The World Bank has good news for Rwanda. Only 4.5 per cent of jobs in low and middle-income countries are at risk from artificial intelligence (AI), against 14.2 per cent in rich countries. I read that a few times. By the third read it stopped sounding like good news.

Our jobs are hard for a model to take partly because they sit low on the ladder. Nobody automates work that was never worth much to begin with. "Your jobs are safe” can quietly mean "your jobs are not worth the trouble yet.” Read the number that way and it changes shape.

To be fair to the report, it does not stop at the risk number. It also finds that 16.2 per cent of jobs in developing economies could get a real productivity boost from AI, close to the 18.7 percent expected in rich countries. Its point is that AI will amplify our workers rather than replace them. That is the strongest part of the case and I am not going to wave it away. But amplified with whose tools? A more productive tenant is still a tenant. The gap that matters here is not the one in the risk column. It is the one in the ownership column.

This ‘good news’ came wrapped in the World Bank’s flagship report, World Development Report 2026: The Promise of Artificial Intelligence, released last month. Its advice to countries like ours fits in three words: adopt, adapt, advance. Take the tools that already exist. Adjust them to local needs. Do not burn money trying to build your own frontier models or big data centres. The report is careful and, on its own terms, mostly right. Rwanda even gets a gold star, the only one of the world’s 25 low-income countries with a national AI strategy.

Adapt-don’t-build is prudent economics and a permission slip to rent forever. The value in AI is not in the wrapper we put around it. It is in the model and the machines it runs on, both owned somewhere else. When we adapt, we keep the thin margin. The owner keeps the fat one. We do the last mile. They keep the road.

This month we opened AI partnership talks with Google. More ‘good news’. But a partnership only counts if both sides still own something when it ends. Bring the users and the data, let them keep the model, and what you have is a customer relationship with a nicer name.

So, the real question under the report is not whether the advice is sensible. I’ll be the first to admit that it is. The real question is, for how long will this be sustainable?

Watch where the money and the attention actually go. Adopt is easy, and we are already good at it. Middle-income countries made up half of ChatGPT’s traffic within six months of its launch. Adapt is a real business, a decent one, the one I am in. Advance is the word everyone claps for in a speech and nobody remembers to put it on a budget line. A country can spend a decade as an excellent adapter and wake up exactly where it started, a very skilled tenant.

None of this makes the World Bank wrong to warn us off trillion-dollar models. It is right. We cannot outspend Silicon Valley on compute and we should not try. But there is a lot of ground between building your own frontier lab and renting someone else’s forever.

The reports points at it. Rwanda’s new procurement rules let the government pay for prototypes instead of only buying finished products off a shelf and let the firm keep the intellectual property. They also drop the bid guarantees and track-record rules that always favoured big incumbents, which in practice opens a door for local firms. Now that’s a builder’s mindset.

The reassuring statistic and the reassuring advice are the same trap in two outfits. Our jobs are safe because they are cheap. Our AI plan is safe because it asks little of us. Safe is not the same as sovereign.

I want that gold star to mean something in ten years. We will know it did the day a Rwandan firm sells the thing it built, not the thing it configured. Until then, when someone tells the poor country not to build, it is only fair to ask who already owns everything worth building.

The writer is an MIT-certified AI professional and the Head of Business and AI at Global Kwik Koders.