The Future of AI Healthcare will be Built in Low-Resource Environments
Zaid Al-Fagih argues that artificial intelligence will transform healthcare everywhere. But the largest marginal gains, and perhaps the first genuinely AI-native health systems, may emerge where governments can build digital infrastructure afresh rather than retrofit AI onto entrenched legacy systems.
Artificial intelligence will radically reshape healthcare in the United States, the United Kingdom, Europe and other advanced systems. That is not the point in dispute. The more interesting question is where it will be easiest to redesign a health system around AI, rather than bolt AI onto structures built for an earlier technological era.
In established systems, the promise is enormous, but so is the drag. AI is being introduced into environments already crowded with electronic health records (EHRs), departmental software, procurement layers, interoperability failures and governance processes designed before generative AI entered clinical practice.
Recent reviews of EHR implementation and AI adoption make the point clearly: digital infrastructure in advanced health systems is often fragmented, difficult to customise and not naturally suited to seamless AI integration. At present, much of the work consists of retrofitting.
Low-resource settings are not easier in every respect. Many face unreliable connectivity, funding constraints, limited implementation capacity and severe workforce shortages. Yet some have an architectural advantage precisely because they are earlier in the digitisation curve. They do not always have to unwind decades of vendor sprawl before they modernise. They can design documentation, referral workflows, coding, decision support and population health reporting as one integrated digital layer from the outset rather than trying to stitch new AI tools onto old systems.
That freedom in procurement matters. A systematic review of EHRs for low-resource settings found that the main barrier to adoption is the cost of purchase and maintenance, which is exactly why open-source options deserve more attention.
OpenMRS, an open-source EHR, was built for resource-constrained environments and states plainly that it offers no licensing fees and no vendor lock-in. OpenEMR is similarly fully open source and free. Both systems meet the global standards for EHRs. Open source does not eliminate cost completely: implementation, training, maintenance and local support still have to be paid for, but it changes both the cost structure and the bargaining position of the health system that adopts it.
The contrast with proprietary systems is stark. Peer-reviewed work has noted that high switching costs help entrench EHR markets and can give vendors leverage to charge heavily for related services.
The headline numbers are sobering: the US Department of Veterans Affairs signed a Cerner contract with a ceiling of $10 billion over ten years; NYC Health + Hospitals invested about $1 billion in its Epic rollout; and Alberta Health Services signed a $459 million Epic software deal for Connect Care, with total project costs expected to reach about C$1.6 billion.
For low-resource countries, avoiding that path early may be one of the most consequential strategic decisions they make. This asymmetry suggests that low-resource settings could become the first places where genuinely AI-native healthcare emerges.
This distinction between absolute and marginal gain matters because health worker scarcity is concentrated in poorer settings. WHO now estimates a projected shortfall of 11 million health workers by 2030, mostly in low- and lower-middle-income countries. In that context, saving even a few minutes per consultation is not a minor efficiency gain.
At Barts Health, around 250 staff took part in an ambient voice pilot; more than half saved at least five minutes per appointment and more than two-thirds reported that consultation quality improved. That is useful in London, where staffing is usually significantly better than in low-resource settings. As a former practising doctor, I know how quickly small delays compound across a day of clinical work. In a chronically understaffed system, it could translate into a trajectory-changing increase in clinical capacity.
The same applies to clinical decision support. In Rhazes’s own studies, we have found that AI performs remarkably well when the appropriate models and architecture are used. The significance of that finding is not that clinicians should surrender clinical judgement to a model. It is that AI can now be engineered to produce management plans tethered to peer-reviewed guidance rather than generic free-text outputs. For low-resource systems, where specialist expertise is often thinnest and high-quality resources are hard to find, that kind of constrained and auditable assistance may prove especially valuable.
None of this should invite a romanticism of scarcity. This trajectory is not inevitable. Many low-resource health systems are not greenfield environments but administratively constrained, fiscally fragile, donor-dependent and operationally overstretched. Institutional readiness depends on far more than the underlying technical architecture.
Low-resource countries are also not blank policy spaces for external vendors to occupy, nor is technological leapfrogging automatically equitable. WHO guidance on AI for health, including its later recommendations on large multimodal models, emphasises transparency, accountability, public benefit and context-sensitive oversight. These are not peripheral considerations to be addressed after deployment. They shape whether AI strengthens public health systems or instead creates new forms of dependency.
For that reason, any credible global framework should rest on three core pillars.
The first is participatory data stewardship. The framework developed by the Ada Lovelace Institute is instructive here: communities should not be treated as passive sources of data, but as stakeholders with progressively greater control over how their data is accessed, governed and used.
The second pillar is local sovereignty. In practical terms, this means national control over health datasets, explicit licensing conditions for any external access, transparent procurement processes, and clearly defined liability regimes.
The third pillar is capacity building. This requires contractual commitments to develop local clinical, technical and regulatory expertise, rather than locking countries into permanent dependence on external bodies.
A WHO-convened coalition could help set baseline rules, but the operative principle should be simple: low-resource systems should not repeat the procurement mistakes of richer ones.
AI will transform advanced health systems. I have no doubt about that. But the first places to build AI into the foundation of care – rather than its periphery – may be those with fewer legacy systems, more procurement flexibility and the willingness to choose open architecture before proprietary lock-in hardens. If that happens, the countries too often treated as late adopters may become the ones that define what AI-native healthcare actually looks like.
Dr Zaid Al-Fagih MBBS MPP BSc is Co-Founder and CEO of Rhazes AI. He previously practised as a medical doctor in the NHS and has worked on humanitarian missions during the Syrian conflict; he holds degrees from Imperial College London and the University of Oxford.
Photo by Google DeepMind
