AI has the potential to overcome barriers that have historically prevented expertise and essential services from reaching large parts of the population in low- and middle-income countries.
Voice interfaces can make digital services accessible to people who cannot read or write. Local-language models can reach communities that global digital platforms have struggled to serve. Image-based AI can help extend expertise in health care and agriculture to places where doctors and agricultural advisers are scarce.
There are already early signs of what is possible at meaningful scale in education, health care, agriculture, and other sectors. Yet the countries where AI could have some of its greatest positive effects are also the hardest places in which to scale it. Barriers range from inadequate infrastructure and missing or poor-quality data to weak delivery mechanisms or gaps in funding.
A working AI model does not automatically create a working AI service. Creating population-scale impact requires leaders to understand what stands between a promising solution and widespread, sustained adoption—and to act deliberately to close that gap.
Look Beyond the AI Model
When an AI service stalls, the instinct is often to look at the technology. Was the model accurate enough? Did it understand the language? Was its performance reliable? Important as these questions are, for AI to advance development, it is key to start with the human using the AI service—not the model.
Population-scale AI impact requires eight essential elements. At the top sits national or regional strategy and the orchestration that give the system direction. Beneath that are the models and applications themselves, supported by accessible compute and relevant, high-quality local and domain data. Often less visible but just as essential are the other elements: fit-for-purpose capital, builder talent (the people who create, adapt, and scale the solutions), sustainable delivery channels, and a deep understanding of how intended users live, behave, and make decisions. (See the exhibit.)
The relative importance of these layers is almost the reverse of the attention they typically receive. As a rule of thumb, the model itself may account for only around 10% of what determines success, compute and data for another 20%, and people, processes, and the surrounding ecosystem for roughly 70%.
Investment and attention, however, often gravitate toward the most visible part of the house: the technology. Yet a technically excellent application can still fail if there is no policy framework to support it, insufficient local talent to build and adapt solutions for end-user needs, no viable route to users, or no durable funding once the pilot ends.
The question for governments, development institutions, foundations, and other funders is therefore not simply how to build better AI. It is how to create the conditions in which useful AI can survive and scale.
Find the Binding Constraints
There is no universal playbook for doing that.
Countries begin from very different positions. One may have strong digital infrastructure but fragmented government leadership. Another may have political commitment and funding but insufficient or poor-quality local data. A third may have data and technical capability but lack the delivery channels needed to reach intended users.
In environments constrained by capacity, capability, and cash, trying to strengthen every element simultaneously is neither realistic nor efficient. Leaders need to look at the system as a whole and identify the binding constraints currently preventing the ecosystem from delivering greater impact.
That changes the strategic question from “What AI should we build?” to “What is preventing AI from scaling here?”
Pouring more money into models when the real constraint is delivery will achieve little. Nor does the 10:20:70 framework imply that compute and data are secondary concerns. In some contexts, they are prerequisites: a regulated service may require local compute that does not yet exist, for example, or an application may depend on digital public infrastructure with which it needs to interface. Until those constraints are addressed, development cannot progress.
Building data infrastructure without sufficient or usable data—or without creating demand for its use—may simply create an expensive asset. Funding pilots without a pathway into government budgets and institutional routines can produce impressive demonstrations that disappear when the funding does.
Diagnosing the binding constraints allows scarce resources to be concentrated where they can have the greatest leverage. National AI roadmap work in countries such as Rwanda illustrates the approach: rather than treating AI development as a collection of individual technology projects, the starting point is to assess the wider system, identify the capabilities and infrastructure already in place, determine what is holding progress back—from infrastructure and data to sustainable delivery channels—and sequence interventions accordingly.
Because those constraints differ by context, the best entry point into the ecosystem differs too.
Choose the Right Door to Scale
In practice, there are three broad ways to begin.
National AI Strategy, Policy, and Orchestration. Where activity is fragmented across government, donors, technology providers, and other institutions, the binding constraint may be the absence of a common direction. National AI strategies and roadmaps can establish priorities, clarify ownership, align investment, and create an orchestration mechanism capable of bringing the wider ecosystem together.
High-Impact Solution Delivery and Building Outward. A high-impact demonstrated solution can provide a focal point around which funding, institutional support, talent, and delivery capacity begin to coalesce. In Rajasthan, India, for example, BCG supported an AI-enabled education program that began with a practical challenge: how to improve learning outcomes at enormous scale. Students were assessed repeatedly, AI was used to identify individual learning gaps and recommend personalized remediation, and teachers were equipped to act on those insights. The program reached approximately 3.5 million kids, with a 10%–12% stabilized year-over-year reduction in children two or more grades behind.
Success depended not just on technology itself but on how it was adopted. In Rajasthan, the approach was embedded in leadership priorities, assessment cycles, teacher routines, budgets, and policy. Effective coordination among stakeholders—parents, teachers, and policymakers—reinforced the ecosystem needed to sustain and build on the program’s early results.
Essential-Element Activation. This means addressing a foundational constraint such as data, compute, capital, or talent and deliberately building the wider ecosystem around it.
India is showing how to address constraints to AI impact at scale. Partnerships with hyperscalers and AI service providers are expanding access to local compute, while structured pricing and major investment in digital public infrastructure are helping create a foundation on which AI applications can be built. As these prerequisites are put in place, attention is shifting toward the people, processes, and ecosystem factors that account for the wider 70% of the challenge.
In Telangana, India, BCG helped design and build the state’s Data Exchange, bringing together public and private data, pre-trained models, curated use cases, compute, and a development sandbox. More than 200 stakeholders across over 50 entities were engaged around it, including state data providers, incubators, startups, and investors. The Telangana AI Rising Grand Challenge attracted more than 420 applications from over 240 startups, with six winners selected to undertake solution delivery with government departments spanning transport, education, and health. Tracking these solutions’ impact will be key to determining next steps.
Build for Permanence, Not Pilots
For governments and development institutions, the ultimate test is not whether an AI application works during a demonstration or grant period. It is whether the service can become part of the system: supported by policy, incorporated into budgets, integrated into institutional routines, maintained locally, and delivered repeatedly to the people who need it.
A coherent national strategy can attract investment and talent. Better access to local data can lower the cost of developing applications. A successful population-scale use case can strengthen political support and unlock funding. Trusted delivery channels established for one service can subsequently carry others. In that sense, successful ecosystem development can become self-reinforcing.
For low- and middle-income countries, the stakes are particularly high. The architecture of the AI economy is being built now. Their ability to benefit from it will depend not simply on access to the most powerful technology, but on whether governments, funders, technology partners, and local actors can build and orchestrate ecosystems around their own priorities.
Imagine a 2030 scenario where those AI gains are widespread: a smallholder farmer gets planting advice tuned to local soil conditions and the week’s rain forecast; a child in a resource-constrained classroom receives tutoring in his native language, tailored to his pace; and a community health worker, backed by an AI-native ministry, spots warning signs of illness in a newborn before the family ever has to travel to a clinic.
The measure of success, ultimately, will not be the sophistication of the AI. It will be whether that AI reaches the people it was intended to serve—sustainably, affordably, and at scale.
