A note by EkStep Foundation, March 2026
A note by EkStep Foundation, March 2026
100 AI DIFFUSION PATHWAYS BY 2030
100 AI DIFFUSION PATHWAYS BY 2030
A race to deploy the most powerful technology in human history to benefit the billions who need it most.
Shouldn't technology reach those who need it the most, fastest?
Shouldn't technology reach those who need it the most, fastest?
“Productivity gains from AI follow a progression—first at the level of the individual, then the team, then the organisation, and finally society. The last step—societal productivity—is the hardest.”
“Productivity gains from AI follow a progression—first at the level of the individual, then the team, then the organisation, and finally society. The last step—societal productivity—is the hardest.”
Nandan Nilekani, Raisina Dialogue 2026
The most transformative technology in human history should reach the people whose lives it can transform the most, and soon, but deploying AI in the real world is extraordinarily difficult. It is expensive, complex, and risky at scale. A successful deployment journey goes well beyond technology - it involves assembling the resources and know-how, making choices, and acting within resource constraints to advance towards a desired outcome.
The most transformative technology in human history should reach the people whose lives it can transform the most, and soon, but deploying AI in the real world is extraordinarily difficult. It is expensive, complex, and risky at scale. A successful deployment journey goes well beyond technology - it involves assembling the resources and know-how, making choices, and acting within resource constraints to advance towards a desired outcome.
A diffusion pathway packages the lived experience of deploying AI at scale into playbooks and toolkits for other adopters, dramatically reducing their learning costs, time, and risk. Since AI is a general-purpose technology that could benefit billions—touching agriculture, health, education, governance, livelihoods, and dozens more—a single pathway isn’t enough. Each sector, country and organisational context carries its own data and resource constraints, trust relationships, and failure modes.
A diffusion pathway packages the lived experience of deploying AI at scale into playbooks and toolkits for other adopters, dramatically reducing their learning costs, time, and risk. Since AI is a general-purpose technology that could benefit billions—touching agriculture, health, education, governance, livelihoods, and dozens more—a single pathway isn’t enough. Each sector, country and organisational context carries its own data and resource constraints, trust relationships, and failure modes.
A diverse group of eleven organisations with overlapping self-interests and united by shared beliefs recently launched an effort to achieve 100 AI diffusion pathways by 2030. Diffusion pathways work. Here is the evidence, the beliefs and an invitation to join.
A diverse group of eleven organisations with overlapping self-interests and united by shared beliefs recently launched an effort to achieve 100 AI diffusion pathways by 2030. Diffusion pathways work. Here is the evidence, the beliefs and an invitation to join.
The Proof
Shared deployment knowledge compressed adoption time by over 90% across three successive adopters:
The Government of Maharashtra (India) built MahaVISTAAR, an AI agricultural advisory service, from scratch: institutional architecture, data pipelines, government alignment, and 30 partner organisations. Over 3 million farmers (of the 16 million) now use it.
Learn more
9
months
Maharashtra — MahaVISTAAR
Built from scratch: institutional architecture, data pipelines, government alignment, and 30 partner organisations. 3M+ farmers.
Built from scratch: institutional architecture, data pipelines, government alignment, and 30 partner organisations. 3M+ farmers.
The Government of Maharashtra (India) built MahaVISTAAR, an AI agricultural advisory service, from scratch: institutional architecture, data pipelines, government alignment, and 30 partner organisations. Over 3 million farmers (of the 16 million) now use it.
Learn more
3
months
Ethiopia — ATI
Drew on Maharashtra's pathway to build a comparable deployment in one-third the time
Ethiopia’s ATI , drawing on Maharashtra’s pathway, built a comparable deployment in one-third the time, without rediscovering which architectures work in the field.
Learn more
67% faster
3
weeks
Amul — Sarlaben
Two deployment cycles of shared learning available 3.6M farmers, 40M cattle, 2B milk transactions/year.
