�鶹������ý

The conversation about artificial intelligence (AI) is dominated by a preoccupation with frontier capabilities, leaderboard performance, and ever-larger models. Yet a more consequential challenge facing the field is the gap between what AI can do and what it is actually doing in the world. This “” is global, but it is particularly acute in the Global South, where AI is being deployed amid resource constraints, limited state capacity, and fragile institutions.

This report sets out to understand AI adoption challenges in the Global South: their nature, their drivers, and the workarounds practitioners are already developing on the ground. Its distinguishing feature is empirical. Drawing on 14 interviews, six structured surveys, and a corpus of 53 case studies, the report builds its case through granular, ground-up evidence—in the process, pushing back against the abstraction that often dominates AI policy conversations.

The report is primarily descriptive rather than prescriptive, but four cross-cutting observations run through the evidence:

• First, AI adoption is fundamentally a problem of alignment with real-world systems, not model capability; contextual fit matters more than benchmark performance.
• Second, adoption is a whole-of-society challenge that cannot be resolved at the technical layer alone.
• Third, constraint is generative as well as limiting: resource scarcity is producing distinctive forms of innovation—frugal engineering, modular compute, edge deployment, participatory data collection—that may have relevance well beyond the Global South.
• Fourth, adoption gaps are unevenly distributed and tend to compound for already excluded populations, particularly women.

The report examines eight parameters that shape AI adoption: four technical (compute, models, data, and infrastructure), and four nontechnical (social and institutional embeddedness, capacity and skills, trust, and funding). For each parameter, the report documents both challenges and—importantly—emerging responses and workarounds. Documenting responses alongside challenges is essential: It surfaces on-the-ground innovations that are already taking shape, and points to where policymakers and funders can most usefully build.

The empirical evidence yields a number of findings, many of which go against conventional wisdom or dominant narratives. Here is a partial list:

• The conversation over compute has the wrong unit. Most Global South deployments fine-tune or run inference rather than train frontier models; their compute needs look almost nothing like the gigawatt headline numbers that dominate conversations in the Global North.
• Open source needs a deployment ecosystem, not just a model. Grassroots organizations still often rely on commercial U.S. providers because open-source alternatives lack the technical assistance and onboarding support that make commercial models easy to adopt.
• AI sovereignty laws can undermine the deployments they are meant to protect. While generally well-intended, data and cloud localization rules often prevent domestic start-ups from accessing the compute they need to successfully deploy AI.
• Embed AI in platforms and workflows users already trust. The friction of asking users to adopt new workflows routinely defeats otherwise capable AI tools; WhatsApp integration in particular consistently drives adoption.
• Voice and image interfaces are essential—but poor connectivity can defeat them. Voice and multimodal input are often the best way to reach low-literacy and digitally excluded users, but the communities that would benefit most often have the weakest connectivity.
• DPI stacks are an underused source of AI-ready data, but require careful governance. Where digital public infrastructure exists, it can produce structured, AI-ready data at scale—but only if governance frameworks for consent and reuse are in place.
• Community data collection is reshaping how AI is built in the Global South. The most innovative approaches go beyond labels to capture idioms and cultural meaning embedded in the data. Royalty and licensing models offer a way to compensate contributors as ongoing partners and create recurring sources of revenue.
• Benchmarking is a hidden source of bias. Even models built on well-collected local data can be misjudged if the benchmarks used to evaluate them contain Global North or other cultural assumptions.

Together, the findings gesture toward the contours of a distinct approach to AI development: one that is frugal rather than capital-intensive, embedded rather than imposed, and attentive to gaps and inequalities. If it coheres, this could emerge as a genuine alternative to today’s dominant paradigms—a model of AI development built for, and from, the Global South. The grassroots deployment work documented in this report, often unrecognized and unsupported, suggests that the building has already begun. Now the institutional, funding, and technical ecosystem needs to catch up.

In late 2018, researchers from Google Health partnered with Thailand’s Ministry of Public Health to deploy an in 11 rural clinics. The tool was designed to detect diabetic retinopathy—a leading cause of blindness—from retinal scans, addressing a serious shortfall in the country’s national screening program. Thailand has roughly 4.5 million people with diabetes but only around 200 retinal specialists. In controlled laboratory settings, the algorithm had achieved more than 90 percent accuracy and performed at a level comparable to human medical specialists. The deployment promised to replace days of specialist review with instant, algorithmic triage and to improve care and outcomes for diabetes patients.

The tool was deployed across 11 clinics and 7,600 patients—results were not as expected. The model had been trained on high-quality retinal images taken in well-lit lab conditions; in actual clinics, where lighting was variable and rooms were sometimes shared with other procedures, more than a fifth of the images nurses captured were rejected as too low-quality to process. Patients whose images were rejected—not necessarily an indication of disease—were referred to specialists hours away, requiring them to take time off work and imposing psychological hardship. Slow internet connections delayed uploads, adding to wait times and frustration. Overall, as a follow-up , a tool designed to accelerate care—and proven to do so in the lab—had instead introduced new bottlenecks when released into the world.

The underwhelming deployment in Thailand is hardly an isolated case. Since the release of ChatGPT in late 2022, as frontier AI labs have competed relentlessly to announce breakthrough models with ever-more sophisticated capabilities, another challenge, more mundane yet arguably more important, has hung over the field. The most pressing task today is not building more capable models but rather deploying and fostering the real-world adoption of models that already exist. Much of the AI conversation, focused on leaderboard performance and swelling parameter counts, overlooks this challenge. Yet many of the most important questions in the field—including the sustainability of the AI investment boom and the likelihood of promised productivity gains—hinge on bridging this “,” or what some have termed a “.” Figure 1, from a recent , illustrates the point: A striking gap remains between what frontier models can in theory do and what they are actually doing today.¹

Figure 1 | Theoretical Capability and Observed Usage by Occupational Category

The conversation about AI’s adoption gap has thus far been largely focused on the Global North—notably in pointing to lagging or uneven workplace usage in the United States. As the Thailand example suggests, though, the underlying dynamic is global. In fact, the challenge may be particularly acute—and, as this report will argue, carries several distinctive dimensions—in the Global South, where AI is being deployed in resource-constrained environments amid sometimes limited state capacity and fragile institutions. According to Microsoft’s , aptly subtitled “A Widening Digital Divide,” roughly 24.7 percent of the working-age population in the Global North now uses generative AI tools, compared with 14.1 percent in the Global South. That gap is growing, with adoption in the Global North expanding nearly twice as fast as in the Global South during the second half of 2025. The risks of this trajectory are evident. If AI delivers on even a fraction of its promised gains, uneven adoption will translate directly into uneven development—setting in motion what the United Nations Development Programme (borrowing from economic historian ) calls “.”

This report sets out to better understand adoption challenges in the Global South: their nature, their drivers, and—especially important—the solutions and workarounds already emerging on the ground. The report’s distinguishing feature is empirical. While some of the challenges and remedies discussed here are known at a high level and have been the subject of valuable prior work, this report builds its case from the ground up, through structured interviews with AI practitioners and a compendium of 53 adoption case studies. This approach allows for a more granular and nuanced account than is typical in the policy literature: concrete characterization of specific bottlenecks; attention to the textures of particular contexts and sectors; and close description of the improvisations that practitioners are deploying, often without recognition or support. One of the premises of this report is that such insights, particular and granular though they may be, can inform broader guiding principles and solutions, and potentially articulate a model of AI development for the Global South that is distinct from the dominant paradigms currently being advanced by the United States and China.

