The Developing World’s Jobs Crisis Was Here Before AI

Developing countries need to create roughly 800 million more jobs over the next decade than they are on track to produce. That gap is already there before a single algorithm has replaced a single worker. Of the 1.2 billion young people who will reach working age over that decade, only about 420 million can expect a job to be waiting. In Africa alone, nearly a quarter of young people will struggle to find work. Robots and large language models are not the reason. The labor markets they are entering were never built to absorb them in the first place. The crisis is decades old. AI is mostly just making it harder to look away.

The two-tier trap

In most developing countries, up to 80 percent of workers are in the informal sector, with no labor codes, no severance, and no mandated benefits. The formal-sector minority enjoys real protections, but the wall between the two tiers is what makes the entire system fragile.

The split is reinforced by how protections are financed. When benefits are tied to formal payroll jobs, hiring becomes expensive. Employers respond predictably: they hire off the books, or not at all. The result is a system that protects a minority generously and leaves the majority to absorb every shock without a floor.

Calling that a design flaw flatters it. It is a structural failure. The informal sector, often romanticized as a hotbed of scrappy entrepreneurship, is largely a product of that structure. As Rafael La Porta and Andrei Shleifer have argued, informal firms are typically small, low-productivity, and run by poorly educated owners. They rarely transition to formality even when registration costs fall. Decades of microenterprise support, business training, and formalization drives have accordingly produced little at scale. Many of these owners lack access to healthcare, education, and skills. Without those, better jobs stay out of reach even when vacancies open up.

AI risks making the trap worse

Sweeping claims that AI will wipe out whole professions overnight have not held up. Radiologists, we were told, were supposed to be gone by now; they are not. Technical capability reaches the labor market only through firms, regulators, consumers, and institutions, and none of them moves at Silicon Valley’s preferred pace. Jobs are bundles of tasks, and automation takes tasks first, not whole occupations. Generative AI is now poking at coordination tasks once thought safely human, which strains the framework but does not break it. Human judgment may be the next place where the AI camel’s nose pokes inside the tent. Because it is still unclear what tasks can be replaced by AI, rebundling tasks into new jobs requires labor markets that can move workers across roles, and split labor markets cannot do that easily.

Daron Acemoglu estimates that AI could add 0.5 percentage points to annual US productivity growth over the next decade—roughly $1.5 trillion in extra US GDP, compounded. Who pockets it is another matter.

A study by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond of 5,179 customer-service agents at a Fortune 500 software firm, 83 percent of them based outside the United States, found that giving agents access to a generative AI assistant raised resolution rates by 14 percent on average, and by 34 percent for novice and low-skill workers. Customer sentiment improved, retention rose, and the AI seemed to spread the senior workers’ know-how to the newcomers. The upside is real. A second study cuts the other way: Voice-based AI interviewers outperformed human ones, displacing one set of jobs while letting firms hire and retain more workers around the AI. Both effects can be absorbed in a labor market with a thick formal core. In a split one, the upside lands with the protected minority and the downside falls on everyone else.

Manufacturing has historically been the main escalator from informal to formal employment in developing countries: it absorbed low-skill workers, raised productivity, and built the payroll base on which contributory social insurance systems rest. Dani Rodrik now argues that this escalator is effectively closed as a broad development strategy. Developing countries are deindustrializing at lower income levels than their predecessors did, with manufacturing employment peaking earlier and at lower shares. Automation, skill-biased technological change, and the fragmentation of global production networks have made modern manufacturing too capital- and skill-intensive to absorb surplus labor at the scale earlier industrializers managed. Some specific countries, like Vietnam, Bangladesh, and parts of East Africa, may still ride the manufacturing ladder for another generation, but as a general path it no longer works. AI makes the squeeze worse. It substitutes capital for low-skill labor, cuts manufacturing’s capacity to absorb workers further still, and chips away at the wage-cost advantage that developing economies have relied on (although garments, with the vagaries of fabric, may still stay predominantly human for a while). The services sector that was meant to provide an alternative ladder for tradeable work (call centers, back-office, business-process outsourcing) is also where AI is making some of its earliest inroads.

Where will the jobs come from? Rodrik’s answer is services. Not the tradeable kind that AI is currently eating, but the more local, labor-absorbing kind that already employs most working people in poor countries: retail, hospitality, care work, hands-on personal services. The economic case is straightforward enough—services will hold the bulk of future jobs anyway. The hard part is raising productivity in those sectors without destroying jobs—a different problem from manufacturing, and the policy toolkit is less developed.

