AI Philanthropy Is a Chance to Entirely Reimagine Service Delivery

My CGD colleagues have recently published a series of blog posts on how to spend a potentially large increase in philanthropy, driven by fortunes created by the AI boom. Recommendations range from scaling proven cost-effective interventions, such as malaria bed nets and severe wasting treatment, to spurring new innovations. While I agree that there's major scope to scale up proven interventions, this conversation has yet to grapple with a core feature of the AI zeitgeist: transformative AI could create entirely new problems and opportunities that the development sector is poorly positioned to address or seize. Alongside persistent challenges like malaria and poverty, we may face new challenges, like widening inequality and the erosion of traditional growth pathways. At the same time, AI is giving us the opportunity to accelerate advances in sectors like health and education.

AI philanthropy should bet big on innovation to discover novel solutions that both capitalize on AI and address the problems it brings. This blog argues for a specific kind of new solution—one that involves a wholesale reimagining of public service delivery in low-and middle-income countries (LMICs). The goal is both to find more cost-effective programs and to transform service delivery in ways that make new impactful development practices possible. Before diving in, I illustrate the difference between old and new problems and solutions in the table below:

* The New Solutions category includes methods to promote and discover innovations since such solutions have not, by definition, been identified.

What if the largest gains come from redesigning the factory floor?

Colleagues often ask me what is likely to be the most cost-effective AI-enabled intervention, and I am excited by the many impactful AI applications that will likely be developed. However, my concern is that the lens typically used to identify both new and existing cost-effective interventions may miss the promise of general-purpose technologies like AI. The largest gains from such technologies may come not from individual applications of the technology but from widespread diffusion and organizational redesign.

Consider electricity—when it first reached American manufacturing, factory owners often replaced central steam engines with electric motors while leaving the cramped floor configurations of belts and shafts that those engines required in place, leading to underwhelming returns. Only later did firms fully reconfigure the factory floor around a new sequence of production, combining the redesigns with cheap electricity to raise productivity. Electricity didn’t just make power more available; it enabled new and transformative ways of working. A century on, AI is tracing a similar path. Using US Census Bureau data on American manufacturers, researchers find AI adoption follows a "J-curve": firms suffer short-term losses as they undergo costly reorganization and widespread diffusion before gains arrive.

This is intuitive when looking at past general-purpose technologies. If you were charged with electrifying a rural health facility today, you wouldn't just electrify the X-ray machine or blood bank—whichever application was likely to be most cost-effective. You’d electrify the whole facility. An electrified facility would do more than power multiple medical devices; it would provide the chance to change the usual way of doing business by extending working hours and augmenting staff roles, ultimately elevating patient care. The same holds for another general-purpose technology, personal computing: its full value emerges only when used across multiple applications in, say, a hospital. Wide diffusion also establishes a platform that unlocks future innovation. For example, digitizing both patient records and a hospital’s medicine inventory allows you to use predictive tools that preempt medication stockouts.

This is more than a call for bundling multiple AI-powered applications—general-purpose technologies deliver the most impact when workflows and roles are redesigned around their widespread use. Consider a junior doctor’s assistant. Transformation is not simply giving her an AI decision-support tool, AI scribe, and AI patient interface to perform the same tasks more efficiently. AI could enable her to take on triage previously performed by senior nurses or doctors, while scribes automate documentation and other tools handle other time-consuming tasks she may previously have had to do. Task-shifting already improved outcomes and reduced costs before AI; AI could compound these gains.

Going beyond health, we could also imagine a redesign in education. Schools facing an existing problem of chronic teacher absenteeism could use AI to teach classes that would otherwise go untaught, with scarce teachers focusing on supervision and one-to-one support—filling gaps that rural recruitment has failed to close for decades. This fundamentally recasts a teacher from instructor to supervisor. Mobile AI tutoring could extend caregiver-supervised instruction into homes. This may again change the caregiver’s role in a child’s education.

Because the effect of each use depends on the portfolio of other uses and redesigned staff roles across the facility, deploying one application (or a bundle), holding all else constant, may bear little relation to the technology's value were it to fully diffuse through an organization. Focusing just on which individual AI application is most cost-effective may miss the possibility that the greatest returns come from multiple applications and workflows in a facility entirely redesigned around AI.

