Agents in the Developing World: Leapfrog or Divide?

Table of Contents

• The Two Stories Everyone Tells
• What "Leapfrog" Actually Means Here
• Where Agents Genuinely Leapfrog
• Where the Divide Gets Worse
• The Infrastructure Reality Check
• Who Captures the Value?
• The Language and Data Problem
• What Would Tip It Toward Leapfrog
• Insights Most People Overlook
• References

The Two Stories Everyone Tells

There are two clean narratives about agentic AI in the developing world, and both are sold with conviction.

The first is the leapfrog story. A farmer in rural Kenya, a clinic in Bihar, a small importer in Lagos, each gets an autonomous agent that does the work a scarce, expensive professional used to do. No need to build out the institutions the rich world spent a century assembling. Skip straight to the outcome. The reference point is always M-Pesa: Kenya went from "most people have no bank account" to "most adults move money by phone" without ever building a branch network. The pitch is that agents do for cognitive labor what mobile money did for payments.

The second is the divide story. Agentic AI-as-a-service runs on foundation models trained in a handful of labs, on GPUs concentrated in a few countries, billed in dollars, governed by terms of service written elsewhere. In this telling, developing economies don't leapfrog, they rent. They become consumers of intelligence they can't build, on infrastructure they don't own, paying a margin to companies whose moat keeps widening. The productivity gain is real but it leaves the country almost as fast as it arrives.

Neither story is wrong. The mistake is treating them as competing predictions rather than as two outcomes that depend on specific, decidable conditions. Whether a given deployment leapfrogs or divides comes down to where the binding constraint was, and where the money lands. Let me unpack both.

What "Leapfrog" Actually Means Here

Leapfrogging isn't "getting technology faster." It's skipping a developmental stage entirely because a new technology removes the need for the intermediate one. Mobile phones let countries skip copper landlines. Solar-plus-battery lets villages skip the grid. The condition is always the same: the old path was expensive and slow, and the new technology routes around the thing that made it expensive.

For agentic AI, the question is what intermediate stage agents let you skip. The most credible answer is the professional-services layer, the lawyers, accountants, tier-one support reps, junior analysts, radiologists, and back-office clerks that a developed economy accumulates over decades. These are exactly the roles the broader GaaS labor debate keeps circling, and in rich countries the conversation is about which roles agents augment versus replace. In a country that never had enough of these professionals in the first place, the framing flips: agents don't displace a workforce, they supply one that was never there.

That's the real leapfrog thesis. Not "agents make existing workers faster" but "agents provide a professional class that the economy could never afford to train and retain." A district hospital that could never staff a radiologist gets diagnostic triage. A microbusiness that could never afford a bookkeeper gets one that costs a few cents per task. This is where the marginal-cost-near-zero nature of agent labor matters most, the economics that look merely disruptive in San Francisco look transformative where the alternative was nothing at all.

Where Agents Genuinely Leapfrog

Be concrete, because vague optimism is how this topic usually gets written.

Healthcare triage and diagnostics. Sub-Saharan Africa has roughly one doctor per several thousand people in many regions, versus one per few hundred in the OECD. An agent that handles intake, flags urgent cases, drafts referral notes, and answers routine questions doesn't replace a doctor, it makes the doctors who exist reachable by ten times as many patients. The constraint was professional scarcity, and the agent attacks it directly.

Agricultural advisory. Per-outcome pricing fits agriculture unusually well. A smallholder doesn't want a SaaS subscription; they want an answer about a specific pest on a specific crop this week. An agent priced per query, delivered over SMS or a voice call in the local language, routes around both the cost and the literacy barriers that killed earlier "agtech" apps.

Government service delivery. Filing for a permit, checking benefit eligibility, navigating a bureaucracy, these are pure language-and-procedure tasks, the kind agents do well. A capable agent layer over a creaky e-government backend can deliver a citizen experience the underlying system could never produce on its own.

Micro-enterprise back office. The single-person or five-person firm that runs on WhatsApp can suddenly have invoicing, inventory reconciliation, supplier follow-up, and customer support handled by agents. This is the developing-world version of the agent-leveraged small team, the same "small team, big output" pattern that's reshaping startups, applied to the informal economy that employs most of the Global South.