Section I provides methodological background, as well as some cross-cutting observations and recurring themes that run through the empirical evidence. The evidence itself is divided into a technical section (II) and a nontechnical one (III); at times the distinction blurs, but the intent is to highlight the fact that AI adoption is an all-of-society exercise, not just a matter of successful technical deployment. In both empirical sections, we discuss a number of parameters that our research suggests play a significant role in AI adoption. For each parameter, we discuss both how it is currently limiting AI adoption and the responses that on-the-ground practitioners are already developing. Together, these challenges and responses provide the basis for Section IV, the conclusion, which tries to identify open questions and directions for further research.

Few fields—in any sector, at any point in history—have been as enveloped in hype as AI. This has many distorting effects. Hype inflates investments and valuations, and it stokes both utopian expectations and dystopian anxieties. In addition, hype also crowds out the kind of grounded, evidence-based analysis that should inform how states and others govern AI: the questions they ask, the trade-offs they weigh, and the priorities they settle on in allocating resources and attention. When it comes to policymaking in particular, hype leads to a version of what Esther Duflo and Abhijit Banerjee, writing of development policy, call “the : ideology, ignorance, inertia.”��

A key goal of this report is to push against the “three Is”: to slow down, look closely at what is actually happening on the ground, and build an empirical foundation on which more informed policy conversations can rest. The particular focus is AI deployment in the Global South, with an aim to better understand adoption challenges and how to address them. If the evidence base is thin for AI policy in general, then it is particularly so in countries across Africa, Asia, and Latin America, where data is scarce and narrative often outpaces analysis. Sometimes we are told that AI will leapfrog legacy infrastructure and accelerate development; at other times, that it will entrench dependency and widen the gap between digital haves and have-nots. One line of thought holds that the Global South is structurally disadvantaged and will fall further behind in the AI era; another insists that the region is leading in frugal innovation, building locally relevant tools, and pioneering use cases the Global North has barely begun to explore. Which of these narratives is true?

The answer is that both probably contain a measure of truth: The real picture, as always, is complicated and uneven. We need to see that picture clearly in order to understand the real trade-offs, how to prioritize resources and, more generally, how to design policy that enhances rather than diminishes the agency and well-being of AI users (and nonusers) across the world.

The evidence presented in this report is a first step in drawing that picture. It draws on three main sources of research, conducted over the first five months of 2026:��

1. Fourteen lengthy interviews and six structured surveys with AI practitioners, builders, and policymakers working on deployment across the Global South. These conversations focused on the day-to-day realities of building and sustaining AI deployments; they surfaced a number of obstacles and also identified glimmers of solutions that are emerging from these deployments.
2. A collection of 53 case studies of AI deployments, assembled from a mix of published reports, practitioner documentation, and some unpublished fieldwork. The corpus was used to elicit specific anecdotes and examples, and it was also analyzed in the aggregate to identify cross-cutting patterns in adoption barriers. The results, summarized in Table 1, do not always match what we found in the interviews (compute, for instance, surfaces less often in these case studies than in our interviews) but they provide a high-level overview of recurring challenges. Methodology for this aggregate analysis, as well as the case studies analyzed, is described in the Appendix.��
3. Finally, desk research, drawn from the growing literature on AI in developing contexts, was used to situate and contextualize the primary evidence.

Several caveats are worth stating up-front. The evidence base is not geographically or sectorally representative: It skews toward countries where the ecosystem is more active and the institutional framework is more robust (India, Brazil, and East Africa in particular). It also may overrepresent sectors such as agriculture, health, and low-resource language development, where the number of deployments has been greatest. Finally, it is biased toward organizations that have survived long enough to be written about; the failures that never made it to publication are almost by definition underrepresented.

This report is a starting point—an effort to build a more grounded picture of a landscape that has been described too often in the abstract. We hope to build on this work in the future with additional interviews and case studies. In the conclusion, we return to some open questions and directions for further research that would help strengthen the evidence base.

Table 1 | AI Deployment Case Study Analysis: Major Challenges Faced��

Note: Percentages reflect the share of case studies that describe each challenge as a substantive obstacle. One case study can be coded for multiple challenges.

This report is conceived as a landscaping exercise. Its primary intent is to map the terrain of AI adoption in the Global South and to surface challenges and workarounds. As such, it does not advance a single thesis. Instead, we highlight four cross-cutting observations that run through the empirical material. More orienting claims than arguments, they recur throughout the evidence and provide conceptual underpinnings for the following sections. They also provide scaffolding for our observations in the conclusion.��

First, AI adoption is best understood not as a problem of model or other technical performance but as a problem of alignment with real-world systems: infrastructure, data, institutions, society, and political economy. The frontier that matters for development outcomes in the Global South is not model capability, it is contextual fit.

Second, both the problems and the solutions we highlight are whole-of-system. Adoption challenges cannot be resolved at the technical layer alone; they require attention to the institutional and social environments as well. At the same time, technology continues to play a central role, though often with a distinctive flavor. Repeatedly, we found evidence of experimentation with locally relevant data, smaller and fine-tuned models, edge deployment, offline-capable systems, and other approaches and architectures that are emerging from or being refined in a part of the world that has long been confined to the periphery of technical innovation.

Third, constraint is generative as well as limiting. AI adoption in the Global South is everywhere shaped, and often hemmed in, by resource constraints. But we found that constraints are not only barriers, they also drive distinctive forms of innovation. Ultimately, we believe that some of these resource-constrained innovations may have relevance well beyond the Global South.��

Fourth, barriers to AI adoption are in most cases compounded for already excluded populations. Women, rural and remote communities, the less educated, and other historically marginalized groups face the adoption gap in sharper form than others. Evidence of gender disparities in particular surfaced repeatedly in our research. This calls for an approach to AI deployment and adoption that moves beyond generic “user” categories and is more intentional about the populations being served and how they are reached.

One of the central observations of this report is that AI adoption is a whole-of-society challenge. At the same time, the multidimensional nature of the endeavor—encompassing society, policy, and political economy—should not obscure the fact that technology remains core to whether and how AI is used across the Global South.��

This section examines four dimensions of AI technology: compute, models, data, and infrastructure (covering electricity and connectivity). While there are other relevant technological parameters—the software and orchestration layers that allow models to be deployed and chained into working applications would be a plausible fifth dimension—these four are the variables that surfaced most often in our research.

Almost half a decade into the generative AI revolution, with the technology having now reached the world’s population, access to compute remains a stubborn and foundational bottleneck. The numbers across the Global South are stark. Africa, home to 18 percent of the world’s people, for less than 1 percent of global data center capacity and an even smaller share of GPU infrastructure. India generates roughly a fifth of the world’s data but only around 3 percent of global data center capacity. The United States and China together contain cluster performance.

