AI adoption in global health is moving from pilots to platforms across NGOs, ministries of health, and funders. This playbook maps the highest-value use cases for low- and middle-income country (LMIC) health systems: AI-assisted diagnostics for TB, malaria and imaging; disease surveillance and outbreak signal detection; supply chain and stockout prediction; community health worker (CHW) decision support; and patient triage. It sets out where AI earns its keep, how to sequence deployments against constrained budgets, and the operational realities of intermittent connectivity, thin data, and workforce shortages that separate a demo from a program that saves lives at scale.
From pilots to programs under a tighter budget
Global health entered 2025 under acute financial pressure. The abrupt freeze and restructuring of USAID cut roughly $40 billion in annual assistance, and modeling published in The Lancet warned of up to 14 million additional deaths by 2030 if the funding is not restored. The Global Fund and Gavi both signaled replenishment shortfalls into 2025 and 2026. Against this backdrop, AI is being asked to do more with less: extend a shrinking workforce, cut waste, and target scarce commodities more precisely. Buyers are no longer impressed by accuracy metrics alone. They want cost per outcome, evidence of local validation, and proof that the tool survives outside a research clinic and inside a district health office with unreliable power and a thin bench of trained staff.
The clinical case is real and maturing. WHO endorsed computer-aided detection (CAD) software for TB screening on chest X-ray in 2021, and field studies across high-burden settings show CAD matching or exceeding human readers where radiologists are scarce or absent. Malaria and blood-smear microscopy AI, ultrasound guidance that lets low-skilled operators capture usable images, supply-side stockout forecasting, and CHW-facing decision support have all crossed from paper into real deployment. The task now is disciplined selection and sequencing, not enthusiasm. The organizations that win are those that pick one defensible use case, integrate it into an existing workflow, and prove it before adding a second. Funders increasingly ask for a coverage and outcome baseline before a single license is bought, and for a plan showing how the tool reaches the last mile rather than only the well-connected referral hospital.
A use-case triage grid for LMIC deployment
Score every candidate use case on burden addressed, data availability, connectivity tolerance, and workforce fit. High scorers get funded first. The rest wait for a data or infrastructure prerequisite to be cleared, so scarce money is never spent on a tool the ground cannot yet support. Revisit the grid each quarter, because a use case that was blocked on connectivity or reference data this year may become viable once that prerequisite is met, and the ranking should move with reality rather than staying frozen at the first assessment.
| Use case | Value signal | Readiness prerequisite |
|---|---|---|
| TB and imaging CAD | WHO-endorsed, offsets radiologist shortage in high-burden districts | X-ray hardware, on-device or edge inference for low connectivity |
| Malaria and blood-smear AI | Faster, more consistent microscopy at peripheral labs | Standardized image capture, quality-controlled reference labels |
| Disease surveillance | Earlier outbreak signal from routine and event-based data | Clean case reporting feeds, DHIS2 integration, epidemiologist oversight |
| Supply and stockout prediction | Fewer stockouts of essential medicines, less commodity expiry waste | Consumption and logistics data at facility level |
| CHW decision support and triage | Safer referral decisions where clinicians are absent | Offline-capable app, validated protocols, clear escalation path |
Sequence for impact, not novelty
- Start with one WHO-endorsed or guideline-backed use case such as TB CAD, so procurement, clinical sign-off, and payer conversations are defensible from day one rather than improvised later.
- Anchor every deployment to a measurable health outcome, such as cases detected, referrals completed, or stockout days avoided, agreed with the ministry before launch rather than reconstructed after.
- Design for offline first: assume connectivity and power drop, and require on-device or edge inference with store-and-forward synchronization so the tool works at the rural sites that need it most.
- Co-locate the tool with the workflow it changes, so a positive AI read triggers a defined clinical action rather than producing a report no one reads or owns.
- Run a shadow-mode phase where AI outputs are logged but not acted on, to measure real-world performance on the local population before it ever drives a clinical decision.
Why global health AI pilots stall
- Pilotitis: dozens of six-month demos that never integrate into DHIS2 or district routines and quietly die the moment grant money ends.
- Validating on a US or European dataset and then deploying on populations, devices, and disease prevalence the model never saw during training.
- Assuming connectivity: cloud-only tools that fail precisely at the rural facilities with the greatest unmet need and the weakest bandwidth.
- Ignoring the human workflow, so the AI produces a signal but no one is responsible for the follow-up action it is supposed to trigger.
Track outcomes, not activations
- Additional true cases detected per 1,000 people screened versus the prior standard of care.
- Referral completion rate for AI-flagged patients, not just the number of flags the model generates.
- Stockout days avoided and commodity expiry reduced at supported facilities over the baseline period.
- Cost per confirmed case or per outcome, tracked continuously against the pre-AI baseline.
Frequently asked questions
How is AI actually deployed in low-resource health settings today?
The most mature uses are WHO-endorsed TB detection on chest X-ray, AI-assisted malaria microscopy, ultrasound guidance for non-specialists, disease surveillance, supply and stockout forecasting, and community health worker decision support. These succeed where they offset a specific workforce shortage and integrate into an existing workflow rather than adding a parallel one that competes for staff time.
Does AI work without reliable internet?
It can, but only if designed for it. On-device or edge inference with store-and-forward synchronization lets tools run at facilities with intermittent power and connectivity. Cloud-only architectures fail at exactly the rural sites with the highest need, so offline capability is a selection criterion you screen for, not a nice-to-have you hope to add later.
How should a ministry or NGO choose a first use case amid funding cuts?
Pick a use case that is guideline-backed, addresses a high local disease burden, tolerates poor connectivity, and fits the existing workforce. Tie it to one measurable outcome agreed in advance. A single well-integrated deployment that survives past the grant cycle beats five simultaneous pilots that impress in a demo and vanish at grant-end.
Related reading
Go deeper on this sector and topic.