Why AI for good still isn’t scaling — and a new effort to fix it

Kanika Bahl is stepping down from her CEO role at Evidence Action, a U.S.-based nonprofit organization that has scaled low-cost health interventions to reach more than 500 million people in Africa and Asia, to lead a new initiative within the organization focused on a different scaling challenge. She’ll be leading the AI Access Initiative, a new effort — to be spun out as a separate organization — that will work to strengthen the AI for good ecosystem in order to ensure the technology delivers meaningful benefits for the 3.5 billion people living in poverty in low- and middle-income countries. Under Bahl’s leadership, Evidence Action has grown into an 850-person organization and become a favorite among tech sector donors associated with effective altruism, a philanthropic approach that uses rigorous analysis to direct resources where they can do the most good. That stems in part from a disciplined focus on finding what Bahl calls the “unicorns” of international development — programs that are lifesaving, low-cost, evidence-backed, and scalable. Now, she is applying that same lens to a new frontier, trading the helm of a large, established organization for the challenge of building an AI-native effort from the ground up. Bahl brings a unique dual perspective to this work as a leader in evidence-based scaling as well as a founding member of Anthropic’s Long-Term Benefit Trust, an independent body that the artificial intelligence company formed to help align its corporate governance with its mission to ensure that advanced AI serves humanity. “It became increasingly clear, this was about a year ago, what an inflection point we're at, and how fundamental the next few years will be to determining whether AI becomes a greatest leapfrog technology in history, or whether it widens global inequality for decades to come,” she said during a recent Devex Pro briefing. “We need to have coordinated action now to build the ecosystem. Otherwise, we really just risk repeating the history of what we've seen with even just the most basic health technologies, where it could take decades to scale in lower-middle-income countries, versus just one to two years,” she added. Her plans for the AI Access Initiative include scaling a small set of AI-enabled “big bets” she believes could reach tens of millions of people over the next five years, alongside work to strengthen the government systems and enabling environments needed for national adoption. Doing so, Bahl said, will require more roles like the one she’s hiring for — people who can bridge AI labs, researchers, implementers, and governments to turn technical breakthroughs into real-world impact. The decision to step aside from day-to-day leadership at Evidence Action followed an effort by Bahl to understand why AI for good efforts had struggled to move beyond experimentation. She spoke with some of the smartest people she knew in the space, including experts now advising the AI for Access Initiative, such as Anthropic CEO Dario Amodei, Nobel laureate Michael Kremer, and Alphabet President of Global Affairs Kent Walker. Eventually, she got Evidence Action colleagues involved in the effort, which led to the publication of an AI for good cross-sector analysis, a snapshot of initial research into AI big bets, which found that most LMIC impact will hinge less on novel models than on the unglamorous enablers, including access to hardware, local data, and government pathways to scale. Despite growing interest and investment in AI, Bahl found there have been too many disconnected pilot projects, and too few systems built for national adoption, creating what she called “digital graveyards” — AI projects that die at the pilot stage because they were never set up for scale. “Without coordination, there's a real risk that upstream technical actors, with all of the best goodwill, build upstream solutions to just languish LMICs, because they're not well designed for users,” she said. Bahl said she sees three key issues with the current AI for good ecosystem: fragmentation across partners and geographies, pilots without planned scaling pathways, and inadequate funding mechanisms. “Today, funders are spending $250,000 per grant in AI,” she said. “I think that needs to be at the scale of $25 million, and moving a lot more quickly than the typical grant cycles.” Bahl and colleagues at Evidence Action, which is yet to name its new CEO, used some of the same criteria from its work across nine countries, including evidence, cost-effectiveness, and tractability, to narrow the list of AI use cases down to three potential big bets: clinical decision support to improve diagnostic accuracy in overstretched health systems; AI-generated weather forecasting for smallholder farmers, where higher-resolution forecasts can translate quickly into income gains; and direct-to-consumer health information to reduce delays in care-seeking. Bahl was explicit about what the AI for Access Initiative will and will not prioritize in the near term. “We are not necessarily focusing on the poorest of the poor first,” she said. “We are focusing on the places where there is the most tractability and simplicity.” She acknowledged that this approach presents tradeoffs. But the goal, Bahl explained, is to learn quickly, demonstrate impact at scale, and use those lessons to extend AI-enabled interventions into more complex and harder-to-reach contexts over time. She also highlighted the opportunities for partnership with the AI Access Initiative, especially as the team is likely to run into challenges they will need to iterate on — with governments, AI labs, and technical staff. “AI is moving very rapidly, and we think it is critically important to start learning by doing,” she said. Devex Pro members can watch the full briefing recording for deeper discussion on AI risks and safeguards, aligning incentives with frontier AI companies, and building AI fluency in development organizations, including a robust audience Q&A.