On the ground, these disparities translate directly into deployment constraints. , a Ugandan company, was forced to operate with just two donated GPUs for training, leaving none available for inference (the stage at which the model actually reaches users). In Uruguay, a university research lab being limited to fine-tuning models with a maximum of 7 billion parameters (some frontier models now exceed a trillion) due to limits on the national computing cluster. Rodrigo Durán of (Chile’s National Center for Artificial Intelligence) estimated that compute access is the primary bottleneck for the vast majority of AI companies in Latin America and that the constraint is even worse for universities and civil society organizations, which lack the philanthropic cultures that cushion nonprofit organizations in wealthier countries.

AI sovereignty policies play a role, too, if inadvertently. Data localization and similar laws are often reasonable in intent but can create compliance traps that undermine the very goals they are meant to serve. Several organizations told us they find themselves caught between sovereignty aspirations and infrastructure realities—legally prohibited from using foreign cloud providers but lacking the local compute to run models themselves. AymurAI, a tool for processing judicial documents in Latin America, illustrates the bind: The system is designed to run locally so that sensitive judicial data stays protected, but judicial offices lack the necessary hardware, forcing the organization to fall back on less capable models calibrated to limited compute.��

, a Rwandan company that builds capacity for low-resource languages, faces a different version of the problem: Its team collects community language data across Africa but encounters legal obstacles transferring it back to Rwanda to train the models that would arguably benefit the originating countries most. The deeper irony is that because larger—and often foreign—companies are better positioned to operate local compute facilities, data localization laws can disadvantage the very domestic start-ups and communities they were meant to nurture. They also foreclose the kinds of cross-border compute-sharing arrangements—regional clusters, pooled GPU access, federated training across jurisdictions—that could offer smaller and middle-power countries a way around their individual infrastructure limits.��

Solutions: Despite these stark disparities, compute did not in fact rank among the chief obstacles identified in our case studies (Table 1). This may reflect the fact that solutions are emerging—though they require rethinking the traditional paradigm of compute. AI compute in the Global South may come to look fundamentally different from the large-scale, centralized, and capital-intensive infrastructure that dominates mainstream AI discourse.��

One insight recurred across our interviews: Organizations deploying on the ground are finding they may not in fact need nearly as much compute as the traditional paradigm assumes. Part of this reflects an important but often elided distinction between two very different compute needs. Training (the process of building a model from scratch) is enormously compute-intensive; by contrast, inference (the stage at which a trained model actually serves users) requires far less compute. Most organizations we spoke with are not training frontier models: They are fine-tuning open-source models for specific use cases or for running inference on models built elsewhere. Their compute needs, accordingly, look very different from the headline numbers that dominate AI infrastructure debates.

This reality is reshaping how many Global South organizations think about compute. Kate Kallot of , a Kenya-based company, said that governments are often told that “the only way to do AI is to build gigawatt factory data centers” costing hundreds of millions of dollars. Her company takes a different approach, scaling compute incrementally for smaller models targeted at particular use cases and problems. Building modularly—“literally shipping containers” of GPU capacity, stacked as necessary—Amini has been able to help governments deploy robust AI capacity with initial investments of under $5 million. Crane AI Labs arrived at a similar insight, reducing its model size from 4 billion to 1 billion parameters and dramatically cutting compute requirements. Both examples illustrate the kind of frugal, need-based innovation emerging across the Global South—a recurring theme in our research. As , a Nigerian digital health platform, put it in a survey response: “Frugal engineering is not a choice for us. It is the condition under which we operate.”

Compute is often framed as a technical issue, but some of the most effective solutions are to be found in policy. India’s investment in public compute is among the clearest examples. Under the , the country has made more than available to nonprofits and start-ups at subsidized rates—part of a broader interest in public compute across the Global South, including in , , and . Several interviewees noted that this has substantially eased compute constraints in India’s social impact AI ecosystem, especially for resource-constrained organizations.

One interviewee offered an important caveat to this generally positive picture, with global implications. Access to GPUs is necessary, she said, but not sufficient. Despite technical availability, smaller and more remote organizations—those removed from the corridors of power—often lack the human connections or even awareness that would allow them to benefit from public compute. The real unlock, she said, is ecosystem awareness: knowing that public compute exists, knowing how to access it, and—importantly—knowing what has worked for other organizations (i.e., best practices). Case studies, she suggested, can offer road maps and frameworks for adoption and often matter as much as the hardware itself—an indication that the empirical work this report calls for is not only a precondition for good policy but even a precondition for use.

Challenges:

• GPU and compute scarcity
• Sovereignty policy traps

Solutions:

• Frugal engineering: modular and edge compute, smaller models
• Public compute
• Focus on inference and fine-tuning, not frontier training
• Regional collaborations and compute pooling

The situation with models is more nuanced than with compute—the problem is more uneven, less about a single bottleneck, and solutions are correspondingly varied. Global South organizations trying to deploy AI models face at least two core challenges: achieving contextual fit, and avoiding dependency or lock-in. Solving for one does not necessarily solve for the other, and both require a mix of technical, social, and policy responses.

The contextual fit problem is by now well documented: Proprietary models, trained predominantly on English-language and Global North data, struggle for relevance amid very different linguistic, cultural, and sectoral contexts. The dependency problem is somewhat different. It arises when organizations build critical workflows on top of models they do not own or control, creating exposure to decisions made elsewhere. The problem compounds over time as switching costs rise. The concern was most often raised in our interviews by officials and policymakers expressing concerns about national sovereignty, but it applies equally to organizational sovereignty and agency. Like countries, companies and NGOs seek to avoid being locked into a small number of dominant providers. This problem is global— are full of users complaining about deprecated models and unannounced changes—but the risks can be particularly acute in the Global South, where small organizations often lack the technical capacity and financial means to adapt to new APIs, pricing tiers, and model behaviors.��

Solutions: The dominant workaround to the first challenge (insufficient context) is to layer smaller, locally trained models on top of larger foundation models or, alternatively, to fine-tune foundation models on local data. The approach is particularly common in language, health, and agriculture, where local relevance is especially critical. As Rikin Gandhi, CEO of Digital Green, , “Agriculture is hyperlocal: soil type, rainfall, altitude, pests, and markets vary village to village. Model learning must stay close to those realities.” Many organizations we interviewed are now pursuing this strategy, with meaningful results.��

But model stacking or fine-tuning leaves the second challenge (model dependency) largely unaddressed. Smaller contextual models are typically grafted onto imported foundation models—a strategy the AI Now Institute has, in a related context, as “barnacles on the hull of Big Tech.” The structural dependency remains, along with the sovereignty and market concentration concerns that come with it.

Open source is frequently proposed as a solution to this problem, and our interviews suggest it has genuine promise. But open source is also more complicated than sometimes advertised. Many of the organizations we spoke with used proprietary models despite the availability of open-source solutions. This was for at least two reasons: a lack of trust in the security, performance, and sustainability of open-source solutions; and a shortage of technical capacity to implement and adapt open-source models. One interviewee from India emphasized that open-source models—including China’s DeepSeek—were primarily being used by larger enterprises with in-house technical skills. Smaller organizations, unable to adapt and deploy open source, were left with little alternative but to rely on commercial solutions.

