State infrastructure in a box: can AI help countries import good government?

In the spring of 2020, as COVID-19 lockdowns spread, Togo faced a problem familiar across the developing world: It wanted to rush emergency cash to its most vulnerable citizens, but had no reliable way of knowing who or where they were. The last census was nearly a decade old; there was no social registry. In a developed country, finding those citizens would be a simple database query; in much of the rest of the world, identifying them often requires assembling fragmented and outdated records.

Working with University of California, Berkeley researchers and the charity GiveDirectly, the government trained machine-learning algorithms on satellite imagery and mobile-phone records to estimate poverty down to individual mobile-phone subscribers, then paid them by mobile money. This was the rural phase of Novissi, Togo’s pandemic cash-transfer programme. It reached roughly 140,000 people that conventional systems would have missed. A function that normally takes decades of administrative infrastructure was accomplished in weeks, raising the question: If state capacity has long been understood as something states painstakingly build, can they instead import it?

For much of the 20th century, development economists believed poor countries were poor because they lacked things: capital, machines, roads, schools. Yet trillions in aid produced disappointing results, and by the 2000s the consensus had shifted. The binding constraint was not capital but institutions, and specifically state capacity: the prosaic ability of a government to collect taxes, enforce contracts, deliver services and administer rules competently.

The theme runs through two decades of development thinking. Economists Daron Acemoglu and James Robinson argued that the gap between rich and poor societies turns on whether their institutions are inclusive or extractive – work recognized with the 2024 Nobel Prize (economics). Francis Fukuyama devoted two volumes to how societies “get to Denmark” – a state that is capable, lawful and accountable, arguing that sequence matters most. Economist Lant Pritchett issued a sharper warning: In the “capability trap”, poor states adopt the laws and best practices of rich ones while remaining unable to perform the underlying functions. Loading too much onto these hollow institutions too fast makes them buckle – what Pritchett calls premature load-bearing. Real capability, he said, had to be grown slowly from within; it could not be transplanted.

Artificial intelligence unsettles that conclusion. Most discussion of AI in government – in the OECD’s surveys and think-tank reports – treats it as a productivity tool for already-capable bureaucracies. The more provocative idea is that AI might substitute for bureaucratic depth itself, as many core functions of a modern state are at bottom large-scale information processing; precisely what machine learning excels at. What makes this new is that AI can draw on data generated outside the state: satellite imagery, phone records, payment flows. This breaks the old circularity in which knowing your citizens required the very administrative machinery you lacked.

Most of this evidence so far comes from digital systems rather than AI itself but the mechanism is the same: Once machines solve a state's information problem, capability follows. India's Local Economic Intelligence Platform offers an early glimpse, using integrated economic data to help local governments understand and shape their economies. An IMF working paper finds digitalization of businesses and tax administration is associated with tax-to-GDP gains of up to three percentage points, particularly when firms and government digitalize in tandem. Such gains could be transformative in countries collecting less than 15% of GDP in taxes, a level often cited as the minimum needed to sustain basic state functions.

India’s digital identity and payments stack has, the government estimates, saved roughly $37 billion (₹3.48 lakh crore) by stripping fictitious and duplicate beneficiaries from welfare rolls; in Malawi, biometric identification expanded credit to borrowers the formal system could not see. In each case the machine supplies analytical muscle the bureaucracy lacks.

The most important objection is conceptual: A state that can process information is not the same as a state that is trusted, and even on its own terms imported capability is fragile. The Togo researchers themselves found the algorithm worked best as a supplement to conventional targeting, not a replacement for them. Layered on to a dysfunctional bureaucracy, AI may accelerate throughput without improving fairness – premature load-bearing in modern dress, the appearance of a capable state without its substance.

The deeper risks are political. If a foreign vendor builds and runs the system that collects a nation’s taxes or decides who receives welfare, who is accountable when it errs, and where does sovereignty reside? History offers a warning: China's Maritime Customs Service was foreign-led for almost a century after 1854. Though exceptionally effective at collecting revenue, its customs receipts were pledged to foreign loans and indemnities, and the institution became an enduring symbol of the unequal-treaty era. Competence could not overcome the perception of compromised sovereignty.

Donor-run project units in modern aid can reproduce a smaller version of that dilemma: imported administrative capability that delivers projects efficiently, but too often bypasses rather than builds domestic institutions. Digital identity systems also err in both directions – excluding the eligible, admitting the ineligible. Delivery can earn trust, but only when citizens can see how decisions are made and contest them when they are wrong. The goal must also be trust, not just capability, because in government, trust is the licence to govern.

Can AI let countries import administrative capacity that historically took centuries to build? Partly, and conditionally. AI can compress the timeline, and for the poorest states that is a real and humane gain. But the IMF finds that digital tools deliver most only where local systems are ready to absorb them. AI is less a shortcut around the state than an accelerant for states already willing to reform.

Capability imported is not legitimacy earned. Conflating the two is the era’s most seductive error. A government can rent the machinery of administration; it cannot rent the consent of the governed, the rule of law or the accountability that makes a capable state trustworthy.

The countries that thrive will be those that use AI to build legitimacy rather than bypass it – deploying imported capability transparently and under domestic control, while the slower work of institution-building continues underneath. Those that mistake a functioning interface for a functioning state may find they have automated only the appearance of governance. Telling the two apart may prove a defining development challenge of the coming decade.