The World Bank says the AI that matters most for poorer countries is small, cheap and local, running on phones people already own, even with weak signal.
The World Bank wants poorer countries to stop chasing giant AI models and build small AI instead: narrow, cheap tools that run on ordinary phones, cope with weak signal and speak local languages. Chief knowledge officer Paschal Donohoe says this could compress centuries of development into decades.
It is already being tested. Teachers in Ethiopia get lesson feedback from an app, students in Lima use an AI career guide, and Ukraine plans a math tutor in its school app before year's end, aiming for 2,000 schools by mid-2027.
The next test is whether these pilots can scale beyond a few hundred users.
The most useful AI for billions of people may never touch a giant data center. The World Bank is now urging developing countries to build what it calls small AI: narrow, cheap models that live on an ordinary phone, speak the local language and keep working when the signal drops.
Why should anyone in the US, UK or Canada care? Because it is a direct challenge to the idea that AI progress means ever-bigger models and ever-bigger power bills. The bank argues that a farmer with a basic Android phone and a patchy connection can get real value from AI today, and that this, more than chatbots, is how the technology could change incomes and jobs across much of the world.
Small AI is a model trained on a narrow set of data to do one job well, rather than a general assistant that tries to answer everything. Think of a tool that reads a soil test and explains the result, or one that gives a teacher feedback on a recorded lesson.
Three traits define it in the World Bank’s framing. It runs on a phone people already own. It copes with low or on-and-off connectivity. And it works in the community’s own language, not just English.
Paschal Donohoe, the bank’s chief knowledge officer and a former Irish finance minister, made the case during the annual World Bank and International Monetary Fund meetings in Bangkok in mid-October. His pitch was bold: development that might once have taken centuries could, with the right tools, happen within decades.
The idea is not just a slogan. Donohoe pointed to Indian farmers in AgriConnect, a World Bank agriculture program, who wanted soil analysis explained in their own language. There will never be enough agronomists to visit every small farm in India, he argued, so a phone-based tool has to fill the gap.
Education is where the bank has shared the most detail:
Similar pilots are underway in Zambia, El Salvador, Senegal and the Seychelles. The bank’s stated aim is to help reach 13 million young people by 2030 by moving the tools that work beyond the pilot stage.
Small AI is designed for weak infrastructure, but it cannot run on none. The bank’s own world development report, published in August, found that almost a third of rural schools in sub-Saharan Africa had no electricity, and more than two-thirds lacked a steady internet connection.
Kristalina Georgieva, the IMF’s managing director, spelled out the preconditions on a Bangkok panel: electricity, digital networks, people ready to adopt the tools, and governments able to protect users from bad actors. Get those basics right, she argued, and AI could spread as widely as the mobile phone did. She used the phrase “as democratic as the cellphone” to describe that hope.
Language is another hurdle. World Bank President Ajay Banga has said the Gates Foundation is helping the bank adapt simple AI tools to local languages, which he described as essential if people are going to use them at all. Funding is the third. Banga has made pulling private investment into emerging economies a central goal, at a time when several rich countries, the US and UK among them, are cutting aid budgets.
The bigger worry behind all this is employment. About 1.2 billion young people are expected to enter the workforce in developing countries over the next decade. Georgieva warned that AI is likely to boost or wipe out around 40% of jobs worldwide over time, and that the first people to feel it will be young workers.
Her point travels well beyond poorer countries. Entry-level roles built on routine, repetitive work have long been the first rung of a career. If AI removes that rung, she said, young people lose the usual way up. That concern is just as real for a graduate in Toronto, Leeds or Ohio as it is in Lima.
The small AI pitch is the hopeful flip side: tools that make a teacher, farmer or student more productive, rather than replacing them. For readers in richer countries, the takeaway is that the most important AI story of the decade may not be the flashiest model, but the cheapest one that works on a five-year-old phone.
The bank has since taken the message on the road. Banga used a Global AI and Digital Summit in Seoul, titled Small AI. Big Impact, to launch a digital strategy built around jobs, with panels of ministers and founders explaining how their pilots could grow.
The test now is scale. A coach for a few hundred teachers is easy to praise. Rolling it out to whole school systems needs steady power, cheap data, local-language models and money that keeps flowing after the pilot ends. Watch the Ukraine tutor rollout and the 2030 youth target as the clearest signs of whether small AI can grow up.
It means narrow AI models built for one local task, such as soil advice or teacher feedback, that run on a basic phone, cope with weak internet and work in local languages.
It overlaps. The World Bank’s focus is less on the chip and more on purpose: cheap tools that work where cloud access is unreliable.
The current pilots target developing countries, including Ethiopia, Peru and Ukraine. No launches in North America or the UK have been announced.
You're reading the quick version.
Read next
Data center ballot measures: voters back limits in 33 places
From outright bans in Ohio to a 180-day pause in Maine, voters chose limits almost everywhere. Power bills, not AI itself, drove the vote.
From outright bans in Ohio to a 180-day pause in Maine, voters chose limits almost everywhere. Power bills, not AI itself, drove the vote.
Openreach says the old copper phone network shuts on Jan. 31, 2027, with no further delay. About a million lines still need moving. Here's what to check.
Beijing proposed new limits on exporting AI, then cleared its two biggest challengers to keep selling abroad within days. Here's what the rules do and don't touch.
The Daily Brief
Every morning, the news that matters from AI, phones, apps and the people shaping tech. Explained in plain English. Free.
One email a day. No spam, unsubscribe anytime. Privacy policy