As this interviewee pointed out, models operate in an “ecosystem.” She told us that, in her experience, a large number of grassroots Indian organizations used a commercial U.S. provider despite being aware of the risks and shortcomings, simply because this company offered credits for inference and a training and deployment apparatus—including salespeople and technical assistance—that made accessing and building on top of the model easy. Without this broader ecosystem, the promise of open source remains difficult to manifest.��

Challenges:

• Model relevance
• Model dependency��

Solutions:

• Build and layer small local models onto proprietary models (“barnacles on the hull of Big Tech”?)
• Open-source models
• Model ecosystems

Data emerged as the most common challenge in our structured case study analysis, cited by 79 percent of the 53 examples in our case study repository. It was also mentioned as an obstacle by nearly all our interviewees. Broadly, Global South organizations face two types of data challenges: those related to representativeness, and those related to accessibility.��

Data representativeness is closely linked to the contextual fit challenges discussed above. Although model relevance cannot be reduced entirely to training data—architecture, fine-tuning, and deployment context all play a role—the problem often starts with data. Many organizations we interviewed emphasized both the importance and the difficulty of sourcing locally relevant data. One respondent working on a low-resource language model in Rwanda told us the largest parallel corpus available—documents containing original and translated text—came from old Jehovah’s Witness texts, written in stilted, archaic language that is unsuitable for a modern chatbot. �ٲ��ٲ���é�Ա���� noted that it is very hard to find high-quality data that is not gender biased.

The problem of data representativeness typically stems from an absence or paucity of data. Often, however, data exists but remains inaccessible. This can happen for a variety of reasons, but many of our interviewees spoke of technical barriers to AI-ready data. According to the World Bank’s Statistical Performance Indicators index—which includes several proxies for AI-ready data—low-income nations high-income nations, and the gap is widening.��

A lack of AI-ready data can take many forms. Some of our respondents spoke of non-digitized or unstructured data, but the challenge can manifest in less intuitive ways. Rodrigo Durán of CENIA told us that when his team asked Latin American institutions for data to train LatamGPT, a regionally focused large language model (LLM), they received structured SQL databases—precisely the wrong format for training a language model, which requires unstructured natural language text. Durán also drew attention to the governance dimensions of data accessibility, telling us that his work was limited by the absence of standardized data access protocols across governmental organizations, meaning that data sharing arrangements had to be individually and painfully negotiated on a bilateral basis. We heard similar complaints from several others, who described navigating a thicket of emerging and often unclear data handling and privacy laws.

Finally, the data challenge applies not only to training data but to the benchmarks used to evaluate models. Happy Buzaaba, a researcher working on multilingual AI with a focus on low-resource languages, described co-creating a benchmark for 17 African languages that he ultimately concluded could inadvertently perpetuate cultural assumptions from their source datasets: Although the questions were translated into African languages, they were adapted from English-origin benchmarks and therefore did not fully reflect African cultural contexts. Overlooking the benchmark layer risks perpetuating biases beyond the model itself: even LLMs built on well-collected, contextually rich data are subject to distorted evaluation if the benchmarks used to assess them reflect Global North assumptions.

Solutions: Organizations are taking a range of approaches to solving data supply problems—some primarily technical, others more oriented around policy or society. On the technical side, synthetic data has emerged as one workaround when real-world data is scarce. Crane AI Labs, for instance, uses synthetic data for their low-resource language speech models; in particular, they use voice cloning to generate synthetic voices with specific accents, for which there is a paucity of data. , a Filipino-language government chatbot, is another organization that was forced to make up for a lack of voice data by relying on synthetic solutions.

Synthetic data, however, carries well-documented risks (recursive contamination, compounding errors), and several organizations we interviewed said they avoid it for these reasons. A second technical approach is building APIs and technical pipelines to pull data for inference. Rather than relying on the models themselves to contain the most up-to-date and relevant information, organizations are able to link the models to local, real-time external data sources. Digital Green, for example, uses live weather data via APIs to inform responses in its multimodal agricultural chatbot platform; others have built similar pipelines to access live market prices.��

One emerging approach is to repurpose data already collected through digital public infrastructure (DPI). One Indian government official told us that DPI investment had been transformative for AI deployment: Standardized electronic health records and unique citizen IDs mean that AI-ready, structured data already exists at scale. The bottleneck in this approach is not data supply per se but data governance (see Durán’s point, above): establishing the reuse rights, interoperability standards, and access protocols that allow DPI data to flow into AI systems. India’s recently released “” take a step in this direction, recommending extensions to existing consent-based data-sharing protocols, new data portability rights, and tighter integration between AI development and India’s DPI stack. This reflects a broader, cross-national push to develop the legal and technical protocols that govern how data can be reused, shared, and accessed for AI development.��

Technical and policy approaches of these kinds are promising for data that already exists but is inaccessible or poorly formatted. However, much of the data that Global South AI systems need does not exist in any written, much less digitized, form—cultural knowledge, oral traditions, Indigenous languages—and here the solutions emerging on the ground are fundamentally human and social. One approach that came up repeatedly was the use of participatory or community data collection. This method treats data creation as a social process: going into communities, building trust, and co-creating datasets with people whose lives the AI is meant to serve. The approach is being used across the Global South in a variety of ways. , a pan-African collective, runs workshops stretching over days in which native speakers co-create corpora for under-resourced languages; collects voice samples in low-resource languages, often through short WhatsApp recordings.

Digital Umuganda has rethought what data collection (and data itself) means. Rather than asking contributors simply to label or describe what they see—the standard annotation approach—Digital Umuganda asks them to narrate a story. Contributors shown an image of a local football stadium do not just say the local word for that stadium, they create a narrative around it, drawing on the significance of football in their community. The result is data that carries not just labels but context (local idioms and cultural associations) embedded directly into the dataset. This richer, more contextually grounded data helps address not just the supply problem but also the deeper challenge of relevance and cultural fit.��

Challenges:��

• Data representativeness (contextual, cultural, and linguistic fit)
• Data accessibility (machine readable, technical and policy silos)
• Distorted benchmarks and evaluation

Solutions:

• Synthetic data
• Pulling data from live APIs
• Repurposing DPI data (with governance safeguards)
• Community and participatory data collection

While much of the AI conversation focuses on compute, models, and data—our own focus so far—deployments in the Global South are often shaped by more fundamental infrastructural challenges. Many of the organizations we spoke with cited two in particular: electricity and connectivity. Infrastructure was also the third most-frequent constraint that arose in our case study analysis (Table 1).��

AirQo, a Ugandan company that uses AI to monitor air quality across African cities using IoT sensors, illustrates the electricity constraint. Deo Okure, AirQo’s head of research and global partnerships, described unreliable power supply as a primary bottleneck for deploying their hardware. His team’s response has been to integrate solar modules directly into hardware design.��

Connectivity is an even bigger infrastructural constraint, and it comes in several dimensions. The first is basic access: although many of the organizations we spoke with operate in regions with at least some internet connectivity, this is by no means universal, especially in rural or mountainous areas. Even when connectivity exists, pricing can be prohibitive, requiring innovative workarounds. Digital Green found that high cellular (internet) data costs in both Kenya and Ethiopia were a significant barrier to adoption of FarmerChat, its multimodal AI chatbot for agricultural support, and responded by negotiating zero-rating partnerships directly with telcos, making the app free for users to access.

