Everyone Can Buy the Matatu. Almost Nobody Gets the Route.

I was wrong, and I was wrong loudly, in a WhatsApp group, with a voice note.

The claim I made in 2024 was that AI would flatten opportunity for anyone with a laptop and a data bundle. A designer in Kisumu, a paralegal in Kampala, a bookkeeper in Rongai, a Kenyan nurse's daughter in Manchester who codes at night. Give them the same model that the Palo Alto person is using, and watch the salary gap start closing like a matatu door.

The tools arrived. Some of the gap closed. And then I watched a friend with genuinely excellent output spend eleven months failing to get paid international rates, while a mediocre operator with a Stripe account and three referrals cleared twice as much.

So I have been asking the same annoying question ever since. Not who has the tool. But who controls the road the tool drives on.

First, the strongest version of the optimistic case

Let me make the argument I no longer fully believe, but properly, without strawmanning it. Because it is a good argument and half of it is true.

Historically, the thing that kept a talented person in Nakuru from earning like a talented person in Toronto was not intelligence. It was access to accumulated professional context. The Toronto person had colleagues who had seen a hundred contracts. They had templates, house style, a senior who caught the mistake before the client did. That surrounding scaffolding was worth more than the individual skill, and it was geographically trapped.

Large models unbundled that scaffolding and put it in a chat box. The senior reviewer is now available at 3am for the price of a bundle. Legal drafting, financial modeling, code review, English that reads like it was born in it. The apprenticeship that used to require proximity to a rich country's office became downloadable.

This is not hype. I have seen a self-taught developer in Mombasa ship a payments integration that a mid-tier agency would have quoted six weeks for. I have seen a small logistics firm in Tema replace a consultant's forty-page report with an afternoon and a good prompt. The competence is real. The output is real. The optimists are describing something that genuinely happened.

And for the diaspora it is sharper still. The Somali accountant in Minneapolis who was quietly capped by an accent on the phone can now write proposals that arrive with no accent at all. The friction that discrimination lived inside, the polish gap, the register gap, the does-this-person-sound-like-us gap, that friction thinned.

Good argument. Now let me break it.

The matatu was never the expensive part

Anyone who has been near the industry will tell you this for free. A matatu is a purchasable object. Save, borrow from the chama, take a loan against your mother's plot if you are feeling brave, and you can own fourteen seats and a sound system.

What you cannot buy at Kariobangi is the route.

The route has a SACCO. The stage has a marshal. The corridor has an informal understanding about how many vehicles run it and who is permitted to add another. There is a daily fee. There is a person who decides the queue order at rush hour, and that person's decision is the difference between four full trips and two half-empty ones. The vehicle is capital. The route is permission. Capital is cheap and getting cheaper. Permission has never once gotten cheaper in the history of anything.

The model is the matatu. Distribution, payments, identity, and default trust are the route. We got very excited about a price drop in the one part of the business that was never the bottleneck.

So look at what actually stands between a brilliant worker in Accra and a brilliant worker's income.

Payments. The client's procurement system accepts vendors from a list of countries. Your country is not a philosophical exclusion. It is a dropdown. Someone maintains that dropdown. Who?

Identity. The platform wants a verified business entity, a tax identifier it recognizes, a bank in a jurisdiction its compliance team has heard of. The AI wrote you a flawless proposal and the flawless proposal died at KYC.

Distribution. Clients do not run open auctions for talent. They ask three people they already know. Referral is the stage marshal of white-collar work, and it does not read your portfolio.

Defaults. When the procurement officer types the request into their own assistant, which vendors surface? Whoever the model was trained to consider reputable. Whoever bought the placement. Whoever existed in the index. This is the quiet one, and it is going to be the loud one by 2028.

Every one of these is a route toll. And notice the shape of the change: when the vehicle gets cheap, the toll goes up. It has to. The scarce thing reprices itself. Ten thousand new matatus on the same corridor does not create ten thousand new incomes. It creates one very wealthy stage marshal and a lot of drivers fighting over the same rush hour.

The optimists were right that AI unbundled the scaffolding. They just assumed the scaffolding was the moat. It was the inventory.

Which means the diaspora story is not the one we tell

Here is what I did not expect. The diaspora advantage in this new arrangement is not skill and it is not proximity to opportunity. It is route access. A bank account in a jurisdiction the dropdown likes. A phone number that does not fail the fraud filter. Twelve years of colleagues who will take the call.

That is a genuinely valuable asset and it is also, awkwardly, an arbitrage. The cousin in Reading is not out-thinking the cousin in Ruiru. She is running a legal, ordinary, deeply human toll booth on a road the cousin in Ruiru is not allowed to enter. Both of them know this. Neither of them says it at Christmas.

I do not think this is a scandal. I think it is a map. Because if the bottleneck is the route, then building matatus faster is not the answer, and no amount of upskilling fixes a dropdown menu.

What builders should actually be doing

Stop competing on the layer that dropped to zero. Everybody has the model. Nobody's advantage lives there anymore, and any business whose entire pitch is we have access to a good AI is a business selling seats on a road it does not own.

Compete on the layer that is still permissioned. Build the SACCO, not the vehicle. Concretely: the collective invoicing entity that lets fifteen freelancers present as one compliant vendor. The reputation ledger that travels with a worker across platforms. The local payment rail that makes a Nairobi freelancer legible to a Rotterdam procurement system. The trust infrastructure, in other words, which is boring, which is unglamorous, which is where every shilling actually sits.

And build your own defaults while defaults are still cheap. Own the phone number. Own the customer list. Own the file. If your entire client relationship lives inside somebody else's platform, you are renting the stage and you will find out the fee when they decide to tell you.

A small note from the lab

Half of why we started building a business agent that answers a real phone line, rather than a chatbot that lives in an app, is precisely this. The phone number is a route you own. Nobody can deprioritize you in the ranking of a number that is already saved in a customer's contacts. It is unfashionable infrastructure. That is usually the tell.

The takeaway

I still believe AI is the most serious economic tool to reach this continent since mobile money. I was just measuring the wrong thing, celebrating a collapse in the price of engines while the price of roads quietly tripled.

What to do this week: take your best client relationship and ask one question about it. If the introduction, the payment, and the ongoing conversation all flow through a channel you do not control, you are not a business. You are a very good driver on someone else's route. Pick one of those three and move it onto ground you own. Start with payment. It is the one they take away first.

Buy the matatu, by all means. Everyone should. Just do not confuse owning a vehicle with owning a livelihood, and do not confuse a cheap engine with an open road.

Related reading from the lab: our piece on why AI infrastructure debates are really power-bill debates; the field notes on user-approved files and who owns your data; and the argument for judging every new AI tool by whether it survives the market stall.