Lecture notes

Day 8

Lectures from Washington, D.C., Wednesday, August 5, 2026.

Washington, D.C.

Wednesday, August 5, 2026

3 lectures

Open the full notes (PDF)

Lecture 1

Official Launch of the World Development Report 2026: The Promise of Artificial Intelligence

A lecture by Gaurav Nayyar

Director, World Development Report 2026, World Bank Group.

The bottom line

The World Development Report 2026 reframes AI from a technology story into a development story. Its message to developing countries is threefold: be optimistic, because AI's capabilities can compress into a decade what once took a century; be strategic, by starting from adoption but treating adaptation to local data and conditions as the main task; and be pragmatic, making cost-conscious choices as enablers, users, and regulators of AI rather than chasing frontier systems or an "AI arms race."

The mindset: why developing countries should be optimistic

Most of what we read about AI concerns advanced economies; the report deliberately asks what AI means for development. It treats AI as a general-purpose technology (joining steam power, electricity, computers, and the internet) best understood through three features: the capabilities it brings, the concentration in its production, and the complements it needs to work in practice.

That optimism rests on capabilities that are scarce in many developing countries. Because AI can help solve complex problems, it "can do in a decade what might otherwise have taken a century." Early evidence is concrete: AI-generated weather forecasts in India help both optimistic and pessimistic farmers adjust their production and investment choices; Kenya's judiciary uses AI to allocate court cases to mediators better than the previous manual process; and in Bangladesh, AI-supported medical imaging is raising patient screening by almost a third to 40% per day amid a global shortage of radiologists. Most of these gains come through public-service delivery and often from older AI models (predictive analytics, remote sensing, early-warning systems) not chatbots; meanwhile job disruption is less imminent than in high-income countries, where a larger share of jobs is amenable to automation.

The strategy: adopt, but above all adapt

The good-news story has two caveats. Concentration is one: nearly all leading AI companies sit in just two countries (the United States and China) across every layer of the value chain. The report's answer is not to chase "AI sovereignty," which is too expensive for most developing countries and would only fragment the technology; it is to navigate the dependency with interoperable systems and to use others' massive investments to customize models locally. The bigger risk is a lack of complements: without the basic infrastructure, skills, and institutions, adoption (and therefore AI's productivity and growth benefits) will be slow to take off, which is why projected gains are larger in advanced economies at current adoption rates.

This is why adoption alone is not enough. AI is unique in that it learns from data, and data must reflect local languages, needs, and conditions; simply importing a tool from abroad may not work and can introduce bias. Advancing the frontier is relevant only for a small number of well-resourced countries; for everyone else, adopting available AI is the logical start, but adapting it to context unlocks the value. This echoes the "investment–infusion–innovation" arc of the middle-income-trap report, with adaptation as infusion, except that with AI, adaptation must begin almost in parallel with adoption.

Evidence at a glance

Excitement tracks development needs. Across about 30 countries, Asia and Africa are more excited than concerned about AI, while Europe and the Americas are more concerned than excited.
Capabilities in action. India (AI weather forecasts for farmers), Kenya (AI case allocation in the judiciary), and Bangladesh (AI medical imaging raising screening by almost a third to 40% per day) show AI addressing long-standing development gaps.
Analog complements bind hardest. In a firm survey across several developing countries, businesses of all sizes cite lack of knowledge, information, awareness, or finance (not just digital gaps) as what holds AI adoption back.
Pilots are not proof. Of about 10,000 studies on AI in healthcare, fewer than 1% offered meaningful evidence that AI improved services, the world is full of experiments, not scaled solutions.
The least prepared may gain most. About 800 voluntary AI standards are already published or in development; yet while all high-income and most middle-income countries have a national AI strategy, only 1 of 25 low-income countries does.

The policy agenda: enabler, user, regulator

Make it your own. The report reframes a technology question as a development-policy question. Governments have three roles to play, and the right sequence is to build foundations, learn what works, and earn trust, all under tight budgets.

Enable. Provide the foundational infrastructure and skills (connectivity, electricity, local-language data, and some compute) recognizing that analog complements matter even more than digital ones.
Use. As major buyers of AI, governments should reform procurement so it is continuous rather than one-and-done and not simply cheapest-bid, with frameworks to procure, evaluate, and scale the solutions that work.
Regulate. With limited capacity and fast-moving technology, start with voluntary rather than mandatory standards, avoid rushing to legislate economy-wide, and make standard-setting more inclusive.
Favor open source, avoid lock-in. Open models let countries truly customize; India, the UAE, and others have built local language models on open-source foundations that outperform frontier models in their own languages, but keep systems interoperable to avoid dependence on any one vendor. The report is the third in a trilogy, the middle-income trap, standards, and the promise of AI, built on one framework; next year's turns to jobs.

