I started my career when the internet in Dakar was something you bought by the hour, in a cybercafé, on a machine still warm from the last customer. Since then, I have been promised the end of the world as we know it roughly every four years.
Web 2.0 was going to change everything. Then came social networks, when every pitch deck I saw between 2008 and 2012 seemed to have the word "community" on slide two. Then there was SoLoMo, if you were around for that one: social, local, mobile. Silicon Valley invented a word that sounded like a dish my aunt makes and declared it the future. After that came an app for everything, chatbots, and eventually blockchain, which at various points was going to fix payments, land registries, elections, and possibly the weather.
The cycle was remarkably consistent. A wave would break in California and reach us eighteen months later. Conferences would rename themselves, investors would start funding whatever carried the new keyword, and founders — good founders, people I respect — would bend their companies to fit the theme. Eventually the money would move on, the wave would flatten, and everyone would quietly remove the buzzword from their deck. Watch that loop enough times and skepticism becomes your default setting. Ask my team; I am not an easy man to excite.
So believe me when I say that I have spent the last three years looking for reasons to dismiss AI, and I have run out. The difference, as I see it, is that most of the previous technology waves changed how we communicated about work, sold work or moved information around. AI is beginning to change the cost of actually doing the work, and I think that puts it in a different category.
There is another reason I think this wave matters particularly for Africa. Many of the previous technological revolutions quietly assumed infrastructure that we did not yet have. E-commerce assumed widespread card payments and reliable street addresses. Streaming assumed cheap bandwidth. The sharing economy assumed a large base of underutilized assets that could easily be discovered and transacted online. As a result, many of those waves reached African markets late, in diluted forms, or remained more aspirational than transformative.
AI has a much lower threshold for adoption. In many cases, all it asks for is a phone and a question. It can read the blurry scanned PDF, help draft the customs letter, extract information from an invoice, or answer in French at midnight without anybody having designed a formal rollout programme for it. Nobody organized AI adoption for the tailor in Sandaga or the transit agent at the port. People found it themselves, the way people here tend to find tools that are genuinely useful. For the first time in my career, I have watched a major technology begin to adopt itself.
That is why I believe the wave is real. But believing in a technology wave is not the same thing as having a strategy for it. Twenty years of watching technologies arrive, peak and disappear have left me with three convictions about where I think the durable opportunities will be.
First: don't build anything a keynote can delete
Every time one of the major AI labs holds a launch event, there are startups somewhere discovering that a product they spent months building has just become a feature of somebody else's model. This is not entirely new; platform shifts have always created this risk. What is different now is the speed at which it can happen.
If your product is essentially a thin interface over somebody else's model, a significant part of your roadmap is being written in San Francisco by people who do not know your company exists. That does not mean foundation models are not enormously useful. It means that access to a model is not, by itself, a defensible business.
The more interesting question is what you can own that a model update cannot easily absorb. That may be a deeply embedded workflow, distribution, proprietary data, customer relationships, or operational knowledge that is difficult to reproduce. In my view, those assets need to come first. The AI should make them more valuable rather than be the only reason the company exists.
Second: build where the internet can't see
The large models are trained primarily on information that has been digitized, published and made accessible. A significant part of African economic life does not meet those conditions. We operate through WhatsApp voice notes, paper waybills, informal addresses, phone calls, handshakes and cash. Many of the decisions that determine how a business actually functions have never appeared in a database, much less in a training corpus.
That creates an interesting inversion. What has traditionally been described as Africa's digitization problem can, in some cases, become a source of competitive advantage for the companies doing the difficult work of capturing that information. A frontier model may know very little about it, but a company embedded in the workflow can learn a great deal.
This is why I continue to be attracted to boring problems. I mean genuinely boring problems: problems that are so local and operationally difficult that somebody sitting thousands of kilometres away is unlikely to spend a weekend trying to solve them, but large enough that solving them properly can move an industry. Buzzwords attract competitors. Boring operational work tends to repel them.
At Chargel, we have spent four years doing work that few people would describe as glamorous: digitizing freight one truck, one corridor and one paper waybill at a time. Moustapha wrote in our last post that the future of African logistics will not be the autonomous truck; it will be coordination. One consequence of building that coordination layer is that we have gradually accumulated a view of logistics that is very different from what you can obtain from a map or a public dataset.
Route optimization is a good example. When you first approach the problem as an engineer, the objective appears relatively straightforward: calculate the available routes, incorporate distance and estimated journey time, and recommend the optimal one. In practice, we learned that a driver would sometimes ignore the route our system considered best and take a longer one. The natural engineering response is to assume that the driver is making a suboptimal choice. Once you spend enough time understanding what happens on the ground, however, you discover that the driver may know something the system does not.
