[Updated April 2026]
This post shares a mapping of AI and machine learning tools in CHW programs in low- and middle-income countries, carried out as part of a scoping review I am working on with a team at Georgetown University and the Community Health Impact Coalition. The mapping was originally published in December 2025 and has been updated to reflect corrections to program classifications and additional verification. 38 programs met the inclusion criteria. A few highlights:
Most programs are in Sub-Saharan Africa and South Asia, with large gaps elsewhere. Sub-Saharan Africa accounts for 23 programs (61%), South Asia for 14 (37%). India leads all countries with 8 programs, followed by Kenya (6), and a cluster of five countries at 4 programs each: Bangladesh, Nigeria, Sierra Leone, Tanzania, and Rwanda. Latin America has 5 programs, all in Brazil and Guatemala. East Asia and the Pacific has one: ThinkMD, deployed in Indonesia. Middle East/North Africa and Central Asia have no programs in this mapping.
LLM/GenAI is the dominant technology type, accounting for 58% of programs. 22 of 38 programs use large language models or generative AI, most launched between 2023 and 2025 alongside the availability of GPT-3.5/4 and open-source alternatives. 9 programs (24%) use computer vision for tasks such as interpreting ultrasound, scanning malaria slides, or screening for cervical cancer. 12 programs (32%) use traditional ML. 14 programs (37%) combine more than one technology type.
This is a field of pilots. 26 of 38 programs (68%) are active pilots. Only 8 have reached meaningful scale: 7 categorized as actively scaling and 1 at national scale (Afya-Tek, integrated into Tanzania's national Unified Community System). The original version of this post reported 34% of programs as showing signs of scaling. That figure has been corrected: 12 programs were reclassified from "active scaling" to "active pilot" after comparing self-reported status against documented geographic footprints. Programs that described themselves as scaling were often operating in a single district or receiving funding to expand, rather than having expanded.
The evidence base is thin, and thinnest for the technology type most programs are using. Only 1 program has RCT evidence evaluating its AI component (safe+natal). 11 programs (29%) have any peer-reviewed evidence. No LLM/GenAI program has been evaluated in a randomized trial; 15 of the 22 LLM/GenAI programs (68%) have only organizational reports. 8 programs have trials underway, including RCTs for ThinkMD, kSanté, and two others.
I intend to update this spreadsheet as I learn about new programs. If you know of any I missed, please send them my way.
I hope this is useful!