The development sector often dismisses such blue-sky thinking as Silicon Valley hype. But reimagining service delivery need not be outlandish. It can build on evidence—for example, that edtech often fails because of weak sustained adoption rather than the technology itself, and that adult supervision can substantially increase learning gains from computer-assisted learning. It must also confront operational realities in LMICs: a reimagined rural hospital may use AI to provide top-quality clinical advice, but patients still need access to quality-assured medicines.

How do we fund the reimagining of service delivery?

Many companies, health facilities, and schools in high-income countries have much larger profit incentives than those in low-income countries to go through the painful process of redesigning their business processes around AI. My concern is that if public financing and philanthropy doesn’t correct for this lack of profit incentive in development settings, and instead focuses just on individual cost-effective interventions, we may fund highly impactful applications but miss getting on the J-curve altogether.

I'll end somewhat unsatisfyingly, without a list of specific funding opportunities—because we first need to design the right vehicle to drive this transformation. Major innovation prizes offer partial precedents, having incentivized task-shifting and delivery model redesign for disease verticals (Gates Foundation Grand Challenges), workflow redesign for governments (Google.org's AI for Government Innovation challenge), and tech-enabled changes to individual interventions (Global Learning XPRIZE). None to my knowledge has targeted AI-driven, wholesale redesign of service delivery, including a systematic rethink of staff roles and responsibilities in LMICs.

One no-regrets first move is to convene the right people to design mechanisms that incentivize this transformation—work that demands perspective from multiple domains. Technologists and management consultants are already thinking about AI-driven transformation in high-income-country businesses; a 2026 Deloitte survey of 500+ US leaders found 74 percent expect nearly half their business processes to be redesigned around AI agents within four years. But they often lack context on LMIC schools and hospitals, implementation constraints, and the development evidence base. Development practitioners bring that expertise but think less in terms of general-purpose technologies, and many who lived through earlier telecommunications and digitization transformations have yet to update their thinking for AI. Meanwhile, LMIC governments I hear from are still struggling with digital transformation even as the AI revolution accelerates. Bringing these perspectives together will require deliberate effort—an agenda my team at CGD is considering pursuing.

While a diverse group will help the conception of a mechanism to incentivize transformation, I suspect that the right vehicle will have the following broad outlines:

• Large prizes or other pull mechanisms will likely be part of the picture. We do not yet know what redesigned service delivery with facility- or organization-wide AI diffusion looks like, and there may be multiple highly cost-effective configurations. Because we cannot specify the solution ex ante, we cannot simply fund its implementation. Instead, funders may need to set a target—perhaps an impact threshold alongside a defined standard for AI-centered transformation—and reward organizations that meet it.
• We might need to fund a social entrepreneurship fellowship or some program that embeds innovators in existing systems. A business school professor once pointed out to me that startups and venture capital face this problem all the time: knowing a transformative solution is possible without knowing its configuration ex ante. Incubators and fellowships address this by funding entrepreneurs to experiment and pivot until they find what works. Discovering these “unknown unknowns” may similarly require embedding entrepreneurs in LMIC schools and health facilities to observe pain points firsthand and continually reimagine how AI could redesign business as usual.
• We will likely need to fund a portfolio of bets. By definition, there is no proven solution to fund. Many efforts will fail as organizations iterate toward a handful of configurations that work. AI philanthropy should therefore fund a portfolio of redesign efforts, expecting failures and accounting for them in expected-value calculations.
• We have to think carefully about what’s influenceable with outside funding. Transformation will face lots of pushback, and most development actors can influence only discrete parts of an organization or facility. AI philanthropy therefore needs partners that “own the factory floor”—such as private healthcare facilities or schools, or an enterprising LMIC minister willing to pilot a portfolio of transformations.

I hope readers don’t misunderstand my argument—I love a good cost-effective intervention and have worked on both proven and novel ones. I expect the development sector to build many cost-effective AI interventions, which should be funded. I'm also not arguing for general "systems strengthening" investments over interventions—I'm making a narrower case, specific to how general-purpose technologies deliver returns.

Ultimately, alongside funding what works, we should capitalize on the distinctive potential of general-purpose technologies and explore entire redesigns of service delivery around AI. Doing so may produce configurations that are more cost-effective than standalone interventions. If you would like to support this redesign, please reach out.

Acknowledgments: Many thanks to Leah Rosenzweig, Daniel Bjokegren, Markus Goldstein, Yolanda Yang, Oliver Hanney, Joseph Levine, Mehdi Sharifkazemi, Tim Ohlenburg, Shah Wani, and Ravi Nirmal for comments.