World Bank links lower AI exposure to deeper dependence

Summary

The World Bank links lower generative-AI job exposure in developing economies to their agrarian and small-enterprise structure.
The World Bank identifies shortages of highly skilled workers as a channel for AI-driven productivity gains.
Indermit Gill warns that AI could weaken outsourcing as a route to middle-class employment.
The World Bank connects foreign control of advanced AI systems to dependence on technology developing economies do not control.

The way this story is framed determines whether readers see AI as a development shortcut or as another source of economic dependence. The World Bank’s assessment supplies evidence for both readings. Generative AI threatens 4.5% of jobs in developing countries, compared with 14.2% in rich nations. But lower exposure does not mean equal gains. Meaningful productivity improvements could reach 16.2% of jobs in developing economies, compared with 18.7% in high-income economies. The difference turns on who can use the technology, which jobs remain available, and who controls the systems underneath it.

Why the exposure gap does not settle the outcome

The article attributes the gap to the greater weight of agriculture and small enterprises, but it does not establish a more specific mechanism for how those sectors reduce exposure. In that sense, the 4.5% figure describes the structure of the economy as much as the protective effect of the technology.

The comparison with rich countries is stark. Generative AI threatens 14.2% of jobs in rich nations, where more employment involves the office-based cognitive tasks that current systems can assist with or replace. Developing economies therefore begin with less immediate disruption. That is a measure of exposure, not a distribution of benefits.

A relationship map shows why the figures need to be read together. Economic structure helps explain lower job exposure. Shortages of skilled workers create a separate opening for AI. Limited computing power, unreliable electricity, poor internet service and insufficient local-language data constrain the gains that follow. Foreign control of models and cloud services adds another dependency after adoption begins.

The World Bank’s central claim is conditional. AI can help less experienced workers perform more advanced cognitive tasks and shift some work away from scarce specialists. That mechanism could raise productivity without requiring poorer countries to reproduce the expensive frontier-AI infrastructure being built in the United States.

The lower-cost route to productivity

The infrastructure distinction matters because the development paths are not identical. The frontier-AI buildout described in the article involves large data centers, substantial computing capacity and dependable electricity, while narrow applications may operate with more modest requirements. Narrow AI applications can address specific tasks without requiring poorer countries to reproduce the entire frontier-AI supply chain. Indermit Gill wrote that there is “no need for massive investments in data centers or large language models” when AI is used for narrow problems.

Gill also wrote that “AI can help solve important problems involving narrow tasks even when local computing power is limited, electricity is unreliable, and internet service is dodgy.” That is a practical proposition, not evidence that deployment will automatically spread. A tool that works in a connected urban firm may remain inaccessible to rural businesses, public agencies or workers who lack the necessary equipment and training.

The productivity figures show the current limit. The World Bank calculates that meaningful gains could reach 16.2% of jobs in developing economies, compared with 18.7% in high-income economies. The gap is narrower than the job-exposure comparison, but it remains a gap. The World Bank says wider access to computing power and more local data, including data in local languages, could reduce it.

Local data is not a decorative supplement. AI systems learn from the material used to train them. Where local languages, institutions and working conditions are poorly represented, imported systems may perform less effectively or require adaptation that local users cannot afford. The result could be a technology that is broadly available in principle but productive mainly for firms already connected to global infrastructure.

The employment ladder at risk

The aggregate exposure number also conceals concentrated losses. Indermit Gill identifies call centers and entry-level jobs in software, finance and business services as vulnerable. Those industries have provided a route into middle-class employment for workers in some developing economies. If AI reduces demand for those entry-level roles, lower overall exposure could coexist with a serious loss in upward mobility.

Gill wrote that AI “could close off a promising route to middle-class employment in many developing economies.” The concern is not simply that some jobs disappear. It is that the first rung of an employment ladder may be removed before workers can move into higher-value occupations.

The evidence supplied here does not establish how quickly that shift would occur or whether displaced workers could move into other sectors. That remains an open question. The direction of the risk is nevertheless clear: a country can avoid large-scale economy-wide displacement while losing a strategically important category of jobs.

This is why the 4.5% figure cannot serve as a general forecast of social impact. A small percentage concentrated in call centers, back-office services and entry-level professional work could matter more for development prospects than a larger but dispersed figure elsewhere. The relevant question is not only how many jobs face exposure, but which jobs provide access to rising incomes.

Whose systems developing economies will use

The World Bank places a second condition beside the labor question: control. The bank says the most advanced AI systems are controlled by a small number of companies, mainly in the United States and China. Developing countries may therefore gain access to useful tools while remaining dependent on technology they do not control.

That dependence can arise even without a new data-center buildout. A country may avoid the cost of replicating the entire AI supply chain and still rely on foreign models, cloud services, software updates and technical standards. Narrow applications can lower the entry cost without eliminating the underlying supplier relationship.

The World Bank recommends buying models, cloud services and other tools from many countries. Gill adds that the systems should work together and be replaceable without requiring an entire rebuild. Supplier diversity and interoperability would make dependence less rigid, but the material provided here does not demonstrate how effective those safeguards would be in practice.

The choice is therefore not between adopting AI and remaining independent. Adoption itself creates a procurement question. Governments must determine whether imported tools can be switched, whether contracts preserve that option, and whether local data and language needs can be served without locking users into one provider or national technology ecosystem.

What to watch next

Over the next five to ten years, the central divide will be between adoption that broadens capability and adoption that narrows economic options. Assisted catch-up would occur if narrow tools help less experienced workers, computing access expands, local data improves and firms outside major cities begin using adapted systems. An outsourcing squeeze would occur if call-center and entry-level professional work contracts before workers can move into alternatives.

Dependent adoption could accompany either outcome. Developing economies may record productivity gains while relying on systems controlled mainly by companies in the United States and China. The relevant indicators are supplier concentration, cloud-switching costs, procurement requirements for interoperability and the availability of tools that support local languages.

Readers can carry four questions into the next AI story: How many jobs are exposed, and which jobs are they? Who receives the productivity gain? What infrastructure and local data does adoption require? Which companies and countries control the systems that remain after deployment?

The development bargain will be decided not by exposure alone, but by whether AI expands workers’ capabilities without closing the employment ladder or hardening foreign control.

Analytical techniques used in this piece

This analysis applies the methods below. Each links to a short, plain-English explainer you can read and reuse.

Relationship Mapping: Extracts the network of ties among people, institutions, and entities.
Scenario Planning: Builds a small set of distinct, plausible futures to plan against.
Wicked Futures: Explores a long-horizon, deeply entangled future with no clean resolution.