1 of 33

AI Commons for Global Health Workshop

Africa Health Tech Summit 2026

Kigali, Rwanda

1

2 of 33

Workshop Overview

01

Welcome + Background

02

Working Group Share-Outs (Data, Evaluations, CDSS, Procurement)

03

Breakout Group #1 - Refining WG Products

04

Breakout Group #2 - Identifying Additional Unaddressed Problems

05

Breakout Group #3 - Brainstorming Ways for the AI Commons to Serve YOU

07

Closing

2

2

3 of 33

Welcome!

7 minutes: 2:00 - 2:07

​

BACKGROUND

​

Aman Dalmia

Lead AI Engineer

The Agency Fund

Amarynth Sichel

Associate Director, AI Foundations

CHAI

Annie Hartley

Professor

LiGHT / EPFL

Gustav Praekelt

Co-Founder

Turn.io

Rubayat Khan

Director of Health Programs

Endless Network

3

3

4 of 33

Opportunity: Universal, high-quality care is now years away, not generations

​

​

​

BACKGROUND

​

  1. Clinical Intelligence will bring expert-level guidance to the hands of frontline health workers.

​

  • Diagnostic power will help make earlier, more accurate detection where specialists are scarce

​

  • Health-system management can make smarter supply, staffing, and resource decisions at scale

However….

4

4

5 of 33

Challenge: Without intentional, coordinated efforts, AI will fail to deliver health outcomes in LMICs

​

​

​

BACKGROUND

​

  • Built Elsewhere. Most AI tools are trained on high-income data, reflecting languages, disease burdens, and infrastructure of high-income contexts.

​

  • Duplicated efforts. Each wave of technology, from eHealth to mHealth to digital health, brought experimentation that struggled to scale without coordination.

​

  • Path of least resistance. Without intentional effort, innovation flows to where money and data are already abundant.

Why?

5

5

6 of 33

Addressing the Problem

​

​

​

​

BACKGROUND

​

Over 3 days, 23 leaders mapped gaps in the AI for Global Health ecosystem, prioritizing where to take action.

What We Did

1. Aligned on problems blocking AI-enabled health gains in LMICs

​

2. Mapped existing initiatives

against those problems

​

3. Crowded in ideas to address outstanding gaps

​

4. Prioritized and launched working groups

How We Prioritized

1 Speed

Can it move within 18 months?

​

2 Need

Does it fill a real gap in the ecosystem?

​

3 Commons-shaped

Is the gap best addressed by a public good created and/or stewarded by a community?

Who Was Present

Global South Innovators

Frontier AI labs

Global tech companies

Academia

Multilateral institutions

NGOs & implementers

Funders

6

6

7 of 33

Outputs from the Convening

​

​

​

BACKGROUND

​

Mission

the AI Commons for Global Health stewards a thriving ecosystem that provides the trusted and shared tooling, standards, and knowledge required for AI to support improved health care access, quality, and efficiency in LMICs.

​

​

For those with tightly aligned and funded workstreams. Small delivery teams driving the four projects.

01 Working Groups

Open network for anyone working on AI in global health - sharing learnings, standards, and opportunities as the Commons grows.

02 Community of Practice

If this work resonates, reach out to discuss the initiatives and the future of the Commons.

03 Funders

Structure

7

7

8 of 33

The Groups aiming to close the gaps

BACKGROUND

​

Challenge: Governments must procure the best-fit AI solutions to support their health systems.

04 Procurement

Challenge: Existing approaches to evaluating performance for AI-enabled CDSS are time consuming, and not grounded in LMIC contexts.

03 AI Evals

Challenge: Datasets that would help make solutions appropriate for LMIC contexts are often hard to share and hard to access.

​

01 Datasets

Procurement Commons

Description

01

02 CDSS

Challenge: LLM-enabled CDSS requires new architectural approaches and many innovators are learning the same lessons, without access to shared insights.

The Commons set up four Working Groups that are tied to specific deliverables and will be spun down when done.

8

8

9 of 33

01

Data Workstream

AI COMMONS for Global Health

​

5 minutes: 2:07 - 2:12

9

9

10 of 33

Problem

Data Workstream

Data exists…but…

Hard to find

Hard to get permission

Too much work

Risk > Reward

Fear of losing control

10

10

11 of 33

Problem

Hard to find

Data Workstream

Data exists…but…

Hard to get permission

Too much work

Risk > Reward

Fear of losing control

Discoverable

Governance

Simplified contribution

Tooling

Flexible, granular contribution options

Solution

NOT another centralized repository

11

11

12 of 33

PILLAR 1

Data Commons

Sharing data is hard. Finding data is hard.

