Microfinance

AI loan repayment scoring in East Africa

A faith-based microfinance NGO considered ML to predict loan defaults. They rejected at Receive when they found CGAP guidelines already addressed this.

Did not build (Receive)
Outcome

Adopted CGAP's published methodology instead of building a proprietary model.

Lament

Who is hurting?

Loan officers in Nairobi and Kampala are exhausted. They're managing 300+ borrowers each. Default rates are rising — not because borrowers are dishonest, but because drought and inflation are crushing them. Leadership is anxious about institutional solvency. Borrowers feel like numbers in a spreadsheet, not neighbors.

What would Christ see here that you've stopped seeing?

Christ would see that the borrowers are not the risk — the system that demands repayment during a drought is the risk. He'd see loan officers who used to know every borrower by name now reduced to data-entry clerks. He'd see leadership treating a moral problem as a technical one.

AI Assistant · pastoral reflection

What you've named is not a credit-scoring problem. It's a discipleship problem dressed up as a risk problem. The institution wants protection; the borrowers need presence. Before you ask whether AI can predict default, ask whether your officers still have time to visit a borrower's farm. The most accurate predictor of repayment is usually the relationship — and that's the thing automation will quietly erode.

Frame the decision

Tool / system

An ML model trained on five years of loan history to predict repayment likelihood, scoring every new loan application for our loan officers in Kenya, Tanzania, and Uganda.

Problem & people

Loan officers like Grace in Kisumu are managing 320 borrowers and can't tell which ones are at risk until they default. Leadership wants a screening signal so officers can intervene earlier. Real users: 47 loan officers across three countries; ~14,000 active borrowers, mostly smallholder farmers and market vendors.

Reject

Soft reject
Human / process / policy alternative?

Yes — partly. Smaller caseloads would solve more than a model would. The real bottleneck is that we cut field officer headcount in 2023 to reduce overhead. We're trying to use AI to recover what we gave away in staffing.

What is the underlying need?

Stated problem: 'we can't predict defaults.' Real problem: 'we've eroded the relational infrastructure that made our model work, and now we're rationalizing the erosion with technology.'

Who would be harmed if it fails?

Borrowers flagged as 'high-risk' by an opaque model would be denied loans without explanation. In peer-group lending, one denial cascades — the whole group's standing weakens. We'd be encoding a drought year into a permanent credit memory.

Non-technological intervention?

Yes. Rehiring two field officers per region. Restoring monthly borrower group visits. Adjusting repayment schedules seasonally instead of monthly. We haven't tried any of these because they cost more than the model would.

AI Assistant · final reflection

This team did the work most teams skip — they let Receive end the conversation. CGAP's framework is freely available, peer-reviewed, and field-tested. Building their own version would have been ego-driven, not service-driven. The discipline here is naming that 'we could build this' is not the same as 'we should build this' — especially when faithful work already exists in the public commons.

Walk your own

Start from this example as a draft. Every field will be pre-filled — edit freely. Your own context will surface as you go.

These examples are illustrative. Real discernments will be more complex, more painful, and more specific to your context. The library helps you see the shape of the work.

Examples paraphrased from common patterns observed in faith-based development AI projects. Not based on any single real organization.