What we learned from our series on AI.
You can listen to this podcast on Spotify, Apple Podcasts, or wherever else you get your podcasts. You can also watch this conversation on YouTube, and we've also started a Substack for these episodes.
Across our series on AI, Deena Mousa and I spoke with researchers and policymakers on topics relevant to the AI and development debate. In this episode we took stock of our main takeaways from these conversations, explored some of the many unanswered questions, and recapped what Deena learned at the AI Impact Summit in India.
What happens when AI meets the world?
There are fifty million cases backlogged in India’s courts. An organisation called Adalat AI is building AI tools to help, such as their transcription tool. When the founder, Utkarsh Saxena, described the deployment experience, one detail stood out. Judges, confronted with a new piece of technology that could transform their work, were stuck on the most basic questions.
“How do I update Chrome?”
This is a good summary of what this series has been about – what happens when AI meets the real world? This tends not to be the question technologists are answering when they talk in broad strokes about the capabilities of their models, and how fast they are improving, against different benchmarks. But economists ask a different set of questions, what actually changes in an economy when a new technology arrives? Who captures the gains? How long does it take? What gets in the way?
Those second-order questions are the ones that determine whether a technology that works in a lab changes anything for people living in Nairobi or Karachi.
What is the binding constraint?
This is a useful framing to have in mind when looking to apply AI in any development context. Ask: what are the binding constraints here, and does AI relax them?
There is an instinct to layer AI onto whatever process already exists, but the more useful approach starts from the problems which AI is actually good at solving, and how to put AI in a position to help.
The series gave several clear answers about where the constraints actually are. Umar Saif put the first most sharply: large parts of the world are data deserts, where the foundational information AI needs to be useful simply does not exist. No patient records to train a diagnostic model on. No land registries. No systematic crop yield data. Rose Mutiso made the same point from an infrastructure angle: the absence of underlying data isn’t a problem of being behind on adopting the latest model. It’s a problem of missing inputs.
This puts the national AI strategy conversation in a different light. Announcing a national AI policy produces summits and documents, and a new technology park might play well in a TV ad. Digitising land records is not sexy, and bumps up against a lot of political economy constraints. But the latter is what actually creates the conditions for AI to be useful.
What universities actually do
Josh Lerner’s research added another dimension to the foundational investments that matter. The finding that stuck was about place: faculty members who moved from Duke to Stanford became more commercially oriented, while the reverse move had the reverse effect. Roughly 20 - 25% of that shift was attributable to the environment itself, i.e. the norms, the networks, proximity to investors, and the institutional culture around risk-taking.
For LMICs, the implication is that funding more research or training more scientists is not sufficient on its own. The environment in which those scientists work determines whether their ideas become companies, and whether those companies stay. Building that environment is a slower and less legible investment than building a data centre, which is precisely why it tends to be underinvested.
The disruption to development escalators
Perhaps the most uncomfortable finding across the series was not about the obstacles to adopting AI, but about what AI may do to growth models that currently serve as a ladder of development.
The traditional path ran from agriculture to manufacturing to services. India rode services, including software and back-office work. Many other countries had been hoping to board a similar escalator, particularly given the difficulties of replicating East Asian export-led manufacturing growth in 2026. But as Umar pointed out, many basic Fiverr-style tradeable services – copyediting, logo design, SEO-optimised content, basic coding etc. – will be displaced by AI relatively rapidly. The door is slamming shut on the opportunities which India took advantage of around the Y2K scare, and it’s unclear what, if anything, will replace that bottom rung.
Raghuram himself was optimistic about India, as their labour cost differential persists even when both workers have AI. But India has progressed beyond the bottom rung of basic service exports. Whether that optimism extends to lower-income countries just trying to get onto the ladder is a genuinely open question.
So developing countries are in a double bind: AI is harder to adopt in LMICs because the prerequisites are missing, and simultaneously it is threatening the growth model they are already running.
What is the bottom rung in 2030?
Manufacturing worked for a very specific combination of reasons: it was tradeable, scalable, absorbed large numbers of workers with modest skills, and generated productivity growth through learning-by-doing. Some services were able to replicate these properties. What AI does to services, and could do to manufacturing (pending cheaper, better robots on the factory floor), is precisely target the tradeable, routine, scalable parts of both.
Bruno Caprettini’s research on the industrial revolution is useful here. The disruption from threshing machines was less damaging near industrial towns, because conditions existed for workers to move toward emerging alternatives. The policy implication is therefore not to identify the new escalator and build toward it. It is to make sure the conditions exist for workers to match with whatever opportunities do emerge. Strong education systems, social protection, connectivity, and urbanisation that puts people in proximity to economic activity.
Access versus capture
Josh Lerner’s research also pointed to a distinction that tends to get lost in AI and development discussions: the difference between access and value capture. Access means you can use a technology. Capture means you participate meaningfully in the value it creates, the jobs, the tax revenues, the firms. The historical pattern suggests those can diverge for decades. A world where every country gets access to excellent AI tools is entirely compatible with a world where the majority of financial gains accrue to a small number of hubs for decades.
The partial counterpoint Josh identified is South-South knowledge exchange. When China rose as a major innovation hub, the effect spread: entrepreneurs in other emerging markets began adapting business models forged under similar constraints in China, e.g. limited credit card penetration, unreliable infrastructure, sparse data. The constraints in one emerging economy are often much closer to those in another than to those in the United States, so Chinese innovations travelled in ways Silicon Valley innovations did not.
AI working through humans
Across the concrete cases in the series, a consistent pattern emerged: AI performing best when it works through, rather than instead of, people. Adalat AI is not replacing judges, it is removing the administrative burden that essentially turns them into scribes, enabling them to hear more cases.
And for the smallholder farmer who has never had access to an agronomist, the student whose teacher has forty children in a class, the counterfactual to AI assistance is not a professional being displaced. It is nothing. That is a fundamentally different economic calculation, and one that points to AI’s particular potential in many development settings.
What remains unresolved
The measurement problem Anton Korinek and Josh Lerner both flagged is a key challenge. Counting job postings, running benchmarks, and tracking AI usage data only captures who uses AI and how often, not what actually changes inside a job when someone does. A lawyer using an AI drafting tool might bill the same hours, work less hard, or take on more clients. And usage data is almost entirely invisible on displacement: the worker who was not hired because the firm used AI instead, the BPO contract that was not renewed. Those effects may be the most economically significant ones in the short term, especially in LMICs.
The cost trajectory changes many calculations, but perhaps not in the direction one might hope. Umar’s point that capable models may be running on phones within five years is plausible. If that happens, the access barrier largely dissolves. But that makes the data and infrastructure argument more urgent, not less. Cheap models running on bad data, in languages they were not trained on, without institutional scaffolding, do not represent a development opportunity.
And then there is the question of whether development economics as a discipline is positioned to grapple with any of this at its full scale. Running RCTs on individual AI applications is genuinely valuable, and there will be hundreds of them. But the big-picture questions – scenario planning, macro implications, growth model disruption – require the kind of speculative, system-level thinking that academic incentives currently work against. Oliver Hanney has expanded on this in his new blog - there is no randomising a technological revolution.
If you’ve enjoyed this series, please do share the episodes with your friends and colleagues.
Next up, seven episodes on cities with Kurtis Lockhart!