Bandwidth is a third dimension of the connectivity challenge—one that poses particular dilemmas for the low-literacy and low-digital-fluency communities that rely on voice, image, and video interfaces. Arguably among our most important findings is the fact that voice has an increasingly important role to play in bridging the AI access gap, allowing citizens to interact more naturally with AI systems and playing a central role in community and participatory data efforts. But voice is bandwidth-hungry, and low or unstable connectivity limits the feasibility of precisely the modalities that matter most for underserved communities. Digital Umuganda’s team pointed out that broader language coverage means reaching deeper into rural communities and, in the process, pushing larger audio files through less stable connections.

Solutions: Organizations are responding with a range of workarounds to these infrastructural constraints. In Kenya, have teamed up to deploy satellite-based connectivity (in this case, via Starlink) for AI-powered agricultural services across 450 community hubs. AirQo’s solar integration (discussed above) addresses power intermittency at the device level, a strategy adopted by several other organizations that are essentially sidestepping infrastructural constraints by building their own portable infrastructure. This pattern has a longer history than AI: Projects like and have for years bundled offline content servers with local Wi-Fi into single devices for deployment in schools, clinics, and libraries across the Global South. Many of the efforts we encountered represent the AI-era extension of this lineage.

The most striking example we encountered of this “” approach was 0-LA, built by , a for-profit start-up focused on disaster response and humanitarian aid. 0-LA is a briefcase-sized solar-powered unit that bundles an LLM, database, and web server, connected by its own Wi-Fi network. It is, in effect, a self-contained AI deployment system, the entire technical and infrastructural stack compressed into something a person can carry. 0-LA represents an extreme—and particularly ingenious—illustration of edge AI, and a good example of the constraint-generated creativity that surfaced throughout our research.

Challenges:

• Unreliable electricity
• Limited or expensive connectivity
• Bandwidth too low for voice, image, and video��

Solutions:

• Partnerships with telcos for zero-rating
• Solar integrated hardware
• “Portable infrastructure”: self-contained AI solutions

Many AI initiatives end at technical deployment. Often, however, it is after the technical barriers of compute, models, data, and infrastructure are overcome—or at least mitigated—that the real challenge of AI adoption begins. The findings of our case study analysis bear this out: three of the top five barriers are nontechnical, including the second most-cited obstacle (“Social and Institutional Embeddedness,” discussed below). Tobi Olatunji, an African AI entrepreneur, this way: “The assumption we had to kill earliest was that the technical problem was the hardest one. It is not.”��

AI adoption is a multidimensional process, one that implicates society, institutions, policy, and human behavior as much as code or hardware. The barriers we discuss below are more diffuse and harder to define or categorize than the technical parameters above, and the list is no doubt incomplete. Precisely because technology is so deeply embedded in human and social life, a wide variety of nontechnical variables shape its success or failure. We are guided here by what came up most frequently and most prominently in our research.

Across our investigations, one finding recurred with striking consistency: Technically capable AI systems fail when they are dropped into human and organizational contexts without regard for existing habits, workflows, and cultures. In our structured case study analysis, the challenge of embedding AI into social and institutional life—rather than expecting users and institutions to adapt to AI—places just behind data as one of the most commonly identified challenges (Table 1). It sits at the intersection of technology and society, involving both technical design decisions at the user experience and product level and harder social and institutional questions about how to work within long-standing systems and cultures.��

This challenge is a contemporary expression of a long-recognized reality. Scholars of technology and organizations have long argued that new systems are absorbed into, and constrained by, what Ole Hanseth and Eric Monteiro call the “”—the existing configuration of tools, work practices, and institutional arrangements into which any new technology must be fitted. A have highlighted the particular complexity of adapting technologies largely developed in the Global North to on-the-ground realities in the Global South.��

Solutions: Two broad aspects of the embeddedness challenge, and the workarounds organizations have developed, emerged from the research. The first is the need to meet individual users where they already are. The second is the need to integrate AI within existing organizations and their workflows, processes, and cultures.��

When it comes to individuals, we found that AI adoption efforts encounter immediate and often insurmountable friction when they require users to learn new interfaces or develop new digital habits. Across geographies and sectors, one of the most-cited factors for successful adoption in our interviews was integration of AI into platforms users already know and trust. WhatsApp emerged as one of the most common solutions. As of more than 100 Indian AI deployments put it, “The friction of downloading a new app is high while the friction of chatting with a bot on WhatsApp is near zero.” (a company building AI for health care) and (an Indian language tool) are among the initiatives that have integrated their AI offerings into WhatsApp. It is worth noting that the reliance on a single foreign platform sits uneasily with the dependency and lock-in concerns raised earlier in this report (for example, at the model level).

In addition to platform choice, input and output modes matter greatly as well—especially in a deployment context where many users have low literacy or limited comfort with text-based interfaces. Many of the deployments we researched demonstrated that voice in particular is a critical interface layer. Digital Green found that multimodal input—allowing farmers to ask questions by voice and image rather than just text—dramatically increased usability. Crisis Cognition’s platform enables voice queries and spoken responses specifically to address literacy and digital literacy barriers. The Agricultural Information Exchange Platform (AIEP) initiative, piloting AI advisory services for smallholder farmers in Kenya and India, similarly identified voice-based interaction as essential for reaching low-literacy users.��

At the organization level, we found that the landscape of AI deployment is littered with tools that are technically functional but rarely used because they are seen as distractions, inconveniences, or threats. In India, a functioning AI diagnostic tool for neuroradiology cultural resistance from the specialists it was designed to support; studies seeking to understand why found that up to 90 percent of radiologists perceived AI as a threat to their job security.

Conversely, actively integrating new AI tools into existing workspaces and community processes can spur adoption: , a chatbot that provides sexual and reproductive information, worked with Indigenous Quechua organizations to expand its reach among underserved populations. Many of the organizations we encountered made similar targeted efforts to serve women in particular. For instance, Digital Green with the Swayam Sampurna Farmer Producer Organisation, a community of women farmers in India, to spur adoption of its chatbot by local women.��

The institutional integration challenge can have policy dimensions, too. In India, one legal AI platform, by all accounts highly technically capable, has struggled to achieve meaningful impact without integration into official judicial workflows. To do so, however, would require both cultural change within legal institutions and policy frameworks to nudge or require adoption—a challenge that points to the role of public sector mandates, procurement standards, and regulatory signaling in shaping which AI tools actually reach scale.

