nobody designed to capture.

The model does its job correctly at every point. It evaluates the information it was given accurately. What it cannot do is evaluate the information it was never given, and some of the most important information about whether a borrower will repay was never going to appear in a form.

A credit model can tell you whether a borrower qualifies. A loan officer can tell you whether they will repay. Those are not always the same question, and in Lagos they are often not answered by the same instrument.

What eight years across from borrowers teaches you

Experience does not simply accumulate. It compresses. Eight years of facing borrowers across the desk, watching which applications repaid and which did not, learning which signals preceded problems and which preceded prompt repayment, produces a form of calibration that no training programme was designed to produce and no dataset can replicate.

It is not infallible. But it is a different kind of knowledge from what the model holds: specific, local, current, and conditional in ways a model trained on historical data cannot be.

Where the model and the officer diverge — and what that costs

The model and the loan officer usually agree: applications that score well tend to repay, and ones that score poorly tend to default. When they converge, the model is doing what it was built to do, efficiently and at scale.

Where they diverge is where the interesting things happen — and where decisions are most expensive to get wrong.

These are not edge cases. They are the predictable divergence points between two instruments measuring different things — the model measuring the formal record, the officer measuring operating reality. In markets where the gap between those two is widest — where informal dynamics, seasonal patterns, and network effects determine outcomes more than formal documentation — the officer's knowledge is not a supplement to the model. It is a different kind of intelligence entirely.

What this means for how intelligence should be gathered

The practical implication is not that credit models should be replaced. It's that credit intelligence — the information that lets organisations understand and deploy credit responsibly in African markets — has at least three distinct layers. Each layer requires a different instrument, and the most common failure is using one instrument to do all three jobs.

Most organisations working on credit access in African markets are investing heavily in Layer 1. Many are building Layer 2. Almost none have a reliable way of reaching Layer 3 — because Layer 3 does not exist in any database. It exists inside the loan officers who have been in the room.

The practical implication is specific: before you design a credit product for a new geography, diagnose why uptake is lower than projected, or decide whether to expand a lending programme into a new market, one session with a loan officer who has worked that geography for three to five years will surface things no model and no market research will find. The model is not wrong — it was built from the data it had, and the most important data was never recorded. And the session becomes a decision asset, searchable, retrievable through Ask Kofa, so the next credit decision in that geography starts where the last one finished.

It is in the conversation. It has always been in the conversation.

Aminu Rabiu

Founder, Kofa Insights Limited