A road that appears perfectly usable on a map may become a poor choice during the rainy season. A theoretically faster corridor may involve a border crossing where controls and queues regularly add hours to a journey. A longer road may simply be more reliable for a heavily loaded truck. Conditions can also change much faster than they are reflected in digital maps or formal datasets. In those circumstances, recommending the mathematically shortest route repeatedly does not make the recommendation more correct. The driver chooses the route that works in the real world, and very often he has good reasons for doing so.
For me, the more interesting technical problem became how to capture what the driver knows and make the system better because of it. If drivers consistently reject a route that the algorithm considers optimal, the rejection itself contains information. If journey times change predictably during the rainy season, that pattern can be learned. If trucks systematically accept additional kilometres to avoid a particular border crossing, there is information there too. Over time, the objective becomes more sophisticated than calculating the shortest route. You want to understand which route is most likely to work for a particular truck, on a particular corridor, at a particular time, under the conditions that actually exist.
The same principle applies across the network. We can observe how long trucks actually wait at the port of Dakar, how corridor prices move when the harvest comes in, which transporters reliably show up when they say they will, and how operational behaviour differs from what the formal process says should happen. Much of this information does not exist on the public internet because nobody had a reason to collect it before.
We did not set out four years ago to create an AI dataset. We set out to solve the everyday coordination problems involved in moving freight. But solving those problems creates data, and more importantly it creates context around that data. That context is increasingly important because intelligence is only useful when it understands the environment in which a decision has to be made.
Third: solve for Senegal, ship to Kampala
I do not think African technology companies should try to compete with frontier labs by training larger general-purpose models. We have neither a compelling reason nor, in most cases, the resources to play that game. The opportunity is elsewhere. There are more valuable problems to solve on this continent than there are people seriously trying to solve them, and many of those problems require knowledge that is difficult to acquire from outside the market.
At the same time, being deeply local does not mean building a product that can only work in one country. One of the things we have learned at Chargel is that African markets are neither identical nor completely different. Roads change, regulations change, border procedures change, languages and payment systems change, and assumptions that work in Senegal cannot simply be copied into Côte d'Ivoire, Kenya or Uganda.
What does repeat is the underlying structure of many of the problems. Supply chains are fragmented, significant parts of commerce remain informal, useful data is poorly structured, and infrastructure often behaves differently in practice from the way it appears on paper. The challenge is therefore not to build something for Senegal and reproduce it unchanged elsewhere. It is to solve the Senegalese problem deeply enough to understand which parts of the solution are local and which parts are invariant. Those invariant elements are what allow a local product to become a continental one.
This is also why I believe the relationship between AI and African technology is more interesting than simply giving African companies access to the same models being used elsewhere. The models themselves are becoming increasingly accessible and, over time, many of their capabilities will become commodities. Context will not. A model can become dramatically more capable without suddenly knowing what happens to a particular road in Guinea during the rainy season, how long a truck really spends at a West African border post, or why an experienced transporter makes a decision that appears irrational in a clean dataset.
For decades, the fact that so much of our economic activity was poorly digitized was an obvious disadvantage. AI introduces the possibility that companies capable of digitizing those workflows and understanding the behaviour behind them can turn that disadvantage into something valuable. The opportunity is not to recreate Silicon Valley products with an African label. It is to build systems that cannot easily be built from Silicon Valley because the knowledge required to make them work is here.
There are already examples of African teams demonstrating how far that approach can go. InstaDeep started in Tunis and became a globally significant AI company by applying serious engineering to difficult problems rather than waiting for somebody elsewhere to define the opportunity.
The window now feels unusually open. Compute is being deployed on the continent, governments are developing AI strategies, and African engineers increasingly have access to the same foundation technologies as their peers elsewhere. More importantly, this platform shift is reaching us while it is still early globally. Historically, we have often received major technology waves secondhand, once the infrastructure had been built elsewhere and much of the economic value had already been captured.
For twenty years, "this time it's different" has been one of the most expensive sentences in technology, which is probably why I have avoided writing it until now. But after spending three years trying to convince myself that AI was simply the next turn of a familiar hype cycle, I no longer think that is the case.
If you are building something boring, local and AI-shaped somewhere on this continent, my inbox is open. And if you think I have finally caught the fever I spent most of my career avoiding, I would be equally interested to hear why. Those are usually the conversations from which the interesting things begin.