Challenges:

• Individual discomfort or lack of familiarity with tools
• Institutional resistance or fear
• Policy inertia that resists AI adoption

Solutions:

• Deploy on existing platforms (especially WhatsApp)
• Voice and image input
• Work with existing institutions and communities to reach marginalized groups (Indigenous populations, gender)��
• Policy reform, including procurement and public sector mandates

Part of the solution to integration challenges runs through developing new capacity and skills—but capacity gaps represent significant barriers to adoption in their own right (Table 1). Interviewees across regions consistently cited a lack of AI literacy, and more broadly digital literacy, as one of the most significant impediments to deployment. The problem is especially acute within government, where officials are often making consequential procurement and deployment decisions without understanding, as one interviewee put it, “each brick in the complex system.” Kallot of Amini framed the stakes as higher than just failed deployments, suggesting that governments that do not understand what they are buying are left “vulnerable to extractive models”—a considerable concern for middle power countries in particular, as they seek to navigate between Chinese and U.S. technology solutions.��

The capacity challenge extends beyond just technical skills. Carolina Glasserman Apicella of �ٲ��ٲ���é�Ա���� notes the difficulty of building teams that combine technical AI expertise with domain knowledge—in her case, gender expertise and legal knowledge, which are essential for building tools that are accurate and contextually grounded. In other words, capacity building in AI is significantly more complicated than just training users how to use technology; it is about cultivating teams that can bridge technical and substantive domains and that understand the communities they are building for. It is worth mentioning that this broader notion of capacity—encompassing social and cultural knowledge as much as technical skills—is perhaps not adequately captured in our structured analysis of case studies, which nonetheless identifies a shortage of capacity and skills as the fourth most-frequent impediment to AI adoption (Table 1).��

Solutions: Organizations have taken several approaches to address capacity shortages. Some deploy their own experts directly into client agencies—Amini’s model of temporarily placing technical staff within companies and government agencies is an example. At the community level, in Malawi trained tech-savvy farmers as informal peer support agents who could introduce their Ulangizi chatbot technology to neighbors and friends; these peer support agents help 150 to 200 farmers almost every week.

University partnerships have proven particularly valuable in building capacity, as they provide both expertise and the institutional credibility needed to enter communities. Digital Umuganda, for instance, relies heavily on local academics to conduct surveys and gather contextual data in remote African villages, a model replicated by Bhashini in India and several other grassroots projects.��

Underlying the shortage of capacity and skills is a structural problem that no training program can fix: brain drain. Rodrigo Durán of CENIA cited the far higher salaries offered by large private companies in explaining the difficulty of building AI capacity, particularly in the public sector. Middle power countries, especially those located in the Global South, may be particularly vulnerable to this dynamic. , an Ethiopian engineer and investor, puts it like this: “We are simply training our best minds for export.” Perhaps ironically, rising anti-immigrant sentiment in the West may offer at least a partial solution. The emergence of the United Arab Emirates as a regional AI talent hub—facilitated by favorable immigration policies and tax incentives—offers an example of how talent-friendly environments can begin to compete with the pull of Big Tech salaries and provide an alternative destination for skilled practitioners from the Global South.

Challenges:

• Lack of technical and AI skills
• Lack of domain knowledge combined with technical skills
• Brain drain

Solutions:

• Embed technical staff in partner organizations, especially in the public sector
• Partnerships with academia��
• Peer support agents and other community interventions
• Talent-friendly policies��

AI is to inspire greater optimism in the Global South than in the United States and Europe, where pessimism about job displacement, misinformation, and political polarization has come to dominate public discourse around the technology. Our research nonetheless found a significant undercurrent of skepticism, which functions as a barrier to AI adoption. Some users cited concerns over deepfakes and hallucinations in their reluctance to use the technology. Others were worried about their jobs—local language model builders said they have a hard time enlisting human translators—and still others simply did not believe the technology works (a skepticism that only deepens when users are offered offline or edge solutions that seem to contradict their notions of what AI requires).

Trust can be particularly hard to earn among women, minorities, and other excluded populations. This issue was particularly salient in the case of women and health care. In Lebanon, of 525 young women found that 54 percent cited privacy as a reason for hesitating to use AI chatbots for gynecological and other intimate health concerns. A of an AI mental health chatbot for transgender women and men who have sex with men had similar findings: while many users valued the platform’s perceived anonymity and confidentiality, others cited concerns about providing personal information to systems whose data practices they did not fully trust. These findings are consistent with broader on gender gaps in technology adoption in the Global South as well as with gender differences in AI trust and adoption in the United States.��

Solutions: The most consistent finding across our interviews is that trust travels through people, not platforms or technology. Organizations that have succeeded in building user trust have done so by keeping humans visibly in the loop, often from the very first point of interaction. Dino Rech, from Audere, mentioned the importance of “having folks at clinics that actually talk to you the first time you use the tool.” Digital Green and many others grew in the early stages through on-the-ground partners, and Digital Umuganda’s university partners help build credibility in communities.

Demonstrated efficacy is another way to increase trust. In Malawi, an agricultural chatbot named succeeded in part because farmers could see that recommendations worked. “After following its recommendation, the worms were completely eliminated,” one farmer was quoted as saying in an article. “Since then, I have relied solely on the chatbot’s guidance.” Actively reassuring users about privacy can have similar benefits. A study of sexual and reproductive health chatbot deployments across Kenya, Uganda, India, and Bangladesh found that engagement increased where systems explicitly assured users their information would be kept anonymous. Kenya’s , a WhatsApp-based platform included in the study, opens every interaction by reassuring young users that the conversation stays between them and the bot, building on young Kenyans’ preexisting trust in the security and safety of WhatsApp.

Even when trust is established, it remains a moving target; AI deployers must work to keep earning it. In India, an agricultural platform found that women users, isolated in their communities and with few other social outlets, became perhaps too trusting of the system, using it for personal conversations and sharing sensitive information well beyond the platform’s intended scope. The organizers realized that this created real safety risks and, over time, could erode hard-earned trust (and thus adoption) within the community. Part of the solution in this case was to constrain the model more tightly to its agricultural domain and build in escalation to human support for sensitive or complex queries. The episode also points to a wider imperative: AI builders need to be attuned to new and distinctive forms of safety risk, including those that emerge from the needs of underserved users. Unaddressed safety failures are ultimately deployment failures, too.

Challenges:

• Misinformation concerns, job anxiety, skepticism that AI works
• Privacy fears

Solutions:

• Humans in the loop
• Information accuracy
• Addressing privacy
• Trust as a “moving target”

Funding is an evident and pervasive constraint, one that shapes every other barrier discussed in this report. Most of the organizations we interviewed rely on a patchwork of multilateral, philanthropic, government, and Big Tech funding to get off the ground. The last often comes with complications. Tiago Maluta of the Lemann Foundation notes that Big Tech companies often provide ostensibly “free” AI products and solutions to public educational institutions. However, recipients who do not carefully read the terms and conditions may unwittingly take on risks that outweigh the benefits—for instance, having their data reused for purposes beyond the original scope. Others raised similar concerns about the combination of credits, grants, and in-kind support that often serve as loss leaders and lead generation for private companies.��

Public funding is an option in some countries—India’s $1.2 billion IndiaAI Mission is among the most prominent examples—but governments across the Global South face genuine and difficult prioritization decisions. Two of our interviewees noted the trade-offs governments face in choosing between investing in AI infrastructure and building roads or hospitals. Limited public funding for AI is not necessarily shortsightedness; it reflects the difficulty of weighing competing priorities in resource-constrained environments.

A less recognized but no less challenging problem stems from the realities of “.” Almost every organization we interviewed was operating at pilot stage, and the challenge of moving from proof-of-concept to sustainable deployment was cited as one of the hardest problems they faced. This is a long-standing problem within the development sector, especially when it comes to technology transfers—sometimes referred to as “” dilemma, after the (perhaps apocryphal) carcasses of heavy machinery that littered the Global South during the Green Revolution.

