Table of Contents
Most of what we hear about artificial intelligence comes through the lens of advanced economies: the race to build the most powerful models, the fear that white-collar jobs will disappear. That framing misses AI’s biggest opportunity. In much of the developing world, the problem is too few skilled workers, not too many — too few doctors to diagnose disease, too few agronomists to advise farmers, too few officials to deliver services. AI, which has reached lower-income countries 40 times faster than the internet did, can extend scarce expertise to far more people in far less time.
This blog draws on the World Development Report 2026: The Promise of Artificial Intelligence, which examines what AI means for development through the lens of three interconnected questions: what AI can do, who controls it, and what it needs to work.
Capabilities: a century of progress in a decade
By guiding decisions where expertise is scarce, AI can do in a decade what would otherwise take a century. Consider a few examples:
- Lower-income countries have fewer than 20 radiologists per million people, roughly a fifth of what is considered adequate, and AI is beginning to close the gap.
- In Bangladesh, AI-assisted screening raised the number of patients checked for diabetic retinopathy in a day by 40 percent.
- In India, AI-supported weather forecasts tailored to local rainfall patterns and languages reached more than 30 million farmers, giving most earlier warning of monsoons than ever before.
- In Ghana, the Rori math tutor, delivered by text message on basic phones, produced a year of learning gains at US$5 per student.
Concentration: access today, but on whose terms?
A few companies in a few economies control the most advanced AI models, the chips they depend on, and the data centers that run them. This cuts both ways. Developing economies can customize world-class AI without spending billions to build it from scratch. But the terms matter.
Many AI providers price access below cost to drive adoption, gather data that improves their models, and embed themselves in how firms and governments work. Early ride-hailing platforms followed the same approach: low fares built a user base, and prices rose once the platform became the default. Today’s low prices with AI products may likewise be an introductory offer. Once adoption crosses a threshold and switching becomes costly, developing economies—with limited bargaining power and thin competition-policy capacity—are the most exposed to becoming dependent on a few AI providers.
Complements: the bigger constraint
Without reliable infrastructure, foundational skills, and capable institutions, AI’s benefits will be slow to arrive. Its expected productivity impact in advanced economies is twice that in developing economies, and three-quarters of that gap reflects differences in adoption. Harms spread quickly too: countries with little capacity to detect malicious uses of AI are more vulnerable to cyberattacks.
What governments should do: adopt, adapt, and then advance
Pushing the AI frontier is enormously expensive: hyperscalers’ capital expenditure is projected to exceed US$750 billion in 2026, more than the GDP of many countries. Adopting existing AI is the natural starting point, but off-the-shelf tools are not enough. A clinical decision-support tool trained in high-income countries can recommend what does not fit local medical norms.
Adaptation grounds AI in local realities, which drives adoption.
Adapting AI is also a path to advancing it: India’s Sarvam AI and Africa’s InkubaLM are language models built from scratch for local languages. Such efforts, along with open-weight models, give countries alternatives — and in a concentrated market, alternatives are the foundation of bargaining power.
Make cost-conscious choices
Governments shape AI’s impact as enablers, users, and regulators. In each role, they should plan for how costs evolve after introductory prices end, not just costs today.
As enablers, governments must build AI’s digital and analog foundations. In Sub-Saharan Africa, nearly a third of rural schools lack reliable electricity, more than two-thirds lack dependable internet, and nearly nine in ten 10-year-olds cannot read a simple text. Data matters too: almost half of internet-derived training data is in English, while more than 2,000 languages are spoken in Africa alone. Local-language data is both a necessity and an asset to bargain with.
As users, governments need frameworks to procure, evaluate, and scale up AI. Only 17 of nearly 10,000 published studies on AI in health care show causal evidence of impact in low-resource settings, and 80 percent of digital agencies in lower-income countries have no or only basic ways to evaluate AI tools. The most immediate opportunity lies in proven back-office tools for predictive analytics and decision support (see figure 1). Procurement should guard against lock-in through price protections, data portability, exit clauses, and interoperable systems.
