The Next Leap.

Section 01The Short Version

The developing world did not simply receive the internet. It reorganised around its cheapest interface. A handset became a communications network, bank branch, shopfront and identity layer. The next wave changes a different scarce input: the internet made information transport cheap; generative AI makes some interpretation, drafting, translation, coding and decision support cheap.

• Leapfrogging is recombination, not magic. Countries route around one expensive legacy system by combining a cheaper interface with local distribution, regulation and trust.
• Adoption moves faster than productivity. Phones and payments spread before firms, schools and governments reorganise around them. Power, logistics, management and competition still decide the economic return.
• Lower job exposure is not automatic protection. Low-income countries have fewer office jobs for AI to automate, but connected service workers can face global competition before the wider workforce captures productivity gains.
• The winning strategy is not frontier imitation. Most countries need affordable compute, local-language and sector data, skilled adopters, open standards, reliable power and institutions that can buy and audit systems.
• The distribution question comes first. Productivity can rise while domestic value capture remains thin if models, clouds and data pipelines are owned abroad. The defensible position is the last mile: context, workflow, trust and the customer relationship.

Leapfrogging is not the absence of weak links. It is the relocation of the weak link.

Section 02What Internet Leapfrogging Actually Changed

“Skipping a stage” is a useful slogan but a poor model. Developing economies did not abolish infrastructure; they used a new general-purpose technology to route around the most expensive part of an old system. The result was not a copy of the rich world with fewer steps. It was a different arrangement.

What did not disappear

Physical systems remained expensive: ports, roads, power, housing and factories. Capability remained scarce: a connection did not guarantee literacy, management quality or technical skill. Institutions still mattered: digital systems scaled good rules and bad ones alike, including weak competition, fraud, surveillance and exclusion.

Section 03Six Economies, Six Starting Points

The same technology lands in radically different economic structures. GDP size affects market scale; GDP per person affects purchasing power; employment ratios show participation but not job quality; internet use defines the reachable population.

Nigeria’s employment-to-population ratio is the highest in the table while its GDP per person is the lowest. That is not a contradiction. Where social protection is limited, people cannot afford not to work. The statistic counts activity; it does not tell us whether work is formal, productive, secure or well paid. For AI policy, job quality and output per hour matter more than a headline employment rate.

Source note

World Bank indicators: GDP, GDP per capita and employment for 2024; internet use for 2023, the latest common year in the comparison.

Section 04When the Rail Becomes the Market

India: a stack, not an app

India combined reusable digital identity, expanded bank-account access, mobile connectivity and an interoperable instant-payment system. In August 2026, UPI processed 24.509 billion transactions worth about INR 29.82 trillion. Small merchants could accept instant payments without card terminals; benefits could move directly; and banks and fintechs competed above a shared rail.

Yet the rail did not automatically produce formal credit, profitable small firms or high-productivity employment. Interoperability removed one friction. It did not remove every weak link.

Kenya and Nigeria: access before formality

In 2024 Kenya had 42.3 million mobile-money subscriptions and 381,116 active agents, with average monthly value of KSh 724.8 billion. The leap was geographic: the ledger moved nationally while cash-in and cash-out stayed local. Nigeria reached 74% financial inclusion and 1.95 million agents by the end of 2024. Access widened faster than productive formal employment.

Brazil and Bangladesh: ordinary rails at extraordinary scale

Brazil’s Pix recorded nearly 80 billion transactions in 2025, worth more than R$35 trillion; by year-end, 148 million people and 12.8 million firms had used it. Bangladesh reported about 250 million registered mobile-finance accounts and two million agents in 2025. Accounts are not unique people, but the scale shows how a low-cost mobile layer can coexist with export manufacturing and reach far beyond conventional branches.

AI will need its own agent network: teachers, nurses, extension officers, accountants and community organisations that turn model output into trusted action. The interface may be digital; adoption will still be social.

Section 05The Weak-Links Model for Development

Economist Charles I. Jones asks why a spectacular technology can coexist with ordinary growth. His answer is complementarity: output is constrained by the tasks and systems that improve slowly. The logic is even more severe in developing economies.

A country does not need every link to be world-class. It needs each link to be good enough for a specific use case. That is why an offline crop-disease model on an ordinary phone may create value sooner than a national supercomputer.

The unfinished internet divide matters: in 2024, 93% of people in high-income economies used the internet, compared with 27% in low-income economies, while 2.6 billion people remained offline. Average monthly mobile-data use per subscription was about eight times higher in rich countries. The AI leap begins on an unequal foundation.

Section 06Two AI Futures for Developing Economies

Both futures can happen in the same country. Agriculture may experience an augmented leap while BPO experiences disruption without dividend. Public services may get cheaper while dependence on foreign cloud providers deepens. National averages will conceal the split.

The key question is not “Will AI arrive?” It is: which bottleneck does AI relax, which dependency does it create, and who owns the resulting margin?

Section 07Where the First Dividends Could Appear

The earliest gains are most likely where expertise is scarce, the task repeats, errors can be checked and delivery already has a human or mobile channel.

• Agriculture — advice at scale. Kenya’s Agricultural Observatory Platform already supplies real-time information to 1.1 million farmers. AI can add local forecasts, disease recognition and market advice on basic devices. India’s Saagu Baagu programme reported 21% higher chilli yields, 9% lower pesticide use and 5% lower fertiliser use in an early programme cited by the World Bank.
• Education — a co-teacher, not a robot school. A teacher-guided generative-AI pilot in Edo State, Nigeria aligned sessions to the curriculum. The opportunity is teacher leverage and practice feedback; the constraints are devices, pedagogy, safeguarding and evaluation.
• Health — extend scarce clinicians. Low-bandwidth systems can screen images, structure notes and guide follow-up. The strongest design keeps a nurse or clinician responsible, follows verified local protocols and exposes uncertainty.
• Small business — the invisible back office. Drafting, bookkeeping, translation, procurement search, customer support and simple software become cheaper. The development test is not how many AI start-ups exist; it is whether ordinary firms sell more, waste less and enter new markets.
• Government — administrative capacity. AI can classify cases, translate requests and find anomalies. But bad data and weak appeal rights can turn speed into faster error. Audit trails, human review and contestability are essential.