Amul’s Sarlaben launched, serving 3.6 million farmers, 40 million cattle, and 2 billion milk transactions a year—drawing on two full deployment cycles of shared learning already available.
Learn more
97% faster
Ethiopia’s ATI , drawing on Maharashtra’s pathway, built a comparable deployment in one-third the time, without rediscovering which architectures work in the field.
Learn more
3
months
Ethiopia — ATI
Drew on Maharashtra's pathway to build a comparable deployment in one-third the time
67% faster
Amul’s Sarlaben launched, serving 3.6 million farmers, 40 million cattle, and 2 billion milk transactions a year—drawing on two full deployment cycles of shared learning already available.
Learn more
3
weeks
Amul — Sarlaben
Two deployment cycles of shared learning available 3.6M farmers, 40M cattle, 2B milk transactions/year.
97% faster
What was shared across deployments was specific and actionable: reusable technical components, data governance models, documented failure modes, operational playbooks for institutional alignment, safety guardrails, training approaches refined in the field, all of which cut deployment costs in subsequent deployments. And the compression accelerated as the shared knowledge base grew, because each new deployment added to the stock of reusable assets available to subsequent adopters. The later adopters were better informed, not necessarily better resourced.
What was shared across deployments was specific and actionable: reusable technical components, data governance models, documented failure modes, operational playbooks for institutional alignment, safety guardrails, training approaches refined in the field, all of which cut deployment costs in subsequent deployments. And the compression accelerated as the shared knowledge base grew, because each new deployment added to the stock of reusable assets available to subsequent adopters. The later adopters were better informed, not necessarily better resourced.
What a Pathway Is, and Isn’t
What a Pathway Is, and Isn’t
A diffusion pathway is different from a case study, a white paper, or a best-practice guide.
A diffusion pathway is different from a case study, a white paper, or a best-practice guide.
It’s written from the next adopter’s perspective, not the builder’s. A case study tells you what someone did. A pathway gives you know-how - it tells you what you would need to do, in a specific context, and what assets you can take off the shelf instead of building from scratch.
It’s written from the next adopter’s perspective, not the builder’s. A case study tells you what someone did. A pathway gives you know-how - it tells you what you would need to do, in a specific context, and what assets you can take off the shelf instead of building from scratch.
Two components
A Deployment Playbook:
the documented lived experience of deploying AI, which goes well beyond models and hardware - Who needed to be in the room, which design choices mattered and which didn’t, how institutional trust was preserved, onboarding stakeholders and end users etc
the documented lived experience of deploying AI, which goes well beyond models and hardware - Who needed to be in the room, which design choices mattered and which didn’t, how institutional trust was preserved, onboarding stakeholders and end users etc
A Reusable Toolkit:
technical assets, data pipeline governance templates, safety guardrails, procurement strategies, evaluation benchmarks, training materials, scale-up and stakeholder onboarding strategies that another adopter can use directly.
technical assets, data pipeline governance templates, safety guardrails, procurement strategies, evaluation benchmarks, training materials, scale-up and stakeholder onboarding strategies that another adopter can use directly.
Why AI Deployment Is So Hard Without Shared Pathways
Why AI Deployment Is So Hard Without Shared Pathways
Three bodies of evidence confirm that AI’s real difficulty is institutional, not technological - a live deployment, a cross-sector study, and the historical record.
Three bodies of evidence confirm that AI’s real difficulty is institutional, not technological - a live deployment, a cross-sector study, and the historical record.
A Live Deployment
01
Building MahaVISTAAR required designing sustainable data pipelines that farmers and government institutions could trust, ensuring the system works across Marathi dialects, and catching failure modes (e.g., wrong dosage advice, misread crop conditions) before deployment. Several government bodies, data providers, and research institutions, as well as farm extension workers, had to align before a single farmer could use it. The AI model was the smaller part of a nine-month effort. MahaVISTAAR didn’t happen because one organisation got it right. It happened because thirty organisations decided that getting it right mattered more than getting it done quickly.