​

Let’s make it easier.

Data Workstream

12

12

13 of 33

Data Workstream

Next steps

01 Build with data owners

​

Co-design the site.�

Find and reduce the barriers to sharing

​

→

02 Build the community

​

→

03 Wild success

Connect data and users

Hackathons, events, joint research and publications around shared priorities.

​

Join our breakout session to help us define what success looks like

14 of 33

Help shape the AI Data Commons Working Group.

Bring a dataset, a community, a model, or your ideas.

​

Test it out and shape it

14

15 of 33

02

CDSS Workstream

AI COMMONS for Global Health

​

5 minutes: 2:12 - 2:17

​

15

15

16 of 33

CDSS Workstream

LLMs make CDSS more flexible,

changing architecture needs

Deterministic CDSS

LLM-enabled CDSS

Safety

Guidelines hard-coded in decision logic

Safety

Safety built around the LLM: grounding, guardrails, orchestration

Validation

Logic checked against guidelines pre-deployment

Validation

Monitoring before and after deployment

The problem

Design decisions

New architecture, new design choices: when to use the LLM, when to call deterministic tools

Isolated development

Innovators working out the same norms in isolation, which slows progress

16

17 of 33

CDSS Workstream

A reference architecture, drafted with global innovators

What it is

Modular building blocks for LLM-enabled CDSS

Blocks can be swapped, shared, or built once and reused

How it is organized

40 candidate blocks, eight layers

Ordered by the flow of a clinical encounter

Common or bespoke

Common: shared needs, used by many

Bespoke: specific to one context

Hypotheses to test; common blocks still need local tuning

17

18 of 33

CDSS Workstream

Eight layers follow the flow of a clinical encounter

0

Interaction and input

Voice-first multilingual input, device connections, patient ID and records

1

Clinical reasoning core

History, differential diagnosis, triage, red flags, treatment plan

2

Knowledge and grounding

WHO SMART Guidelines, retrieval grounding, country constitution file

3

Safety and governance

Audit logs, human escalation, de-identification, drift monitoring

4

Orchestration and routing

Care-level routing, autonomy levels, adapters to existing apps

5

Evaluation and monitoring

Vignettes, LLM-as-judge, benchmarks, monitoring (evals workstream)

6

Data and interoperability

Record connectors, data ownership, donations to the Commons

7

Deployment and infrastructure

Offline and low-bandwidth, model-agnostic hosting, cost targets

18

19 of 33

CDSS Workstream

Next steps

1

Validate

Test blocks with landscape data and innovators

[IN PROGRESS]

2

Establish demand

Confirm which common elements innovators want

3

Build common elements

Build shared elements any team can reuse

4

Update and maintain

Keep architecture and shared elements current

19

20 of 33

Join us for the CDSS innovator showcase webinar on October 8

What you will hear

Innovators share the architectural design choices, successes, and challenges behind their LLM-enabled CDSS.

Scan to register

Speakers

Noora Health

Anubhav Arora

The Frontline Knowledge Project

Priyank Jain

Penda Health

Robert Korom

ARMMAN

Zeel Mehta

Qure.ai

Ankit Modi

CHAIRII

Dr. Fred Mutisya

Intelehealth

Neha Verma

20

21 of 33

03

AI Evals Workstream

AI COMMONS for Global Health

​

5 minutes: 2:17 - 2:22

​

21

21

22 of 33

Challenge: good, context-appropriate evals are critical for impactful solutions

THE PROBLEM

▶

THE LIFECYCLE APPROACH

1

Model selection

Which model fits my use case?

2

Application staging

Does my whole application behave safely?

3

Monitoring and CQI

Is it still safe and effective in the field?

THE PROBLEM

Today's evaluations are not appropriate for LMIC health systems

Languages

Context

Infrastructure

…

22

23 of 33

The working group's focus is the tooling layer that connects shared infra assets to real products.