One slice of the services answer crosses borders. As cognitive and digital tasks become cheaper to automate, more value may accrue to work where the human being is part of the service itself, and aging populations in high-income countries are already pulling on that margin. Managed migration corridors are how that demand can meet developing-country supply. The lesson from existing corridors is that language requirements and credential-recognition rules do more work than the headline agreement. Japan’s Economic Partnership Agreements with the Philippines, Indonesia, and Vietnam have brought in nurses and care workers since 2008, but only a few thousand in total: candidates must pass Japanese licensing exams in Japanese, and only about 7 percent of nurse candidates have done so. Germany’s 2023–24 Skilled Immigration Act reforms, which lowered language thresholds for shortage sectors, allowed deferred qualification recognition, and simplified family reunification, issued around 200,000 work visas in their first year. Two different corridors, two quite different results.

The case for reform does not depend on AI

Failures in predicting job outcomes from technological change are familiar. Earlier automation waves displaced workers in specific sectors but did not produce the aggregate job destruction predicted. Bank teller employment in the US kept rising two decades after ATMs arrived, even as the technology eventually reshaped the occupation. Industrial robots reduced direct manufacturing labor in advanced economies but did not collapse aggregate employment. Predictions that offshoring would hollow out 30–40 million American service jobs proved dramatically overstated. Each wave provoked alarm; each was absorbed, and employment kept growing. AI may follow the same pattern, particularly in developing countries where adoption is slower and automation is more expensive compared to labor.

That is a plausible read, and it does not change the conclusion. The 800 million job gap exists whether or not AI lives up to its disruptive billing. The two-tier trap exists. Premature deindustrialization exists. The reform case rests on those facts, not on AI. If AI turns out to be more ATM than industrial revolution, the reforms below lose some urgency but remain necessary.

Protecting workers for change, not from it

Labor policy often starts from a defensive instinct: keep people in their current jobs as long as possible. But work is shifting, and job preservation is the wrong starting point. Workers need benefits they can carry with them, the human capital and skills to shift to new roles, and a sector that is productive enough to absorb new workers.

The first lever is decoupling health coverage, pensions, and unemployment support from formal employment, so they follow the worker instead of the job. That lowers the risk of moving between employers, sectors, or self-employment. It also means paying for coverage through broad-based taxation rather than employer mandates that make hiring more expensive.

The second is skills. Portable benefits handle the downside risk of movement; human capital determines whether workers can move toward something better.

More than half of children in low- and middle-income countries cannot read and understand a simple text by the end of primary school. That is a binding constraint on developing country labor markets, with or without AI. A workforce that has not cleared the basic literacy threshold is poorly positioned for any future, and the future under AI is not one where the bar gets lower.

Beyond the foundations, the picture gets harder. The previous generation of AI—classification, pattern recognition, summarization—left the high-cognitive ground intact, and education systems could reasonably aim to develop students into it. Generative AI is reaching into exactly that territory: drafting analysis, generating arguments, working through novel problems, synthesizing messy inputs. Which cognitive skills will be durably rewarded is now uncertain. But the prerequisite—foundational literacy and numeracy for everyone—is not.

The third lever follows where the jobs will be. If most future formal employment will come from nontradeable services, the central policy question is how to raise productivity in those sectors without destroying jobs in the process. This is what Rodrik’s reframed industrial policy is pointing toward—less about blunt tariffs or factory subsidies, more about the kind of context-sensitive interventions that can lift productivity in care work, retail, and hospitality: tailored workforce training, targeted local public goods, public-private partnerships at city or sector level, and innovation policy designed to push AI toward augmenting rather than replacing labor. None of this is solved policy. There is a great deal of work still to do.

Now for the politics, which is where all this usually dies. Three groups stand to lose. Public-sector unions already hold the most secure formal jobs in the country, with pension and severance terms more generous than what universal coverage would offer. Formal-sector workers have spent careers paying payroll contributions for the benefits those contributions purchased. And the social-insurance bureaucracy is built around employer mandates. Decouple protections from formal employment and you threaten all three. The benefit to the informal majority is diffuse and politically unorganized; the cost to insiders is concentrated, visible, and loud. That is why decades of “obvious” reforms have stalled across the developing world.

From resilience to portability

“Resilience” is not the right word. Too often it means asking people to absorb repeated shocks with a brave face. The policy conversation needs to shift from protecting specific jobs toward helping people move toward better ones. A few countries have started. Brazil’s Bolsa Família shows that broad-based tax-financed transfers can reach the informal majority at scale. Reforms to BPJS, Indonesia’s social security agency, show that contributory systems can be extended toward universal coverage without dismantling existing structures overnight. Rwanda’s universal community-based health insurance shows what near-universal coverage looks like in a poor country. None is a model to copy wholesale, but each broke a political coalition; the orthodox view said could not be broken. The substance of the reforms is not really in doubt. The harder work is building the political coalitions to push them through. AI is not changing what to do; it is shortening the time to do it.