Section 08Jobs: A Small Buffer, Big Bottlenecks

The latest evidence complicates both panic and complacency. ILO estimates put generative-AI exposure at roughly 30–32% of employment in high-income economies and 10–15% in low-income ones. Fewer clerical and professional jobs reduce direct exposure in poorer countries, but weaker connectivity and complementary systems also make the productivity dividend harder to realise. Exposure is task potential, not a forecast of job losses.

1. 1

   The apprenticeship problem

   If entry-level drafting, coding and analysis are automated, where do young workers acquire judgement? Firms may gain output now while weakening the future skills pipeline.

2. 2

   The BPO squeeze

   Countries that exported routine cognitive labour on cost may face global price pressure. The defence is domain knowledge, client trust, regulated workflows and AI-assisted quality.

3. 3

   The informal-productivity test

   The decisive question is whether tools reach traders, farms and microenterprises. Employment may barely move while incomes rise because each hour produces more value.

4. 4

   The gender channel

   Female-dominated occupations are almost twice as likely to be exposed globally as male-dominated ones, reflecting concentration in clerical and administrative work. Transition policy cannot be gender-neutral.

Section 09Country Pathways to 2035

These are directional scenarios, not forecasts. Each starts from the economic structure and digital route documented earlier.

The common move is from consuming general intelligence to embedding verified intelligence in a local workflow. That is where domestic firms, workers and institutions can still own value.

Section 10A Policy Playbook: Build the Complement, Not the Monument

The prestige project is a national frontier model. The development project is making thousands of ordinary organisations more capable.

1. 1

   Finish connectivity

   Treat reliable electricity, affordable data and devices as AI infrastructure. Publish price, speed, outage and usage data by region and income.

2. 2

   Rent compute wisely

   Negotiate diversified cloud access and regional capacity before subsidising prestige-scale national compute. Avoid single-vendor lock-in.

3. 3

   Own context

   Digitise and govern high-value local data: languages, crops, disease protocols, curricula, laws and public records. Clarify consent, licensing and benefit sharing.

4. 4

   Train adopters

   Build capability among teachers, nurses, civil servants, accountants, farmer organisations and SME managers — not only machine-learning engineers.

5. 5

   Buy outcomes

   Public procurement should test accuracy, cost, equity and appeal in real workflows. Pay for verified service improvement, not impressive demonstrations.

6. 6

   Keep markets open

   Require portability, interoperable interfaces and audit access where public functions are involved. Competition above common rails powered earlier leaps.

7. 7

   Protect the first rung

   Create apprenticeships and supervised AI work so automation of junior tasks does not destroy the path to senior judgement.

8. 8

   Measure domestic capture

   Track local wages, firm margins, tax revenue, cloud imports, data ownership and productivity. User counts are not a development scorecard.

Section 11Leapfrogging Without Illusions

The internet era shows that developing countries are not passive recipients of technological change. They adapt tools to missing institutions, invent new distribution systems and sometimes move faster precisely because there is less legacy infrastructure to defend.

But each apparent shortcut relocates the constraint. Mobile networks reduced the need for fixed lines, then made spectrum, devices and data affordability more important. Mobile money reduced the need for branches, then made agents, identity, fraud control and interoperability more important. E-commerce reduced the need for shops, then made logistics, payments and platform power more important.

AI follows the same pattern. It may reduce the scarcity of some cognitive tasks, then make reliable data, verification, domain judgement and accountability more valuable. Countries can gain without building frontier models. Yet countries that own none of the context, infrastructure or distribution may receive cheaper services while surrendering much of the margin.

The internet rewarded countries that became mobile-first. AI will reward countries that become capability-first: not those with the loudest strategy, but those that turn intelligence into reliable action in ordinary institutions.

Section 12Method, Limits & Principal Sources

• Scope. “Developing world” is shorthand for low- and middle-income economies. The six countries are illustrative, not representative.
• Economic data. GDP is current US dollars and sensitive to exchange rates. GDP per person does not show distribution. Employment ratios do not measure hours, formality, productivity or pay.
• Timing. Comparable country indicators use 2024 for GDP and employment and 2023 for internet use. Payment-system figures use later national releases where stated.
• AI exposure. Estimates describe technical exposure of occupational tasks, not adoption, displacement or realised productivity.
• Scenarios. The 2035 pathways are analytical, not numerical forecasts. Agriculture and education programme results may not generalise nationally.

Principal sources

• World Bank Data — GDP, GDP per capita, employment-to-population ratio and internet-use indicators.
• World Bank, Digital Progress and Trends Report 2025 — AI concentration, foundations, adoption and country applications.
• International Telecommunication Union, Facts and Figures 2024 — global internet-use and data-use gaps.
• International Labour Organization — generative-AI job exposure, disruption and gender evidence.
• International Monetary Fund — AI, labour markets and the global economy.
• National payments and finance authorities — NPCI and Reserve Bank of India; Central Bank of Kenya; Central Bank of Nigeria; Banco Central do Brasil; Bangladesh Bank.
• Charles I. Jones, AI and Our Economic Future (2026) — the weak-links lens and contrasting-scenarios discipline.

The complete source notes and tables are preserved in the 15-page PDF edition. Prepared as an independent research report. Free to read and share with attribution. Data cut-off: 24 September 2026.