Solutions: Organizations are responding with a range of sustainability strategies. Amini is pursuing a build-operate-transfer model in which the company sets up a system within government, runs it for two years while building capacity, and then hands over the reins to a local team. Two companies spoke of eventually transferring their platforms directly to government, while others are actively working with private sector partners to embed their tools into existing revenue streams and build profitable—or at least sustainable—business models.��

Edge AI represents another pathway. Evert Bopp, CEO of Crisis Cognition, spoke of designing for sustainability. Its portable device, 0-LA, accrues no ongoing token fees (because it uses on-device models), no internet bills, and runs on solar energy—a direct response to the API-dependency trap and infrastructure costs that make so many AI deployments expensive to run.

A final observation: We found that funding directed primarily at technology consistently underestimates the cost of the human and institutional layers. In particular, training and workflow integration are chronically underfunded relative to the technology itself.

Community participation, vital to the process of building culturally embedded data and locally relevant models, is even less of a priority. When participants are compensated at all, they are typically paid a one-time, token amount. Royalty-based compensation models—akin to those being negotiated by rights holders in the Global North in response to AI training on their work—offer a promising alternative, creating arrangements that are both economically fairer and socially more sustainable over time. Such arrangements remain fledgling, but they are beginning to emerge in and in proposals for” that would negotiate licensing and compensation on behalf of communities. At least one Global South AI lab, South Africa’s, is also pursuing licensing frameworks and reinvestment mechanisms that would license community-held language data and, in theory at least, provide an ongoing revenue stream.��

Challenges:

• Shortcomings of private sector funding (extractive, lead-generation)
• Competing priorities for governments
• Sustainability challenges (“pilotitis”)

Solutions:��

• Transfer to government��
• Edge AI (to reduce costs)
• Community licensing agreements

This report has been deliberately descriptive. Its goal has been to map a terrain that is too often described in the abstract and to surface the textures of what is actually happening on the ground. Although we list emerging workarounds, we have generally resisted the move from description to prescription—in part because the evidence base is not yet wide or representative enough to bear the weight of confident policy recommendations, and in part because we believe the more useful contribution here is to open questions rather than close them.

What follows, then, is not a conventional list of policy recommendations but rather a set of research and inquiry directions. Returning to our four cross-cutting observations, we suggest some areas where further work is most needed and where the conversations sparked by this report might productively focus.

The Thailand diabetic retinopathy deployment with which we opened is a useful reminder: A model that achieved 90 percent accuracy in the lab still foundered when dropped into clinics with variable lighting and slow internet. The pattern recurs throughout our evidence—the legal AI platform in India that struggled to reach scale without integration into judicial workflows, the agricultural tools that worked best when embedded in WhatsApp. In case after case, the frontier that matters for adoption appears to be contextual fit rather than raw capability.

Yet the field lacks shared ways of measuring (or supporting) the notion of contextual fit. How should on-the-ground success be defined for AI deployments—by build, by launch, by use, by outcome? What would meaningful evaluation of contextual fit look like in practice, and who is positioned to conduct it? How might procurement and funding frameworks incorporate fit alongside performance? These are just some of the questions worth exploring as we seek to better understand real-world AI adoption in the Global South and beyond.

We repeatedly found that successful deployments depend on an ability to work across layers. IndiaAI Mission has made 38,000 GPUs available to nonprofits and start-ups, but one of our interviewees pointed out that hardware access is only a beginning: Smaller and more remote organizations often lack the awareness and technical assistance that would let them actually benefit. A similar pattern emerged around open-source models, where many grassroots organizations continue to rely on commercial U.S. providers, despite known risks, because those providers offer credits, salespeople, and onboarding support that open-source alternatives cannot yet match.

These examples suggest that the layers of an AI deployment—hardware, software, data, governance, capacity—depend on each other in ways that are not yet well understood. What does an effective public compute ecosystem look like in practice, beyond GPU access alone? What would it take to build deployment infrastructure for open-source models that can compete with proprietary onboarding? Importantly, this is not an argument for every country to seek to develop capacity across every layer of the AI stack. That would be neither feasible nor desirable. The more useful goals are to understand how the layers interact, to identify which dependencies matter most for which use cases, and then for each country or region to make deliberate choices about where to invest, where to partner, and where to accept reliance on others. The question is less about building the full stack than about identifying points of leverage at which targeted investment yields the greatest downstream effect.

Among the most striking of our findings is the observation that, in the Global South, constraints are often producing genuinely novel approaches. For example, Crane AI Labs reduced its model size from 4 billion to 1 billion parameters to fit local compute realities. Amini ships modular GPU capacity in literal shipping containers, allowing governments to start with investments under $5 million rather than hundreds of millions. Likewise, Digital Umuganda is rethinking what data collection means, asking contributors to narrate stories around images rather than simply label them.

Whether these approaches add up to something coherent—a distinct paradigm of frugal AI development—is a question the evidence cannot yet fully answer. But it is worth taking seriously and investigating further. How might “” patterns from earlier technologies repeat or differ with AI? Under what conditions do frugal approaches outperform scale-driven ones, and where do they fall short? We should also consider what the Global North might learn in this area from the Global South—particularly as energy and compute costs constrain even the richest countries—and what it would take to support Global South AI as a generative source of innovation rather than simply a passive transfer destination.

AI adoption gaps are not evenly distributed. The Lebanese survey of young women, in which 54 percent cited privacy as a reason for hesitating to use AI chatbots for intimate health concerns, points to one form of unevenness. The Indian agricultural platform on which isolated women users began sharing sensitive personal information, suggesting forms of risk and safety challenges that are not yet well understood, is another example. Voice and multimodal interfaces emerged repeatedly in our research as essential for reaching low-literacy and digitally excluded users—yet voice and video are bandwidth-hungry, and the communities that would benefit most often have the weakest connectivity.

These patterns raise both empirical and normative questions. What does disaggregated evidence on AI adoption—by gender, language, geography, literacy—actually show, and how can it be gathered more systematically? What forms of safety, trust, and consent matter most, and for which populations? There are also concerns related to design and architecture: for example, how can AI deployers default to the needs of populations least likely to benefit, and what commercial models could make such design choices viable? More generally, we perhaps need to rethink what meaningful inclusion looks like, beyond just adoption. There are glimpses of possibilities in this paper: benchmarks that are more linguistically and culturally representative, for instance, and data compensation mechanisms that treat community contributions as ongoing sources of income rather than one-time transactions.

Taken together, the four directions above gesture toward the contours of a distinct approach to AI development—one that is frugal rather than capital-intensive, embedded rather than imposed, and attentive to gaps and inequalities. Whether that approach coheres into something recognizable will depend on the choices made by policymakers, funders, and practitioners over the coming years. It will also depend on the kinds of conversations the field is willing to have—about what counts as success, about who is being served, about what the Global South has to teach as well as to learn, and about how to balance the legitimate pull of sovereignty against the practical constraints of feasibility.

The deployment work documented in this paper—often unrecognized and unsupported, far from the glare of frontier labs—suggests that the building has already begun. Now the policy frameworks, institutions, and funding need to catch up. Our hope is that this paper contributes to building a more rigorous evidence base that can help foster this enabling ecosystem.

