What does AI progress mean for health systems in low- and middle-income countries?
Author: Jenny Kudymowa
AI is often discussed as a solution to health system challenges in low- and middle-income countries (LMICs) — from shortages of skilled health workers and weak disease surveillance to poor data quality and limited diagnostic capacity — but near-term impact will depend on how ready systems are to use it. This post explores where AI might plausibly help in the near term, and where it may fall short.
How can AI be leveraged to address health system bottlenecks in the near term?
There are many promising examples of AI for health in LMICs already in use1, a steady stream of new initiatives2, and a growing body of commentary highlighting both opportunities and risks. This attention has led to new funding and implementation efforts, including a $50 million Gates Foundation-OpenAI pilot to deploy AI in African primary health care, and Evidence Action’s AI for Good project focused on identifying and scaling high-impact AI applications in LMICs.
Yet enthusiasm about rapid technical progress coexists with deep uncertainty about real-world impact. It remains unclear where AI is likely to matter for health systems, and where it is unlikely to make much difference. Given how rapidly AI is developing, we focus on the near term, where predictions are at least somewhat grounded. Inspired by a recent blog post by Jacob Trefethen at Coefficient Giving on what AI progress does and does not imply for medical progress, I ask a related but distinct question: Can AI plausibly alleviate health system bottlenecks in LMICs over the next 2–5 years?
What I did: Mapping health system bottlenecks against AI progress
In previous work, the Global Health and Development team at Rethink Priorities (RP) examined both the landscape of existing AI-for-health organizations in LMICs and the structural constraints that limit health system performance. I used this as my starting point, and built on it with a shallow review3 of recent writing on near-term AI capabilities and their likely implications for health systems in LMICs.
I then mapped common health system bottlenecks in LMICs against what current and near-term AI capabilities plausibly can and cannot address. Nobody knows exactly how AI will develop, and progress could turn out to be faster, slower, or different from what might be expected. My goal was not to predict specific breakthroughs, but rather to highlight where AI is likely to matter, where it may help only at the margins, and where it is unlikely to alleviate bottlenecks in the near term.
What I found: Five things to keep in mind about AI and health systems
1. AI progress may be driven by cheaper, more reliable, and better-integrated tools rather than new kinds of intelligence. Recent reports show that AI is becoming dramatically cheaper to run4 and increasingly capable of handling the kinds of messy, real-world data that health systems actually use5 (e.g., forms, registers, images, and free-text notes). Many useful models are now small and flexible enough to run with limited connectivity or local infrastructure, rather than requiring constant access to large cloud systems.6 These trends make it more plausible that AI moves beyond small pilots toward routine use in day-to-day health system operations, including in resource-constrained settings.7
2. AI is most useful where health system bottlenecks are informational, coordinative, or administrative, rather than physical or political. The clearest near-term gains appear where constraints are driven mainly by information gaps, delays, or coordination failures rather than by shortages of staff, facilities, funding, or governance capacity. In areas like information systems, service delivery coordination, and supply chains, AI can support functions such as surveillance, referrals, follow-up, logistics, and reporting.8
3. AI can improve care quality and extend reach, but workforce shortages will still bind.
In the near term, AI is likely to extend the reach of health systems in two main ways: by supporting patient-facing tools such as chatbots, reminders, and remote monitoring, and by reducing administrative burden and cognitive load for health workers.9 By supporting decision-making and freeing up clinician time, these tools can improve the consistency and quality of care delivered by existing staff. However, they do not substitute for hands-on clinical work, and are unlikely to close the large health worker gaps in many LMICs.10
4. AI can move where bottlenecks occur, rather than eliminating them altogether.
By easing access to information, triage, or diagnosis, AI might inadvertently increase demand for care faster than supply can be expanded.11 For example, AI-based screening may identify more patients as needing treatment, but when drugs or staffed clinics are limited, delays and congestion can emerge at later stages of care. In such cases, AI-driven gains at the information or screening stage translate into health benefits only if service delivery capacity, including staffing, supplies, and referral pathways, expands in parallel. This mirrors a broader lesson from health systems strengthening: improvements in one health system building block rarely succeed in isolation, and instead depend on aligned improvements across others.
5. Barriers to adoption matter more than technical feasibility
The primary challenge is often not a lack of AI tools, but the lack of health system readiness to implement them. Limited local data, constraints on costs and digital infrastructure, and regulatory barriers often slow or block adoption.12 Addressing these practical barriers is likely to matter at least as much as further advances in AI capability. Adoption barriers also mean that early gains are likely to be uneven, favoring better-resourced settings unless deliberate efforts are made to support lower-capacity systems. Recent reductions in external technical assistance programs have further weakened implementation capacity in many countries, making readiness support even more critical.13
Find a summary of findings for each health system bottleneck in the table below:
Implications for funders and implementers
Based on these observations, several practical considerations emerge:
Assess whether AI will solve or shift your bottleneck. Verify that downstream capacity can absorb the gains; AI improvements at one stage often increase demand at the next.
Don’t fund AI tools in isolation. Budget upfront for data infrastructure, technical assistance, and system integration; adoption barriers often matter more than technical capability.