McKinsey's work on AI's potential in emerging markets repeatedly lands on the same point: the upside is largest precisely where existing service provision is weakest. The constraint-removal logic is sound. The catch is that none of it is automatic.

Where the Divide Gets Worse

Now the other column of the ledger.

Value leakage. When a Nigerian SME pays a per-task fee to a GaaS provider, most of that fee leaves the country. It pays for model inference (foreign), cloud compute (foreign), and the provider's margin (foreign). The local productivity gain is real, but the economic surplus it generates flows to whoever owns the foundation model and the data centers, which connects directly to the broader question of who captures the productivity gains from agents. Historically, leapfrogging built local value: M-Pesa created Kenyan jobs, Kenyan agents, a Kenyan company. Agentic AI, deployed naively, can deliver the service without building the local industry underneath it.

Compute dependency. You cannot leapfrog something you cannot run. Frontier models require GPU clusters that, as of the mid-2020s, are overwhelmingly located in North America, China, Europe, and the Gulf. A country that depends on agentic services depends on data centers it doesn't host, export controls it doesn't write, and pricing it doesn't set. That's not leapfrogging the infrastructure, it's relocating the dependency from a domestic gap to a foreign one.

A new kind of inequality. The agent-driven inequality conversation usually frames winners and losers within a country. Across countries, the same dynamic scales up: economies that own model development and compute capture compounding returns; economies that only consume agentic services capture a one-time productivity bump and then plateau, dependent. The gap between the two doesn't close. It widens with every model generation.

Resilience risk. A clinic that routes diagnostics through a foreign agent is one API deprecation, price hike, or sanctions decision away from losing the capability. Leapfrog gains that can be switched off from another continent aren't development, they're a service contract.

The Infrastructure Reality Check

The leapfrog story quietly assumes the pipes already work. They often don't.

Agentic workflows are chattier and more compute-hungry than a single chatbot call. An agent that plans, calls tools, retries, and verifies might make dozens of model calls to complete one task. That's fine on fiber in a city; it's punishing on intermittent connectivity, metered mobile data, and the latency of a round trip to a distant data center. The World Bank's digital development work has documented for years that connectivity in much of the Global South is improving but remains expensive and uneven relative to income, a structural fact that agent reliability has to be designed around, not assumed away.

This is why the deployments that actually leapfrog tend to be the unglamorous ones: SMS and voice interfaces, aggressive caching, small models running locally for the common cases with escalation to frontier models only when needed, and agents engineered to fail gracefully when the connection drops. The teams that ignore this build demos that work in a conference room in Nairobi and collapse in a village two hours away. Agent reliability, a recurring theme across the GaaS cluster, is not a luxury feature in these markets. It's the whole game.

Who Captures the Value?

The single most important variable is ownership, and there's a spectrum.

At one end: pure consumption. A country imports agentic services wholesale, pays per task, owns nothing. Maximum speed, minimum local value, maximum dependency. This is the default, because it's the easiest.

In the middle: a local services layer. Domestic firms build the agents, the orchestration, the local-language tuning, the integration with local systems, on top of foreign models. The model is still rented, but the application layer, the customer relationships, and a meaningful slice of the margin stay local. This is roughly where the most promising African and South Asian AI startups are positioning themselves, and it's a defensible spot. It mirrors how Indian IT services built a global industry on top of foreign hardware and software in the 1990s.

At the far end: sovereign capability. A country or region develops its own models (likely smaller, domain-specific, locally-tuned rather than frontier-scale), hosts its own compute, and keeps its data home. Expensive, slow, and out of reach for most, but the only configuration where the gains genuinely stay and compound. India's investments in sovereign compute and open multilingual models are the most-watched experiment here.

The policy lever that matters most is nudging deployments rightward on this spectrum. A government that simply lets foreign GaaS flood in gets the leapfrog and the divide simultaneously. One that conditions adoption on local capability-building, data residency, local-partner requirements, open-model support, compute investment, can capture more of the surplus, at the cost of moving slower. There's no free lunch, but there is a choice.