Building MahaVISTAAR required designing sustainable data pipelines that farmers and government institutions could trust, ensuring the system works across Marathi dialects, and catching failure modes (e.g., wrong dosage advice, misread crop conditions) before deployment. Several government bodies, data providers, and research institutions, as well as farm extension workers, had to align before a single farmer could use it. The AI model was the smaller part of a nine-month effort. MahaVISTAAR didn’t happen because one organisation got it right. It happened because thirty organisations decided that getting it right mattered more than getting it done quickly.
A Cross-Sector Study
02
The Use Case Adoption Framework 4 (UCAF) examined hundreds of AI use cases across agriculture, health, education, and government and found a consistent pattern- technology accounted for only roughly 30% of what reaching scale required. Factors like data readiness, language support, workforce adaptation, accountability guardrails, and organisation-wide alignment accounted for the remaining 70%, and this proved consistent across sectors and continents.
The Use Case Adoption Framework 4 (UCAF) examined hundreds of AI use cases across agriculture, health, education, and government and found a consistent pattern- technology accounted for only roughly 30% of what reaching scale required. Factors like data readiness, language support, workforce adaptation, accountability guardrails, and organisation-wide alignment accounted for the remaining 70%, and this proved consistent across sectors and continents.
The Historical Record
03
Jeffrey Ding’s research on general-purpose technologies shows that the electric dynamo took fifty years to produce widespread economic gains, not because the technology was unavailable, but because adopters independently rediscovered how to organise around it. Michael Mazarr’s RAND study and Sangeet Paul Choudary’s Reshuffle reach the same conclusion about AI: the competitive challenge is social and institutional, not technological. The fifty-year lag is a failure of knowledge sharing. Diffusion pathways compress this lag.
AI Adoption Journey for Population Scale | Carnegie Endowment for International Peace
Why diffusion pathways matter beyond the evidence
Why diffusion pathways matter beyond the evidence
Technology does not distribute and transform lives by itself. Smallholder farmers, under-resourced health workers, and teachers in underfunded schools are rarely the first to benefit from a new capability and often the last. Reaching them requires deliberate effort.
Technology does not distribute and transform lives by itself. Smallholder farmers, under-resourced health workers, and teachers in underfunded schools are rarely the first to benefit from a new capability and often the last. Reaching them requires deliberate effort.
No single actor is enough
Deployment knowledge spans organisations, sectors, and geographies. It only becomes useful when it comes together in a real-world deployment — and is shared for the next one.
The cost of waiting is real
It is measured in decisions made today without knowledge that already exists somewhere in the world. Populations that never see a benefit from AI become resistant to it — constraining markets, trust, and political will.
AI must add to agency
The goal is not algorithms that replace people. It is people empowered by knowledge that was previously out of reach — who understand more and judge better because of it.
3.6M
farmers
Amul System — A Concrete Example
Amul System — A Concrete Example
Across the Amul system, 3.6 million farmers manage livestock with access to only 1,400 vets. The goal is not an algorithm that replaces the vet — it is a farmer who understands more and judges better because knowledge that was previously out of reach is now available to her.
Across the Amul system, 3.6 million farmers manage livestock with access to only 1,400 vets. The goal is not an algorithm that replaces the vet — it is a farmer who understands more and judges better because knowledge that was previously out of reach is now available to her.
A farmer who becomes a better farmer. A teacher who becomes a better teacher. A health worker who makes better judgements. Not people made dependent on technology — people empowered by it.
A farmer who becomes a better farmer. A teacher who becomes a better teacher. A health worker who makes better judgements. Not people made dependent on technology — people empowered by it.
Who Has Committed
Who Has Committed
Twelve founding organisations committed in February 2026:
Twelve founding organisations committed in February 2026:
Anthropic, Carnegie India, EkStep Foundation, the governments of Ethiopia, the Gates Foundation, Google Data Commons, IIIT-B, Italy, Kenya, the Observer Research Foundation, Qhala and UNDP came together, on the sidelines of the AI Impact Summit in Delhi, and set a target of 100 pathways by 2030. Several are already live.
