No Power, No Data, No Impact: The Hard Limits of Africa’s AI Ambition

Artificial intelligence is rapidly becoming part of Africa's development vocabulary. It is being promoted as a tool to diagnose disease, improve crop yields, personalise learning, monitor ecosystems and strengthen public decision-making. However, the most important question is whether those systems can keep working when electricity is unreliable, broadband is expensive, local data are incomplete and institutions lack the capacity to maintain or govern them.

A new review, "Artificial Intelligence for Sustainable Development in Low-Resource African Contexts: A Structured Review and Adaptive Deployment Framework," published in Sustainability by Ikiomoye Douglas Emmanuel and Ebenezer Esenogho of the University of South Africa, shifts attention from AI's theoretical promise to the practical conditions that determine whether it can deliver lasting development value.

AI may support healthcare, agriculture, education and environmental management, but technical success in controlled settings does not automatically translate into better services, stronger institutions or measurable progress on the Sustainable Development Goals. In low-resource environments, the real test begins after the demonstration: Can the system operate offline, use locally representative data, survive beyond donor funding, earn public trust and remain under local control?

The Last Mile Is the Real Test

Much of the global AI industry is built around assumptions that do not hold evenly across Africa: stable electricity, continuous cloud access, high-performance hardware, large labelled datasets and a deep pool of technical specialists. When systems designed for those conditions are transferred without adaptation, they can become unreliable, unaffordable or impossible to maintain.

The review therefore distinguishes between AI adoption and deployment feasibility. Adoption asks whether an organisation can use an AI tool. Deployment feasibility asks whether that tool can be operated, governed, repaired, adapted and financed under real conditions over time. This distinction changes how governments and funders should assess AI projects. Model accuracy is important, but it is not enough.

The review highlights resource-efficient approaches that could reduce these barriers. Model compression, edge computing, TinyML, federated learning and offline-capable systems can lower dependence on powerful cloud infrastructure and continuous connectivity. Yet these techniques are not automatic solutions. A lightweight model still needs maintenance, institutional ownership, user training and a clear purpose. Technical efficiency must be matched by operational resilience.

Sectoral Promise Meets Operational Reality

Healthcare offers some of the strongest technical evidence. AI systems have demonstrated potential in medical imaging, disease detection, screening and clinical decision support. In settings with shortages of specialists, these tools could help expand diagnostic capacity and improve triage.

However, healthcare also carries the highest stakes. Models trained on non-African populations may perform poorly when disease patterns, equipment and patient characteristics differ. Privacy, bias and clinical accountability become critical when an algorithm influences treatment. The review therefore stresses local validation, representative data, human oversight and integration into existing health-system workflows.

Agriculture presents a different opportunity. AI can support crop-disease detection, yield forecasting, pest monitoring and climate-informed decision-making. Such tools could be valuable for smallholder farmers facing drought, soil degradation and volatile weather. Their usefulness, however, depends on local crop, soil and climate data; affordability; offline functionality; and integration with extension services. Without participatory design, farmers may receive technically sophisticated advice that is impractical or poorly matched to local realities.

In education, adaptive learning and predictive analytics could help teachers identify struggling students and tailor instruction. But unequal access to devices, electricity and connectivity could concentrate benefits among already advantaged learners. Student-data protection, teacher readiness and curriculum relevance also determine whether educational AI reduces or deepens inequality.

Environmental applications may be especially useful for remote sensing, biodiversity monitoring, flood risk and climate analysis. Yet even accurate environmental intelligence has limited value if public agencies lack the capacity or resources to act on it. Across all four sectors, the same pattern emerges: AI can improve prediction and monitoring, but development impact depends on the institutions surrounding the technology.

Infrastructure, Data and Governance Are One System

The review treats infrastructure, data, governance and human capacity as interconnected rather than separate problems. Poor connectivity can push institutions toward external cloud providers, raising concerns about data sovereignty and long-term dependency. Fragmented local datasets can increase bias and reduce reliability. Weak procurement rules can lock governments into proprietary systems that local teams cannot audit or repair. Skill shortages can leave public agencies dependent on vendors long after pilot funding ends.

These risks are especially serious in public services. If locally generated health, education or agricultural data are stored and governed abroad, African institutions may lose control over how those data are accessed, reused or commercialised. If systems are too complex to maintain locally, innovation can create a new form of technological dependency rather than institutional capability.

The authors respond with an adaptive deployment framework built around four linked layers: structural enablers, sectoral applications, potential development outcomes and continuous feedback. The framework is not presented as a tested blueprint. It is a decision-support lens for asking whether the conditions for responsible deployment are present, weak or absent.

The approach also challenges the tendency to treat Africa as a single deployment environment. Urban innovation hubs differ sharply from rural communities with limited electricity and connectivity. A system suited to a university hospital or technology hub may require major redesign before it can function in a rural clinic, school or farming community.

For policymakers, context must therefore shape architecture. Rural systems may need mobile-first, low-power and offline designs. Public-sector applications require stronger procurement, accountability and privacy safeguards. Institutionally constrained settings may need simpler tools with clear human oversight rather than highly automated systems that cannot be audited.

From Pilots to Public Value

Africa does not primarily suffer from a shortage of AI pilots. The deeper problem is the gap between demonstration and durable public value. Too many technology projects are judged by launch events, model performance or short-term user numbers. The review argues for a broader standard: whether the system remains useful after external funding ends; whether local institutions can maintain it; whether communities trust it; whether benefits reach rural, low-income, disabled and marginalised users; and whether measurable development outcomes improve.

AI procurement should include lifecycle costs, local maintenance plans, data-governance arrangements, staff training and exit strategies for vendor support. Donors should fund institutional capacity and local data systems alongside software. Universities and innovation hubs should be supported not only to build models, but also to evaluate, adapt and repair them.

Private companies also face a strategic choice. Products designed for high-bandwidth, cloud-dependent markets may struggle to scale in low-resource contexts. Firms that build multilingual, low-power, interoperable and offline-capable systems could create both development value and viable markets. But that opportunity depends on avoiding extractive models in which local users provide data while ownership and economic value remain elsewhere.

The review connects directly to SDGs on health, hunger, education, innovation, climate action, strong institutions and partnerships. Yet it also warns that AI can obstruct those goals when it amplifies exclusion, consumes scarce resources, weakens accountability or shifts control away from local institutions.

The next research frontier is not another catalogue of AI possibilities; it is evidence on what survives. The proposed framework needs testing across countries, sectors and institutional settings. Longitudinal studies should examine whether systems remain operational after pilot funding ends, while evaluations should track affordability, trust, equity, institutional adoption and measurable SDG outcomes.

Africa's AI future will not be determined by who announces the most ambitious strategy or deploys the most advanced model. It will be determined by whether technology fits the infrastructure, institutions and communities expected to use it. In that sense, the continent's AI challenge is not simply to innovate faster. It is to build systems that can endure.