Ensure AI complements rather than replaces core system investments. Even where AI addresses informational bottlenecks, health outcomes depend on adequate workforce, infrastructure, and service delivery capacity.
Plan for weaker implementation support. With reduced technical assistance infrastructure, expect longer timelines and higher costs for system readiness and data quality improvements.
Crucial questions remain
In charting AI’s implications for health systems, several open questions remain:
Which AI applications can deliver impact within current health system constraints? – Identify tools that are technically capable, fit existing workflows, won’t create downstream bottlenecks, and have manageable adoption barriers.
Why do AI-for-health pilots fail to scale, and how can funders prevent it? – Identify the most common barriers when moving from pilots to routine use, including regulatory hurdles, data integration challenges, and workforce adoption.
Which health systems are ready to adopt AI solutions? – Identify countries or settings that already have reliable data, functioning infrastructure, and enough operational capacity for new tools to be used in practice, not just tested in pilots.
At Rethink Priorities, we’ve explored related questions in our landscaping of AI-for-health organizations and our forthcoming report on barriers to launching AI projects in LMICs. We’re keen to collaborate with interested funders or implementers as we continue this work.
Get in touch
Reach me (Jenny Kudymowa, Senior Researcher) at jenny@rethinkpriorities.org, or the Global Health and Development team at ghd@rethinkpriorities.org.
Acknowledgements
This post was written by Jenny Kudymowa. Thank you to Ruby Emerson, John Firth, and Jerome Mayaud for helpful feedback, to Shane Coburn for copyediting, to Thais Jacomassi for bibliography, and to Elisa Autric for review.
Thank you
Thank you for taking the time to read our Substack. If you would like to support our efforts, please subscribe below or share our posts with friends and colleagues.
We’re also always looking for feedback on our work. You can share your thoughts about this publication anonymously or simply reply to this email/post with suggestions for improvement or any questions.
By default, we’re sharing this Substack via email with Rethink Priorities newsletter members. Please feel free to unsubscribe from this Substack if you’d prefer to stick with our monthly, general newsletter.
Prominent examples already in use include AI-based tuberculosis screening via chest X-rays (e.g., Delft Imaging’s CAD4TB+), AI-enabled detection of substandard and falsified medicines (RxAll’s RxScanner), and AI-supported clinical decision and messaging tools for maternal and newborn care, such as those developed by Jacaranda Health.
For example, several pilot programs are exploring LLM-assisted clinical decision support for community health workers, including HEP Assist in Ethiopia, PATH’s AI chatbot initiative in Rwanda, and ASHABot in India.
Here, “shallow review” refers to a non-systematic scan of recent papers, reports, and commentary, mainly aimed at identifying common themes and recurring ideas rather than conducting a formal literature review. Key sources are cited in the footnotes.
Inference costs for models achieving GPT-3.5-level performance have fallen by more than 99% between late 2022 and late 2024 (Maslej et al., 2025, p. 64).
Recent gains on multimodal benchmarks like Massive Multitask Multimodal Understanding (MMMU) suggest that leading models are increasingly able to process mixed, real-world data formats, e.g., text and images (Maslej et al., 2025, p. 138).
Chen et al. (2025, pp. 3-4) document a shift toward smaller and open-source AI models that can be deployed locally or on edge devices, reducing dependence on centralized cloud infrastructure and continuous internet connectivity.
A recent review finds that AI-for-health applications in LMICs are mostly small-scale or pilot projects, but declining costs and improved workflow integration are reducing barriers to broader routine use (Stanford CDH, 2025).
Commentaries argue that AI should complement rather than replace investments in staff, infrastructure, financing, and governance (Matupi, 2025, Panteli et al., 2025). Moreover, reviews indicate that near‑term applications of AI focus on data and information-intensive tasks such as surveillance, decision support, logistics, and administration (Janků et al., 2025; Alaran et al., 2025).
Reviews consistently point to patient-facing tools (e.g., chatbots, reminders), clinical decision support, and administrative automation as the most plausible near-term applications to expand access to healthcare (e.g., Ong et al., 2025; Akbarialiabad et al., 2025), a pattern also reflected in Rethink Priorities’ landscape review of AI for Health organizations (Emerson et al., 2026).
There is broad consensus that near-term AI in health primarily augments existing workers rather than replacing them, improving productivity and decision support but not eliminating the need for hands-on clinical care or resolving workforce shortages entirely (OECD, 2024, p. 24; Sezgin, 2023).
See Stanford CDH (2025, p. 25), which notes that generative AI may increase demand for care in LMICs and stresses the need for corresponding supply-side and system-level investments.
A systematic review finds that many AI tools work well in theory, but real-world adoption is often slowed by practical issues like weak digital infrastructure, limited data systems, and regulatory hurdles rather than a lack of technical innovation (Rahamtalla et al., 2025). See also our forthcoming report on the barriers to launching AI projects aimed at social impact in LMICs.
For example, USAID’s 2025 funding cuts disrupted the Demographic and Health Surveys program and health management information systems across many countries, reducing technical assistance capacity for data collection, system maintenance, and survey design (Henninger et al., 2025).