NextBillion - A WDI Publication

Multilateral development banks (MDBs) have spent the last decade building the foundations of digital economies across low- and middle-income countries (LMICs), by financing broadband networks, data centers, digital ID systems and regulatory frameworks. This work can help to make large-scale AI deployment in these markets a viable investment opportunity.

But laying the groundwork alone will not deliver transformation. As AI moves from a research frontier to a practical tool for governments and service providers across the Global South, MDBs are confronted with a harder challenge: What will it take to actually deploy AI at scale in LMICs, and are they set up to finance that?

Investing in components rather than systems

The AI investment landscape within MDBs today is wide but thin. There are pilots, such as an AI tool for tax administration in Armenia, responsible AI experiments with municipalities in Chile, and innovation challenges for road safety in Asia. There is infrastructure, such as data centers in Thailand and Azerbaijan, broadband expansion across West Africa, and digital identity programs in the Philippines. There are enabling environments, such as cybersecurity support across dozens of countries, regulatory sandboxes, and AI policy advisory.

Each of these initiatives is valuable, and they span across the “AI stack,” including AI applications themselves; enablers like policies, regulations and digital public infrastructure; and foundational technologies like internet access, data and compute infrastructure, and digital devices. Taken together, however, they show that MDBs are still in the discovery phase, investing in components rather than systems, and in pilot programs intended to demonstrate feasibility rather than programs that aim to scale.

What’s largely missing is a financing approach that connects AI use cases to the underlying foundations they depend on (i.e., the enabler and foundational technology components of the full AI stack), and that moves from proof-of-concept to national-scale deployment within a coherent MDB program.

What scaling AI actually requires 

It will take more than funding individual use cases and building data infrastructure to achieve large-scale AI deployment in an LMIC context. Large-scale AI roll-out will require MDBs to get five things right simultaneously, spanning across multiple sectors and working with government ministries that rarely plan together. These include:

• Use cases that are technically sound, locally adapted and designed for the realities of end-users in often under-resourced environments — e.g., community health workers with basic smartphones or smallholder farmers with intermittent connectivity.

• Data foundations or interoperable platforms that connect to existing government systems, with governance arrangements that determine how data is collected, stored and used.

• Compute and connectivity calibrated to the deployment context. Consider a national deployment of AI in healthcare: If it can only run in urban environments, it hasn’t solved the problem of access across the country.

• Devices such as smartphones and tablets in the hands of frontline workers. This remains the most consistently underfinanced layer in the stack, despite being the final link in the chain that decides whether AI-based systems reach traditionally excluded communities or not — and despite the fact that smartphone ownership across LMICs stands at 50%, and just 24% in Africa (as of 2024).

• Enabling policies such as data protection regulations, procurement frameworks, and sector-specific guidelines that allow governments to deploy AI responsibly and at speed.

Many MDBs’ digital and AI strategies articulate these layers clearly but fall short on operationalization. In practice, ensuring that an AI strategy can be effectively executed means addressing these layers in a single, cohesive program as opposed to having them spread across separate projects, teams and financing instruments.

What an AI system with large-scale public benefits could look like 

Consider what an integrated approach might look like in primary healthcare, a sector where the development case for AI is strong, and the infrastructure gap — including both ill-equipped facilities and constraints on provider availability and capacity — is well-documented.

The goal of such an approach would be to create a system that gets ahead of illness instead of just responding to it. Every citizen, regardless of where they live, would receive continuous, personalized health support. That would include preventive care through ongoing monitoring and early detection, delivered by community health workers at people’s doorsteps.

Achieving such an ambition would mean investing across a suite of interconnected AI use cases rather than a single tool, including:

• A clinical decision-support system that helps community health workers diagnose conditions accurately in low-resource settings.
• An administrative layer that reduces the documentation burden on those same workers, enabling them to spend more time with patients and less on paperwork.
• A referral coordination tool that ensures that patients who need higher-level care actually get to the right facility, and that their records follow them.
• A patient tracking system that enables longitudinal monitoring, flagging individuals with deteriorating health or those who have missed critical follow-ups.

Individually, each of these use cases makes an interesting pilot. Taken together, they make up a transformational system.

