As we build the next generation of health technologies, we must ask ourselves: What does true impact look like? This time, we can build with the benefit of hindsight and widen our lens as we do so. If computational tools are to move beyond isolated success stories, they need to be deliberately integrated, governed, and evaluated as part of a broader population health infrastructure.

Precision is often imagined narrowly as clinical personalization, matching treatments to patients based on unique biology. However, what distinguishes precision medicine from precision public health is the promise of prevention at scale and the population as the unit of interest. Realizing that promise requires robust data architecture, systems capable of integrating multifaceted signals, and translating them into timely, actionable insights for entire populations.

Africa is building a foundation for this kind of precision. In Nigeria, for example, the African Centre of Excellence for Genomics of Infectious Diseases (ACEGID) has become a continental leader, providing real-time pathogen surveillance and data that inform population-level responses. ACEGID’s work under the H3Africa initiative, a major pan-African genomics research program launched in 2010 to build genomic infrastructure and capacity across Africa, has expanded genomic infrastructure and training across the continent, making Nigeria a key site for Africa-led genomic science. These efforts represent a broader Nigerian momentum in health innovation, though realising their full potential will require sustained institutional support and integration into national systems.

Strategic coordination is also beginning to emerge at the policy level. The African Union’s 2024 Artificial Intelligence Strategy offers a continent-wide vision for sovereign digital ecosystems. It encourages investment in interoperable data infrastructure and regional collaboration to standardise data governance practices. It also emphasises African ownership of innovation processes and highlights the importance of aligning AI development with public priorities, including health system resilience and responsiveness. Though still early in its implementation, the strategy has catalysed continental policy dialogues with member states on framing and adapting national AI roadmaps to prioritise public interest applications, including health.

Oversight also remains essential. Our evaluation of Swahili machine translation applied to health use cases found that 19 of 71 outputs were considered harmful or misleading. Language, context and cultural nuance cannot be fully captured by models trained on scraped internet text, particularly for underrepresented languages and domains. Human-in-the-loop systems will thus need to continue to provide grounding and accountability, even as automation advances. Masakhane, the grassroots NLP collective, offers an example of how this oversight can be integrated from the start.

The path to impact also requires robust, outcome-centred evaluations that move beyond AI-specific metrics or usability studies. An overreliance on metrics like accuracy, precision, F1-score and recall, while useful for benchmarking model performance, says little about whether tools work in real-world settings or meaningfully improve health. Tools must also be assessed for whether they lead to better care delivery, improved behaviour, or measurable health gains. Where formal AI evaluation is still emerging, best practices from digital health can guide the way. For example, the WHO’s Monitoring and Evaluating Digital Health Interventions framework offers structured guidance that can be adapted to AI contexts.

Examples from other emerging settings provide instructive examples of how data systems can be built modularly, laying the groundwork for scalable, context-aware AI in health. Consider India’s Ayushman Bharat Digital Mission. Rather than relying on cloud-based systems that store all data centrally and require constant internet access, the mission uses a federated architecture. In this model, data is stored locally at the clinic or hospital level and connected through standardised APIs. Patients can manage access to their data through a consent manager. This approach allows facilities with limited connectivity to continue operating and later sync data when possible. It also enables national planning through aggregated, anonymised insights without compromising individual privacy.

Ethiopia’s HEP Assist offers a similarly instructive model for resource-constrained settings. Designed to support frontline Health Extension Workers, the platform provides real-time, step-by-step guidance for screening, diagnosis, and referral decisions across key services like maternal care, child health and infectious disease management. Workers input basic patient information into the mobile application, which then uses built-in clinical protocols to suggest next steps, such as whether a patient should be referred to a higher-level facility or monitored at home. Crucially, the platform functions offline and synchronizes when connectivity becomes available, allowing continuous service delivery even in remote areas. By reducing uncertainty and standardizing care processes, HEP Assist enables more consistent, timely, and appropriate decision-making at the frontlines.