Why most AI pilots in Africa never reach production

We are regularly asked to look at an AI proof of concept that impressed everyone and then quietly stopped. The model still works; everything around it failed.

One: the pilot was never tied to a decision

A model that predicts something is not a model that changes something. Name the decision, the owner, and what they will do differently on Tuesday morning. If that is not immediate, the use case is not ready.

Two: the data was assumed rather than checked

Pilots run on a clean extract; production reads the real system, where customers appear under three identifiers and records are missing the field you depend on. Budget for entity resolution and quality gates in the programme.

Three: nobody owned it in production

A pilot has an author; a production system needs an owner who watches quality, handles what it gets wrong, and can turn it off. Without that, the first unusual output becomes a crisis.

Four: the cost of being wrong was never priced

A wrong recommendation is an irritation; a wrong triage decision is a clinical incident. Where errors are expensive, keep a human in the loop and treat the model as an accelerator.

What we do instead

We run a short assessment first: rank use cases by payback, check the data honestly, settle the human-in-the-loop question, and agree who owns the system. Then we build the first one behind an interface that keeps the model replaceable.