By Anil Wasif

August 5, 2025

The AI revolution being pitched to the Global Majority is a seductive one: it’s the chance for low and middle-income countries (LMICs) to leapfrog again for the first time since the catapult over landlines directly to mobile phones.

But what if that’s not how the story goes?

At the World Bank’s Annual Bank Conference on Development Economics (ABCDE) last week, the case was made for a more grounded approach, one that values context, culture and resourceful adaptation over capital-intensive invention. During the session on AI, experts mostly argued against the keynote by economist and Queens College Cambridge President Mohamed El-Erian, who stressed the urgency of having the “fundamental wiring systems in place”. They contended that in this revolution, the kingmaker is not the algorithm, but the context.

World Bank economist and session chair Gaurav Nayyar noted that the debate around AI maps “almost perfectly” onto our anxieties about populism and distributional concerns—about who wins, who loses, and how influence will be redrawn. The consensus from the panel was that we are making the same mistake with AI that was made with electricity, when everyone thought its primary use was for illumination—replacing gas lamps—while completely missing the real revolution: the electrification of factories.

Princeton computer scientist Arvind Narayanan dismantled the hype, arguing that AI is subject to the same slow stages of adoption as any other general-purpose technology. He pointed to the vast “capability-reliability gap,” reminding the audience of the famous, failed prediction that AI would replace radiologists.

A decade later, their numbers have only grown, with AI serving as an augmentation tool, not a replacement. The reason, Narayanan argued, is that the “task model” of jobs is incomplete; the hardest things to automate exist at the nuanced boundaries between tasks. The real gains, he concluded, come from the unglamorous work of “adoption and diffusion,” making the capital-intensive race to build the biggest AI model a fool’s errand for LMICs.

A roadmap for Arvind’s alternative came from Daniel Björkegren of Columbia University’s School of International and Public Affairs, who provided two powerful examples that distinguished between different eras of AI. Björkegren, whose research involves digital transformation and applied machine learning with a focus on developing economies, began by cautioning on the world of predictive AI, citing the example of AI’s innovative use of mobile phone, social media usage, money transfer and pictures of receipts to create credit scores for the unbanked.

While a technical success for financial inclusion, here we must also recall our  experience of digital credit from the early 2000s, such as over-indebtedness and predatory lending. Similarly, there will be a fresh set of regulatory challenges governments will face, such as the welfare effects of the miscalculation of someone’s credit score, pushing them out of the state’s care. Innovation, he pointed out, often creates as many problems as it solves.

This set the stage for his more recent research on generative AI in Sierra Leone. The problem was a stark information gap. In a country where 95% of teachers use WhatsApp daily, only 3% used a web search to find classroom materials. The reasons were brutally simple: high data costs and a web where only 2% of search results originated from Sierra Leone itself. To its people, the internet was a foreign country.

A government cannot intelligently steer the adoption of technology if it cannot first intelligently understand its own streets, systems, and citizens.

Björkegren’s team introduced a simple AI chatbot on the one platform everyone already used: WhatsApp. Teachers could now ask for help in plain language: “Write a lesson plan to teach two-digit addition,” or “Describe immunization for a fifth-grade class in Sierra Leone.”

The adoption was immediate. The reason was simple. A standard web search, bloated by the advertising models that MIT’s Catherine Tucker warned about on the same panel, costs about 2.5 megabytes of precious data. The AI’s answer, delivered via WhatsApp? Under one kilobyte. The efficiency gain is over 3,000-to-one. The goal was to build a better door to the knowledge that already existed, designed for the specific constraints of a teacher’s budget and a nation’s digital infrastructure.

This is what business scholars call architectural innovation: reconfiguring existing components to unlock new value (long shorthanded in South Asian tech circles as jugaad, meaning an ingenious shortcut or improvised hack). A real-world example recently detailed in The Economist is the cheekily nicknamed “Bangla Tesla” electric rickshaw.

Contending to be the world’s largest informal EV fleet, with over 4 million e-rickshaws on the roads in Bangladesh, the human-EV hybrid has allowed drivers to dramatically boost their daily earnings, in some cases from 200 to 1,500 taka—a more than seven-fold increase. These market and welfare gains from the Bangla Tesla are a textbook example of architectural innovation.

At the same time, flimsy construction and high speeds contributed to over 870 fatal accidents in 2024 alone, alongside public health crises from lead-acid battery recycling – prompting Bangladesh to consider reinstating a previous high court ban that led to widespread protests. Dhaka’s Bangla Teslas also caution us on what happens when rapid transformations, like AI, outpace a government’s ability to see and manage them – revoking a social contract.

This challenge—harnessing the good of an innovation while mitigating the bad—is where state capacity becomes paramount. A government cannot intelligently steer the adoption of technology if it cannot first intelligently understand its own streets, systems, and citizens. To avoid the regulatory blind spots that allow new technologies to create new harms, the state must first build its own diagnostic rigor.

As laid out in the World Bank’s Government Analytics playbook, this means using its own data to see what is happening—analyzing procurement data for corruption, using customs data to detect tariff evasion, or using service delivery metrics to track actual student learning instead of just attendance.

This unglamorous work is the precondition for any effective reform. It is the state’s inoculation against the magical thinking that capital and technology can solve problems that are, at their core, about contextual governance. It gives officials the capacity to be discerning clients and wise regulators.

South Asia’s true competitive advantage is context. It is the innate understanding of what works—and doesn’t—on a crowded street, in a rural classroom, or in a supply chain held together by hope. This revolution will be local. And governments better be prepared.

Policy Columnist Anil Wasif is a public servant in the Ontario government. He serves on the University of Toronto’s Governing Council and the advisory board of McGill’s Max Bell School. Internationally, he serves on the OECD’s Infrastructure Delivery Committee and the World Bank Economic Development Institute’s Community of Practice. He co-owns and manages the Canada-born global non-profit BacharLorai. The views expressed are his own.