Leapfrogging Is Not a Law of Nature
Every deck about African technology has the same slide: we skipped landlines, we skipped bank branches, therefore we will skip whatever is being sold. The slide is true. The lesson people take from it is backwards. Leapfrogging is not a law of nature — it is a checklist, and most projects fail it.
Every deck about African technology has the same slide.
Africa skipped landlines and went straight to mobile. Africa skipped bank branches and went straight to mobile money. Therefore Africa will skip whatever this deck is selling.
I have used that slide. It is not false.
But the lesson people take from it is backwards, and in AI that is going to be expensive.
What actually happened in Kenya
The story gets told as though being behind was itself the advantage. No legacy to defend, so the new thing walked straight in.
That is not what happened.
Before M-Pesa, Kenyans were already using prepaid airtime as an informal currency. You bought credit, sent it to a relative, they sold it on for cash. The behaviour existed and it was widespread. Safaricom did not invent demand for remittance. It noticed a workaround its own customers had built and shipped the sanctioned version.
Three other things had to be true at the same time.
Safaricom held a dominant share of the market, so the network effect existed on day one instead of after a decade of coalition-building.
The Central Bank of Kenya allowed a telecom company to hold customer funds. Most regulators, then and now, would not have.
And Safaricom built an agent network — a large distributed layer of small shops handling cash in and cash out, staffed by people, requiring float management, training and fraud control.
That last one is the least glamorous part of the story and the most important. It never appears on the slide.
So the pattern is not "we were behind, therefore we jumped ahead." It is far more specific: an existing behaviour to ride, dominant distribution, a regulator willing to permit it, and someone doing unglamorous physical work at the last mile.
Leapfrogging is not a law of nature. It is a checklist, and most projects fail it.
Where the alignment is missing, nothing leaps. General e-commerce has not leapfrogged across most of the continent, and not for lack of capital — it runs into addressing, payment trust, returns and logistics, none of which a better app fixes. Plenty of off-grid solar failed too, and mostly not on the technology. It failed on maintenance and collections, which are agent-network problems wearing a different hat.
So what is AI riding?
If the checklist is right, the interesting question is not which model is best. It is: what existing behaviour does this attach to?
In Ethiopia, the answer is messaging. Not apps. Messaging.
Some numbers from my own work, small enough that you should treat them as one data point. Lomi is an AI study tool for students preparing for the national school-leaving exam. There are about four thousand students on it. Roughly seventy per cent arrived through Telegram rather than the web app, and stayed there.
The web app is better. It is not close. It renders mathematics properly, shows diagrams, holds full lessons, has a dashboard. The Telegram bot is a cramped, constrained version of the same thing.
The bot wins anyway.
It wins because it is already open on the phone. Because there is nothing to install. Because it costs nothing to start. Because it survives a bad connection. Because it answers in the language students actually speak.
Every one of those is a distribution property. Not one of them is an intelligence property.
The most over-rated variable in African AI is the model
I have swapped underlying models, rewritten prompts, and improved how the system pulls grounding material from past exams and textbooks. Those produced real, measurable gains in answer quality. I can show you the evaluations.
Their effect on whether students actually studied was close to nothing.
What moved usage was whether a message arrived in the evening rather than the morning. How near the exam was. Whether the thing loaded on a weak connection. Whether the first screen asked a student to type or let them tap.
That was not a comfortable thing to learn as an engineer, and it is the most useful thing I know.
It is also good news, and I want to be clear about why.
Model quality is a purchased input. It arrives roughly equally for everyone, it improves whether or not you do anything, and nobody on this continent is going to win by training a better one. If model quality were the deciding variable, we would be permanently downstream of decisions made in California.
It is not the deciding variable. Distribution is. And distribution is a thing you can actually build an advantage in, from here, with no permission from anyone.
What follows
The AI companies that matter here will not be the ones with the best models. They will be the ones that solve the equivalent of the agent network — the unglamorous operational last mile.
For M-Pesa that was shopfronts and float. For AI I think it is some combination of: living inside the messaging app people already use, tolerating a network that drops, speaking to people who do not type quickly in English, and supplying a reason to open the thing today rather than eventually.
That last one is the piece I understand least, and I am increasingly convinced it is the whole game. I will write about it properly once I have data rather than a hunch.
Two things would change my mind. If a materially better model produced a step change in engagement with no distribution change, my ordering is wrong. If a product with weak distribution but excellent output grew anyway, the checklist is missing something. I am watching for both.
Three questions
When the next deck tells you the continent will skip a step, ask three things.
Which existing behaviour is this riding?
Who is the regulator, and what have they actually permitted?
And who is doing the boring physical work?
If there is no answer to the third one, it is not a leapfrog.
It is a hope.