However, building this system requires a shared infrastructure: a unified data platform that interoperates with the government’s existing health information systems to link a person’s longitudinal health data to their existing national ID, enabling continuity of care between providers and informing the government’s population-level research; data protection regulations that govern how patient data is collected and used; and devices in the hands of every frontline worker.

In addition to infrastructure, capacity building must be a core investment. Community health workers need to be genuinely equipped to use these tools on an ongoing basis, and government health teams must have the ability to manage, evaluate and own these systems without relying on external technical partners to keep them running. At both the grassroots and local governance levels, this requires structured, continuous support that reflects varying levels of digital literacy and local language needs, while accommodating the often-unpredictable realities of fieldwork.

At the same time, project timelines need to shift to ensure that solutions do not become obsolete before they are even launched. With AI performance doubling every seven months, having a six-to-12-month lead time from approval to first deployment, rather than the two-plus years typical of large MDB operations, matters enormously. And active technical support cannot end at go-live; it must continue through the scale-up phase. This can help governments avoid vendor lock-in and dependence on a single provider, manage data governance, and progressively own these systems over time.

Five shifts worth considering to scale systematic AI implementation

Moving from experimentation to systematic AI financing at scale is more an institutional challenge than a technical one. It involves rethinking how MDBs frame ambition, scope programs, structure financing, run operations and deliver advisory services, and includes the following five shifts.

From component funding to sector transformation: This shift reframes the question from “what AI investment can we structure?” to “what would it take to transform this sector through AI?” That reframing shapes everything downstream, including the scope of the program, the mix of instruments, the partners involved and the metrics of success.

Across the full stack, in a single program: This shift requires an institution to take stock of what already exists in a country (e.g., internet connectivity, data systems, workforce capability, regulatory environment) and build a curated set of investments that address the specific gaps between where things are and where they need to be for AI to work at scale.

Financing calibrated to scope: Compute infrastructure, data platforms and AI applications have fundamentally different risk profiles and cost structures — from tens of millions of US dollars for sub-national phases to billions for a national rollout — and a one-size financial instrument is unlikely to serve all three well.

Procurement and operations designed for AI realities: This shift requires procurement that: prioritizes digital public goods, reusable building blocks from comparable countries and MDB-developed tools; treats interoperability as a non-negotiable requirement; and avoids vendor lock-in. It also requires lead times under 12 months; project durations that account for active support beyond the launch date; and evaluation frameworks that capture adoption, use and deployment.

Deep technical advisory front-loaded: This shift involves helping governments: understand AI’s cost reality from the outset; make smart early decisions on model selection, dataset localization and the use of existing digital public goods; and manage the risks that are specific to AI, such as data interoperability failures, low user adoption, and systems that work in pilots but cannot sustain themselves after project financing ends. Two areas deserve particular attention:

• Supporting governments in choosing the right AI architecture: MDBs must help governments work through on-the-ground realities such as network connectivity in rural areas, the processing capability of devices used by frontline workers, power infrastructure, and available budgets. These factors inform choices between frontier, cloud-enabled large language models; offline-capable, cost-effective small language models; or AI-in-a-box solutions that may be equally or more effective depending on the context.

• Helping governments understand the sovereignty stakes of AI procurement: AI procurement carries strategic implications that physical infrastructure never did, and most governments are ill-equipped to navigate them. MDBs should help governments think through where data will be stored, where compute will reside, and who will control the models underpinning critical public services. These decisions, once locked in, are costly and difficult to reverse.

The development opportunity in AI transformation 

MDBs have been helping to build the infrastructure layer of the digital economy across the Global South for years. The next step is to ensure that the communities this infrastructure was built for actually benefit from AI transformation at scale.

AI can extend diagnostic reach in healthcare, strengthen agricultural advisory, and improve government service delivery across other sectors and societal needs, while supporting evidence-based policymaking — at a scale and cost few other interventions can match. As the 2030 deadline for the Sustainable Development Goals approaches, LMICs continue to face an estimated US $4 trillion annual SDG financing gap. AI alone won’t close that gap, but used well, it can make every dollar of development financing go further.

Kunal Walia is a Partner at Dalberg Advisors.

Photo credit: Urupong