Why AI tutors will land harder in African classrooms than anywhere else

There is a familiar argument made by AI tutoring companies built in San Francisco. It goes: every learner deserves a one-on-one tutor; teachers cannot scale; therefore AI replaces the teacher. It is a clean argument and it is also wrong almost everywhere it is sold.

It is wrong because most learners in the world do not want a tutor instead of a teacher. They want a tutor in addition to a teacher who is overstretched. And nowhere is the gap between teacher capacity and pupil need wider than in Sub-Saharan African public education. A typical Zimbabwean primary school class is 50–60 pupils. A typical secondary class around 45. The teacher is doing the best they can. The pupil who falls 20 minutes behind in algebra in week three never catches up — and there is no available adult to spot it.

The leverage opportunity is not “a tutor in every pocket” framed against zero. It is framed against a teacher who is doing six things at once. A well-built AI tutor in this context does three small things very well, instead of trying to do everything badly: it spots the moment a pupil is lost, it generates one good practice problem at the right level, and it lets the teacher see — at a glance — which seven pupils need ten minutes tomorrow morning.

That is not the product the international press writes about. It is the product the schools that have piloted ours are asking us to keep shipping. And it lands because of three things that look like constraints but are actually advantages.

First — bandwidth. A model that has to work on a 256kbps connection in a rural primary school in Mashonaland West cannot afford to be chatty. So we built ours to be terse. It explains in three sentences. It does not roleplay. It does not produce poems unless asked. Constraints made the UX better.

Third — trust. A teacher who has been doing the job for 22 years is not going to be replaced by a chatbot — and shouldn’t be. Our tutor is built to be the teacher’s instrument. The teacher always sees what the tutor said. The teacher can correct it. The teacher’s judgement is the source of truth, and the tutor’s confidence drops when corrected. That is the only design that earns the right to be in the classroom long-term.

There is a longer technical post on the architecture coming — RAG over the syllabus, eval harness on past papers, and the lightweight structured-output layer that turns the model’s reply into a teacher-readable score. For now, the point is the framing.

AI tutoring will land harder in African classrooms than in Silicon Valley ones because the constraint that everyone says makes it impossible — overstretched teachers, big classes, patchy connectivity — is the precise boundary condition that makes the maths work. The product is leverage, not replacement.

If you run a school or a multi-school network and want to talk about a pilot, message us. We are running three at a time, sequencing carefully, and writing up everything we learn.