Lecture 2

How Ready Are We for AI: New Insights from the BReady4AI Index

A lecture by Ufuk Akcigit

Co-Director, Growth Academy; Arnold C. Harberger Professor of Economics, University of Chicago (on leave); Deputy Chief Economist and Director of Private Markets, World Bank Group.

The bottom line

The headlines about AI come from large incumbents laying off workers, but most of the economy (and most of the developing world) is made of very small firms. New microdata show that the smallest businesses are the most likely to adopt AI and that adopters see more job creation than destruction and higher revenue productivity. And a new, living "BReady4AI" index reveals that countries' national AI strategies often do not match their real bottlenecks, making awareness of those gaps the first task of readiness.

Creative destruction meets a general-purpose technology

Akcigit opened where the Growth Academy began, with creative destruction. As with past general-purpose technologies, AI's cost-reducing effect will land first, displacing jobs, before the economy adjusts. Whether that adjustment is benign depends on dynamism, on new entry and young people bringing ideas we cannot yet imagine. If economies instead cling to old goods and simply make them more efficient, that is process innovation, and process innovation costs jobs. The optimism, then, rests on product innovation and on firms that scale.

The trouble is that almost everything we know about AI comes from large incumbents (Amazon, Oracle) yet firms with nine or fewer workers make up about 80% of U.S. businesses, and developing-country firms are smaller still. The left tail of the U.S. firm-size distribution is therefore the most informative window into AI's likely path in developing countries. With co-authors including James Evans and Enrico Colonnelli, Akcigit turned to real-time data from Intuit QuickBooks, the largest online platform for small businesses, with more than 10 million customers, complemented by the team's own surveys.

What the smallest firms reveal

Surveys say about 70% of small-business owners "use" AI, but usage is a fuzzy word. Following the logic that firms pay for what creates value (much as patent-renewal fees reveal a patent's worth), the team looked at spending: only about 10% of small businesses pay $10 a month and just under 1% pay $100 a month. Adoption is real but early and experimental. It is highest in information and professional services and lowest in agriculture, a sobering fact for developing economies where agriculture looms large.

The most striking pattern is a U-shape: adoption is highest among the very smallest firms. A one- or two-person business that cannot afford to hire (adding a worker can nearly double its costs) is the most desperate for help, and AI offers it cheaply. Ambition matters too: firms whose owners say they want to grow are far more likely to adopt than "subsistence" ones. And the outcomes echo creative destruction: most adopters report no change in employment, but among those that change, job creation exceeds destruction and revenue growth outpaces decline, so revenue productivity rises. The main barriers are privacy concerns and, tellingly, lack of knowledge, which correlates strongly with the entrepreneur's education.

Evidence at a glance

Small firms dominate. Businesses with nine or fewer workers are about 80% of U.S. firms; the developing-world equivalent is smaller still, making the U.S. left tail the best available proxy.
Adoption is early. About 70% of owners report using AI, yet only about 10% pay $10 a month and just under 1% pay $100 a month, firms are still experimenting.
The U-shape. The likelihood of adopting AI is highest for one-worker firms, then falls with size, the smallest, most financially constrained businesses adopt most.
Ambition predicts adoption. Transformative, growth-oriented entrepreneurs are much more likely to adopt AI than subsistence firms content to stay stable.
Creation beats destruction. Among firms that change, AI adopters report more job creation than destruction and stronger revenue growth (a net gain in revenue productivity) while power outages and data-center gaps constrain adoption.

The BReady4AI Index, and its cautions

Every country is like a fingerprint. A living index, to be released on bereadyforai.com, lets policymakers track where nations stand as AI evolves; but its measures are indicators of what countries do and say, not verdicts on welfare or productivity.

Measure the whole ecosystem. The index draws on academic papers, patents, GitHub activity, data centers, and government AI grant registries (with syllabi and legal documents being added) across four pillars: frontier creation, global influence, diversity and collaboration, and compute and talent. Early scorecards rank the U.S. first, China second, and the U.K. third.
Read national strategies in three dimensions. Processing 80-plus national AI strategies by frequency, intensity, and direction (expansionary versus safeguarding) shows sharp contrasts, the U.S. strategy is dominated by security and safeguarding, China emphasizes R&D and industrial adoption expansively, and the EU leans toward safeguarding regulation.
Watch for mismatch. Combined with the World Bank's governance indicators, the data show weaker rule of law linked to more emphasis on controlling AI, and (the red flag) some countries with frequent power outages whose strategies barely mention energy.
Do not read expansion as "good." As Akcigit and Evans cautioned, expansionary effort can breed market concentration that kills competition and future innovation, and stated strategies may diverge from actual laws, so the index is a starting point for analysis, not a scorecard of success.
Build the human-capital complement next. The planned education pillar reflects the core message: AI pays off when combined with other skills, so what matters is whether farmers and health workers (not just AI specialists) can adopt it.