Table A1 | Summary of Challenges and Solutions

Table A2 | List of Individuals and Organizations Included in Interviews

Table A3 | List of Organizations Surveyed

Table A4 | List of Organizations in the Case Studies

This appendix presents a quantitative analysis of how frequently specific implementation challenges appear across 53 case studies drawn from 13 curated repositories of AI for social impact in the Global South. Each case study was coded against nine challenge categories derived from the project’s framework.��

The table below shows how many of the case studies mention each challenge category and what percentage that represents.

Table A5 | Challenge Categories Mentioned in Case Studies

Note: Percentages reflect the share of case studies that describe each challenge as a substantive obstacle. One case study can be coded for multiple challenges.

Data

Data challenges are the most pervasive implementation barrier, appearing in 42 of 53 case studies (79 percent). These include limited access to locally relevant or high-quality training datasets, fragmented and siloed data, lack of Afrocentric or region-specific datasets, data quality issues (inconsistent imagery, incomplete registries), and high cost of data acquisition.

Social & Institutional Embeddedness

Social and institutional embeddedness challenges appear in 38 of 53 case studies (72 percent). AI systems consistently face friction when they require users to adopt new platforms, learn new interfaces, or change established workflows. Across sectors, successful deployments integrated AI into platforms users already inhabit—most commonly WhatsApp—or adapted to existing institutional workflows.

Infrastructure

Infrastructure challenges appear in 30 of 53 case studies (57 percent). Limited rural internet access constrains both end-user engagement and model training; unreliable electricity affects IoT-enabled AI and device-dependent health tools. Several projects adopted hybrid edge/cloud architectures to address low-bandwidth constraints. Energy infrastructure directly constrains deployment in sub-Saharan Africa, undermining both offline device use and cloud-based model training.

Capacity & Skills

Capacity and skills challenges appear in 24 of 53 case studies (45 percent). This includes end-user digital literacy gaps (especially among rural women and farmers), shortages of AI/ML professionals in the health and nonprofit sectors, inadequate curricula in Tier 2/3 cities, and the need for sustained training programs for frontline workers. Several cases highlight that even when AI tools are technically sound, adoption fails without companion capacity-building efforts.��

Trust

Trust and adoption resistance challenges appear in 22 of 53 case studies (42 percent). This ranges from concerns about the reliability of AI tools and skepticism among farmers about working with researchers to develop AI tools to professional resistance among radiologists and judicial staff fearing job displacement. Trust building is particularly important in contexts where AI tools serve as intermediaries in high-stakes decisions (health care, legal aid, financial services).��

Models

Model-related challenges appear in 20 of 53 case studies (38 percent). The dominant challenge is the underrepresentation of African, South Asian, and Indigenous Latin American languages, dialects, and cultural contexts in foundation models.��

Funding

Funding and sustainability challenges appear in 17 of 53 case studies (32 percent). Most organizations rely on philanthropic grants or donor funding with no clear path to commercial sustainability. Common concerns include the high cost of developing foundation AI models, long cycles for government adoption and cost-sharing, difficulty securing long-term concessional capital for infrastructure-intensive models, and short grant terms that terminate projects before they reach sustainability.��

Policy

Policy and governance challenges appear in 14 of 53 case studies (26 percent). This includes fragmented regulatory frameworks across African countries, evolving regulatory pathways for AI-enabled medical devices, data sovereignty constraints limiting cloud deployment, and the absence of data protection policies that create enabling environments for AI deployment.

Compute

Hardware and compute challenges appear in 7 of 53 case studies (13%). Key issues include the cost of GPU access and cloud compute licensing, dependency on donated compute credits, and pay-per-token LLM charges that make scaling expensive. While less frequently coded as a standalone challenge than other topics, compute constraints interact significantly with the data, models, and infrastructure challenges above.

Thirteen source documents were analyzed:

• (GSMA)—11 case studies
• —16 case studies
• —8 case studies
• —3 case studies
• —1 case study
• —1 case study
• —5 case studies
• —2 case studies
• —1 case study
• —1 case study
• —1 case study
• —1 case study
• —2 case studies

Case study repositories and sources were collected by the authors through an internet search of AI for social impact case studies in the Global South. This search yielded 13 sources that contained varying numbers of case studies (descriptions of organizations implementing AI). Each source was then uploaded in PDF form into Claude (Sonnet 4.6). All case study documents were read in full by Claude. An organization was included as a case study if the organization is located in the Global South and the document describes at least one concrete implementation challenge the organization faced. A topic was coded as present (1) if the text explicitly describes that topic as something the project struggled with, had to work around, or actively managed. A passing reference or positive mention without a described challenge does not qualify. Each case study can be coded for multiple topics. Coding criteria can be found in Section 4.3.��

Claude outputted a document with all coding decisions evidenced by specific quotes or paraphrases from the source text. Authors then uploaded this document into an advanced thinking Claude model (Opus 4.7) to fact check the outputted coding and the evidence and to flag both hallucinations and challenges that were present in the source text but not included in the output document. These hallucinations were removed and missing codes were added.

As a final validation metric, the authors then manually fact checked to validate the coding on 75% of cases. The authors confirmed that each coded piece of evidence in the document was, in fact, present in the source document.��

Nine challenge categories were defined and applied:

• Data: Challenges with data availability, quality, local relevance, annotation, bias, or collection.��
• Infrastructure: Challenges with internet access, bandwidth, electricity, power supply, or energy infrastructure that constrained training, deployment reach, or scale. Connectivity and energy are combined into this category. Projects that built offline solutions but still reported infrastructure affecting model training, deployment reach, or scale are coded as present.
• Models: Challenges related to using or adapting AI models—including Western-centric training data, poor performance in local languages, hallucination risks, and need for fine-tuning for local contexts.��
• Capacity & Skills: Challenges with workforce skills, digital literacy, end-user training, or shortage of AI/ML professionals in the development/nonprofit sector. Includes org team, end users, and government partner capacity.
• Funding: Cases where funding was uncertain/insufficient or where the project flagged difficulty with long-term sustainability, dependence on donor grants, or difficulty finding commercial models.
• Policy: Challenges with regulatory frameworks, governance structures, institutional barriers, data governance policies, data sovereignty laws, or regulatory uncertainty. Data governance is coded here, not under Data.
• Trust: Challenges with community resistance to using the AI tool itself or reluctance to trust AI-generated outputs.
• Compute: Challenges with GPUs, cloud access, computing costs, API call/token costs, cloud credits, or device availability (e.g., imported hardware).
• Social & Institutional Embeddedness: Challenges fitting AI into existing workflows, platforms, interfaces, or institutions. Includes friction from requiring users to adopt new platforms, interface/modality mismatches, and difficulty displacing legacy practices. For this topic, we noticed a fine line between organizations encountering it as a challenge and proactively working around it. For that reason, we coded this challenge as present if an organization mentioned integrating their solution into existing systems, even if they did not explicitly state it as a challenge.

• The sample is not random; case studies were selected for publication because they involve successful or notable implementations, which may underrepresent the most challenging or failed deployments.
• Cases from the health compendium and gender empowerment casebook were submitted to a government-organized summit, which may bias toward successful or government-affiliated projects.
• Challenge coding reflects what organizations chose to report, not necessarily the full range of challenges they faced.