TOOLING LAYER

AI Commons x OHS-SF working group

PRE-DEVELOPMENT

​

Read model results for a real use case

PRE-DEPLOYMENT

Staging harnesses and schema adapters, for FHIR and custom data models

POST-DEPLOYMENT

​

​

Telemetry, review queues and CQI playbooks

▲

BUILDS ON

SHARED INFRASTRUCTURE

gHBench (contributing initiative led by PATH, MLCommons, WHO, OHS-SF, CHAI, Google, Anthropic and others)

Judges

Patient emulators

Clinical data

Federated harness

Enhance, don't duplicate

Distil into open assets

Road-test with the field

THE TOOLING LAYER

24 of 33

The “AI Evals” Working Group

OPEN HEALTH STACK SOFTWARE FOUNDATION

FHIR foundations

OHS Player

PILLAR

AI Commons

AI Evaluation Working Group

Hosted in OHS-SF

1

Align

Map the three-phase lifecycle with implementers, researchers and funders.

2

Learn

Inventory who is doing what, so nobody rebuilds the same harness.

3

Deliver

White paper, technical roadmap and funding proposals for the Governing Board.

4

Sustain

A co-led technical group meeting fortnightly to steward open connectors and playbooks.

AI EVALS WORKING GROUP

Kicking off�Oct 20th

24

25 of 33

04

Procurement Workstream

AI COMMONS for Global Health

​

5 minutes: 2:22 - 2:27

25

25

26 of 33

Procurement Workstream

Governments must select best-fit AI solutions for health systems

Current processes

Many governments currently lack clear procurement guidelines for selecting AI-for-health solutions

The challenge

With nascent systems, solutions might not be selected based on evidence

Our goal

Governments have tools that help them choose based on evidence and rigor

27 of 33

Key insights from

initial scoping

Vendor Perspectives

  1. Partnership not procurement�Public sector procurement of AI for health tools often runs through partnership models, rather than procurement.

�

  • Missing evaluations and evidence�Evaluations are a key bottleneck, meaning solutions are selected sometimes based on networks not evidence.

�

  • Long, unclear decision cycles�Decision chains are long and unmapped, and staff rotation hampers progress; interviewees mentioned ~3 year decision cycles

27

27

28 of 33

Procurement Workstream

AI requires different procurement approaches

Usual purchase

So government should ask for

Deterministic; only one test is needed

Dynamic; can give a different answer for the same question

Very high accuracy on the specific task of interest, making it almost certifiable

Fixed operating cost

Costs change based on queries and other factors

Variable cost accounted for in the contracts

Relatively stable products

Key components change rapidly as new models are released

Solutions without any vendor lock-in

AI

29 of 33

Procurement Workstream

Next steps

Starting in Asia

01

Learn the challenges from people who run procurement

→

02

Build a playbook for procurement that helps select best-fit solutions

→

03

Test it with real government buyers, and improve it

In partnership with WHO-SEARO, Indian School of Business and Ashoka University

30 of 33

Our ask from you

What we need

An introduction to national or state governments, or anyone planning to buy health AI at scale

Timing

Teams about to make a buying decision now

What matters most

Willingness to try our playbook and tool, more than the contract size

31 of 33

Breakout Session #1: Refining Current Products

45 minutes: 2:27 - 3:13

​

​

​

AI COMMONS for Global Health

​

Objective: Gather feedback on current artifacts that AI Commons working groups are currently building to co-create and further refine product vision.�

​

Guiding Questions:

  • What elements of current products are most helpful?�
  • Are the assumptions about user groups and needs accurate?�
  • Other…

31

31

32 of 33

Breakout Session #2: New challenges and YOUR role

30 minutes: 3:13 - 3:55

​

​

AI COMMONS for Global Health

​

Objective: �

  1. Shape how the AI Commons for Global Health can best serve YOU, as an innovator, implementer, government, funder, multilateral, or healthcare professionals.�
  2. Identify unaddressed ecosystem bottlenecks that will prevent AI from contributing to health improvements for people living across the continent and beyond.

​

Guiding Question:

  • How can the AI Commons for GH best engage and serve YOU and what would you like to contribute to the AI Commons?�
  • What unaddressed shared challenges will prevent AI from having a transformative health impact in Africa and beyond? Are the challenges best addressed by collective action and public goods?

32

32

33 of 33

Thank you!

​

Please fill out this form to let us know how you’d like to be involved.

​

​

​

AI COMMONS for Global Health

​

33

33