The World Bank's World Development Report 2026 recommends that developing countries adopt and adapt existing AI tools rather than develop their own models. The first time I read that sentence, something in me bristled. We are being told not to build. To take what already exists. To stay on the side of those who use.
I read it again. I had stopped on the wrong verb.
There are two of them in that recommendation. Adopt, and adapt. We hold on to the first because it flatters or wounds our pride. The second does all the work, and almost nobody talks about it.
Adapting is not giving in
Adopting a model means taking an engine. Adapting means building the vehicle around it, for the roads here. The engine, I will not be manufacturing this year. Togo is not ranked among the most advanced countries in this field; it sits far behind the United States, China or the United Kingdom. That fact no longer offends me. It frees me from the worst possible project: retraining a model with nothing comparable in resources against those who have already done it.
Because an engine on its own carries nobody. What is missing is the body of the vehicle. And the bodywork is either drawn here or it is not drawn at all.
Where local value is created
I worked alongside a seamstress who takes all her orders on WhatsApp. Measurements, fabrics, deadlines, deposits: everything lives inside a conversation, mixed in with voice notes and photos. No model in the world arrives with the solution to that problem. The model can read text. It does not know what a deposit means in her shop, or why a customer who writes "same as last time" has just handed over half the specification.
That is exactly where the work is. The real language of the exchanges, not the French of the textbooks. The vocabulary of a mechanic down the street who names his parts differently from the catalogue. The rhythm of a trader at the Grand Marché who writes up his sales in the evening, by hand, because that is how it has been done for years. Field data that nobody else owns.
That is what adapting is. It is not subcontracting. It is half the product.
What I would put in place on Monday morning
Choose a single process, the most painful one. Not an AI strategy. A process: the quotes, the follow-ups, the order entry. The one that costs the most hours. We measure what it costs today, in minutes, over a week. Without that starting figure, everything else is an opinion.
Write the vocabulary before writing the code. One page, two columns: the word used in the trade, what it actually refers to. I do this with every client now. It is the most profitable document of the project, and it costs one hour.
Treat the outage as the norm. Not as an exception to be handled later. Whatever has been entered must survive a loss of network and an abrupt shutdown. A burnt-out machine taught me more about a system's resilience than any documentation ever did.
Wire up payment from the very first version. Here, a tool that does not talk to mobile money stays a demonstration. The integration is not a bonus module, it is the condition for usage to hold beyond the second week.
Keep the interface simpler than the model. Perceived value does not come from the power behind it. It comes from the number of steps you have removed.
Acknowledging what already exists
In parallel, the report insists on investment in digital infrastructure and administrative capacity. And it has to be said: we are not starting from nothing. The Novissi precedent exists, and it is an achievement. A data analysis laboratory is in development. These are real prerequisites, and I would rather name them before listing what is missing.
What is missing is named too: the cost and stability of electricity, rural connectivity. Those two obstacles will not be settled from my office. But they change the way I design. A tool built for a permanent connection and uninterrupted power is not a tool for here — it is a tool for somewhere else, imported without being adapted.
And that is the only mistake the recommendation still leaves us room to make: adopt without adapting, then be surprised that nothing takes hold.
The model will come from elsewhere. The use of it is written here.