Lecture 3

Learning from 80 Years of Lending: Economic Returns and the People Behind World Bank Projects

A lecture by Somik Lall

Co-Director, Growth Academy; Director, Strategy and WBG Institute for Economic Development, Development Economics Vice Presidency, World Bank Group.

The bottom line

The best way to know what an institution does is to see where it puts its money. Eighty years of World Bank infrastructure lending have delivered high economic returns (about 24% on average) and, surprisingly, the returns are highest in the poorest-governed, most capital-scarce countries. What makes the difference is not the country's institutions alone but the World Bank's own performance, and above all the individual project leaders who make investments deliver on the ground. The uncomfortable corollary: the institution rewards those people far too little.

Put the money where the mouth is

You can learn from research and from practitioners, but the surest test of what an institution believes is where it actually lends. The World Bank's first loan was a one-page letter from the government of France on 25 June 1946, $500 million for reconstruction, with no economic analysis and no rationale, just a request for help. Eighty years later, the question is whether all that lending has paid off. To answer it, Lall and colleagues in the research and development-economics group synthesized the social (economic) rates of return on IBRD- and IDA-funded infrastructure over time and across sectors.

This is deliberately different from the work presented the day before by Peter Henry, which benchmarked the IFC's financial return against yardsticks like the S&P 500. The IFC is an impact investor, not a competitor to standard investors; the IBRD's mandate is economic and social value, not private-market profit. So this exercise measures economic value, and every input is public: the underlying investment data can be downloaded and re-run through a reproducibility package. As Lall put it, "your country paid for this, this is your data." It matters because infrastructure is a long-lived bet: a road or dam delivers for 60–70 years, a very different frame from a health or education project whose benefits show up over 8–10.

Systems, and the people who beat them

The standard aid-effectiveness literature says aid works only where governance, rule of law, and institutions are strong, which implies that fragile and conflict-affected economies should earn low returns. Lall calls that nonsensical, because those are exactly the places that need concessional finance most. What makes high returns possible in hard places is the system an institution puts around a project, and, crucially, the individuals who run it. Much of the measured return cannot be attributed to systems at all; echoing John List, it depends on people who are hard to scale: the task team leaders who work inside project management units, manage procurement, and make sure the money builds the road rather than ending up in a Swiss bank account.

To see this, the team built an unusually deep dataset: economic rates of return for completed projects over 60 years (available for about half of them), matched for the first time to loan-disbursement records from borrowers' own treasuries, to 50 years of confidential CPIA governance scores, and to the scraped career histories of 1,600 individual task team leaders drawn from archives reaching back to the 1940s.

Evidence at a glance

Returns are high and rising. Economic rates of return average about 24% (median 18.5%); after a structural break around 1997 they rose from roughly 15% to 21%, the opposite of the usual finding that infrastructure returns decline over time.
Poorly-governed countries do better. Countries below the median CPIA earn higher returns than better-governed ones, suggesting scarce capital is flowing where needs are greatest.
Institutional performance pays. In weak-governance countries (CPIA below three) base returns run about 23%; strong World Bank supervision and stronger project units add roughly 12 percentage points.
Timing is asymmetric. Getting the first disbursements out early raises returns, but rushing a project to close does not, the final stages are where monitoring catches the roads and bridges that were never built.
The leaders are exceptional. About 30% of task team leaders hold PhDs; only 5% are hired locally; 10% come through the Young Professional Program (about 100 chosen from 200,000 applicants). A typical infrastructure project runs about $90m over roughly 2,600 days (seven years) at around 3.5% interest.

What it means for the institution

Value comes from capability, not just capital. The Bank adds value less by moving money than by building the supervision, accountability, and teams that turn a loan into delivered returns, and it should reward the people who do that work.

Fix the talent gap. The Bank practices negative assortative matching, sending its best people to its hardest environments, where they deliver returns near 40%, yet those same experts average about 1.5 promotions against a possible six. Create tracks that let them keep technical expertise while gaining real influence.
Do not confuse speed with success. Pressure to compress implementation timelines is misguided; development is a long business, and cutting supervision short is where projects fail.
Treat the estimates as conservative. Returns count only directly attributable social value (for example, reduced travel time and trade costs), excluding second-order effects such as land prices, so the true benefits are likely larger.
Keep a dual focus. Maximize economic value through better project design in high-need contexts and pursue broader institutional reform; the two are complements, not rivals. Getting results on the ground is hard, and the participants' own task is to build a chain from problem to action to scale.