The Language and Data Problem

Here's a constraint the Silicon Valley framing routinely misses: most frontier models are dramatically worse in low-resource languages.

A model that's near-expert in English might be mediocre in Swahili, weak in Amharic, and barely functional in a regional dialect spoken by millions but written by few. Agentic reliability degrades fastest exactly where the language data is thinnest, which is to say, in much of the developing world. An agent that confidently gives wrong answers in a language its users can't easily double-check is worse than no agent at all. This is a safety and trust problem, and it maps onto the broader public-trust-in-autonomous-agents question with a sharper edge: trust is hardest to earn precisely where verification is hardest.

There's a flip side that's genuinely hopeful. The data needed to fix this, speech, text, and labeled examples in underrepresented languages, can only be produced locally. That's a real economic opportunity: the developing world isn't just a market for agents, it's the only source of the data that makes agents work in its own languages. Whether that data is harvested cheaply and exported, or collected, owned, and licensed by local institutions, is one of the more consequential and least-discussed forks in this entire story. It determines whether language data becomes another extracted commodity or a durable local asset.

What Would Tip It Toward Leapfrog

Pulling the threads together, the deployments that leapfrog rather than divide share a recognizable profile:

• They attack a genuine scarcity, a missing professional class, rather than displacing existing workers.
• They're engineered for the actual infrastructure: low-bandwidth interfaces, offline-tolerant, small-model-first.
• They keep a meaningful slice of value local through a domestic application layer, local data ownership, or sovereign compute.
• They're built in the local language with local data, not bolted onto an English-first model and hoped for the best.
• They're resilient to the switch being flipped from abroad, through open models, multi-provider design, or local hosting.

Deployments that hit most of these leapfrog. Deployments that hit none of them deliver a quarter of cheap convenience and a decade of dependency. The technology is identical in both cases. The difference is entirely in how it's structured, which is exactly why "leapfrog or divide" is the wrong question if you read it as a prediction. It's a design brief.

Insights Most People Overlook

1. The real leapfrog isn't cheaper labor, it's a professional class that never existed. Commentators frame agents in the developing world as a cost-cutting story (do the same work cheaper). That misreads the situation. In economies short of doctors, lawyers, and analysts, agents aren't undercutting a workforce, they're supplying one. The displacement debate that dominates rich-world coverage is almost backwards here. The risk isn't that agents take jobs; it's that they take the rungs of the ladder that junior workers used to climb into those professions.

2. Per-outcome pricing is a better fit for the Global South than for the West. Subscription SaaS assumes a buyer with predictable cash flow and a credit card. Most of the developing world's economic activity is informal, lumpy, and cash-based. Per-task and per-outcome agent pricing, pay a few cents for one answer, maps onto that reality far better than any seat license. The pricing model the GaaS industry invented for enterprises may turn out to fit emerging markets even more naturally.

3. Compute geography is the new resource curse, in reverse. The classic resource curse is having a valuable thing extracted by foreigners. The agentic version is not having the valuable thing (compute, models) and renting it forever. A country can have brilliant engineers and a thriving app layer and still export most of its AI surplus because the inference happens on someone else's silicon. Sovereignty debates that focus only on data miss that compute location may matter more.

4. Low-resource languages are simultaneously the biggest barrier and the biggest opportunity. Everyone notes that models are worse in underrepresented languages. Fewer notice the corollary: the only place to fix it is locally, which makes language-data collection a genuine domestic industry, if it's structured as ownership rather than extraction. The countries that treat their linguistic data as a national asset to be licensed, not a free resource to be scraped, will capture value that's invisible in today's deployment-first thinking.

5. "It can be switched off from another continent" is the quiet dealbreaker. A capability that depends on a foreign API, foreign pricing, and foreign export-control decisions is not development, it's a subscription. The most overlooked design criterion for agents in the developing world isn't accuracy or cost. It's whether the capability survives the provider changing its mind. Resilience, not performance, is what separates a leapfrog from a leash.

References

• McKinsey Global Institute, The economic potential of generative AI
• World Bank, Digital Development Overview
• GSMA, State of the Industry Report on Mobile Money
• a16z, The Economic Case for Generative AI and Foundation Models