Anthropic, Carnegie India, EkStep Foundation, the governments of Ethiopia, the Gates Foundation, Google Data Commons, IIIT-B, Italy, Kenya, the Observer Research Foundation, Qhala and UNDP came together, on the sidelines of the AI Impact Summit in Delhi, and set a target of 100 pathways by 2030. Several are already live.
Videos & Articles
Videos & Articles
AI Diffusion & the 100by30 Commitment
AI Diffusion & the 100by30 Commitment
INDIA AI IMPACT SUMMIT 2026
INDIA AI IMPACT SUMMIT 2026
AI Diffusion at Population Scale
AI Diffusion at Population Scale
EkStep Foundation · Feb 2026
EkStep Foundation · Feb 2026
Watch on YouTube
Watch on YouTube
Use case adoption framework
Use case adoption framework
Published with Carnegie India
Published with Carnegie India
AI Adoption Journey for Population Scale
AI Adoption Journey for Population Scale
Shalini Kapoor & Tanvi Lall · Carnegie Endowment for International Peace · January 2026
Shalini Kapoor & Tanvi Lall · Carnegie Endowment for International Peace · January 2026
Why AI adoption at scale remains elusive. Introduces the Use Case Adoption Framework (UCAF) — moving AI out of pilot purgatory and into scalable deployment. Identifies cross-sector horizontal enablers: data readiness, language, voice, workforce reimagination, and accountability guardrails.
Why AI adoption at scale remains elusive. Introduces the Use Case Adoption Framework (UCAF) — moving AI out of pilot purgatory and into scalable deployment. Identifies cross-sector horizontal enablers: data readiness, language, voice, workforce reimagination, and accountability guardrails.
The Invitation
There are four ways to contribute to the effort:
There are four ways to contribute to the effort:
01
Illustrate a pathway
If you’ve deployed AI at scale or know of a deployment, document the lived experience - the architectural decisions, the institutional work, the failure modes that only surfaced in the field.
If you’ve deployed AI at scale or know of a deployment, document the lived experience - the architectural decisions, the institutional work, the failure modes that only surfaced in the field.
02
Use a pathway
If you’re ready to deploy, use what exists. Your starting point will be far ahead if you do so.
If you’re ready to deploy, use what exists. Your starting point will be far ahead if you do so.
03
Strengthen a pathway
If your expertise is in AI safety, workforce adaptation, data governance, or language support, contribute to and strengthen a pathway that benefits all deployments built on it.
If your expertise is in AI safety, workforce adaptation, data governance, or language support, contribute to and strengthen a pathway that benefits all deployments built on it.
04
Build a new pathway
If you see a sector, country or community where no pathway exists yet, build it there. Document the journey and share. The knowledge you create will travel to contexts you can’t anticipate, serving people you’ll never meet.
If you see a sector, country or community where no pathway exists yet, build it there. Document the journey and share. The knowledge you create will travel to contexts you can’t anticipate, serving people you’ll never meet.
100 Pathways to AI Diffusion by 2030
Will your organisation be
Will your organisation be
part of building it?
part of building it?
FREQUENTLY ASKED
QUESTIONS
Why Pathways? Why Diffusion?
The problem AI diffusion solves, and why structured pathways are the answer
Q0: Why does AI diffusion matter, and why do we need pathways?
A
What Is a Pathway?
Definition, structure, examples, and quality
Q1: What is a pathway, and what is it not?
Q2: Can you point me to an example pathway?
Q3: What does “100 pathways actually mean? Is there a quality bar?
Q4: How is safety built into pathways?
B
Participating in the Commitment
Who it’s for, how to join, and the value proposition
Q5: Who are pathways for?
Q6: How do I join?
Q7: Why would I participate in this commitment?
C
Running the Commitment
Operations, targets, and sustainability
Q8: What does joining the initiative require? Is there a financial commitment?
Q9: Is the approach by sector taking a sector s use cases global?
