AI needs to focus on the systems that drive development

Artificial intelligence has arrived in the global development world. Some present it as a ladder that low-income countries can use to climb into prosperity; others see another imported technology destined to fail in dusty clinics and on messy data systems. 

But the question needs to be reframed: the issue is not whether AI “works”; it is how we can focus it on the right questions.

Is it enough for development policy to identify effects?

Development policy, at times, muddles two things: effects and drivers. An effect is something that changes when pushed. A driver is something that, once pushed, changes many things at once. Effects move outcomes at the margin. Drivers reshape the terrain on which outcomes are produced.

Development strategy hinges on this difference. A fertiliser subsidy may increase yields this year, while a land registry that secures property rights would change investment patterns for decades. A tutoring programme may lift some test scores, but a reform that improves teacher recruitment would alter the quality of an entire generation of schooling. Both work – only one transforms.

We have become very good at identifying effects. We can now show with precision that countless variables affect income, health, and learning. But development does not suffer from a shortage of evidence that something matters. 

It suffers from a shortage of forces that matter enough. An intervention is not valuable because it produces a result in isolation. It is valuable because it explains a meaningful share of what goes wrong or right in the real world.

The difference is subtle but important. A policy can be causally correct and practically inconsequential. Some effects are real, but may be tiny or not scalable. Many operate only when conditions are controlled. But development does not occur in controlled conditions; it occurs in fragile systems, under noisy constraints and dynamic political incentives. What developing countries need are not just clever nudges, but also structural shifts.

Using AI to target the drivers of development

This is where artificial intelligence plays a role. Given enough data, AI can detect and reproduce patterns at scale, generating results faster than institutions can absorb them. But pattern is not priority. A model can highlight behaviour without judging its consequences, optimise predictions without understanding the purpose, and flag a change without knowing whether the change matters.

The task, then, is not to ask whether AI can help shape development outcomes, because it already can. The goal is to make sure it targets drivers rather than merely producing effects. If AI is applied to whatever is easiest to measure rather than what is hardest to shift, it will accelerate activity without building capacity. 

What now follows is a bet on where AI is most likely to change trajectories, not just totals; distributions, not just averages; institutions, not just indicators.

How can AI transform information systems at scale?

Many systems in the world fail because of a lack of information. Consider the smallholder farmer (often some of the poorest people in the world): millions of them make costly decisions daily – when to plant, what to grow, whether to sell – with little guidance. 

Weather shifts, pests spread, and prices wobble, but extension services are thin, insurance is rare, and market information is unreliable. Similarly, in healthcare, patients also require accuracy. In many countries, people die not just because of a shortage of clinics and hospitals, but because care is inconsistent, diagnoses are delayed, and drugs are misprescribed. 

These failures have a common cause: information arrives late,  or not at all, and it is often incomplete. Forecasts, guidance, and alerts powered by AI can change this asymmetry by reaching millions almost for free.  When information is timely and accurate, it converts uncertainty into income and error into prevention.

Information can also be an excellent multiplier by raising the return on every other resource. Economies do not advance only by acquiring more capital, but by organising knowledge more effectively. 

A clinic with the same doctors becomes more competent. A farmer with the same land becomes more productive. When the fog lifts, productivity follows.  This is what makes information a driver, not merely an effect. It does not just shift outcomes at the margin, but compounds advantage over time. Development, in this view, is less a story of accumulation than of coordination.

But even the most precise information systems are useless without reach . Information that cannot travel does not exist – a forecast unread is no forecast at all.  AI must be able to speak local languages, run on low-cost devices, and overcome bad connectivity. In development, where the real innovation is reliability, not sophistication, AI doesn’t need to be brilliant; it needs to be consistent at scale.

Can AI strengthen state capacity where it matters most?

The binding constraint in much of the developing world is policy implementation. Governments struggle to collect data, enforce rules, pay staff, track supplies, and deliver benefits. While policies exist on paper, they dissolve on contact with reality. When states are weak, everything else is fragile. This is where AI’s quietest and perhaps deepest promise lies: as an infrastructure for governments.

In many countries, the machinery of the state remains paper-bound, disconnected, and slow. AI can automate workflows, flag irregularities, and support routine decisions. These gains are small individually, but large in aggregate. A ministry that closes cases faster wastes less. A department that audits in real time deters theft.

Governments in developing countries often operate with fractured information: health in one system, land in another, taxes somewhere else. Governance thrives on integration – and so does AI. 

Linking registries, cleaning databases, and reconciling records are not glamorous reforms, but they are foundational ones. A government that knows who its citizens are, where assets lie, and where money flows governs differently from one that does not.

Guardrails matter a lot here. AI deployed without standards – without data protection, oversight, or appeals – can magnify dysfunction. Sovereignty matters too – governments that ignore where models run and who controls them risk exporting information and importing dependency. 

The answer is not digital autarky, but a strategy that involves tough procurement, investment in local capability, and standards that bind vendors to public purpose.

If AI strengthens state capacity, it reinforces everything else. Development takes place not when tools arrive, but when institutions absorb them.

Using AI to improve measurement and feedback in public systems 

Policy works when learning is quick, but it rarely is in reality. Evaluations in developing countries are expensive and slow. Results arrive years late, by which point the world has changed. AI can change this by enabling surveys by phone, real-time analysis, and failure to be flagged early. The importance of this does not lie in producing more reports, but in changing how systems behave. 

Take modern machines. They did not succeed because they avoided failure. They succeed because failure is visible – through aircraft black boxes, cars reporting faults before engines fail, and  software logging its crashes. 

Public policies, by contrast, operate largely without diagnostics, and as a result, malfunctions are discovered long after damage is done – if at all.^[1]

AI offers governments something that engineering made routine decades ago: internal telemetry. It has the potential to help school systems track attendance and learning daily, not annually; allow health systems to notice medicine shortages when they begin, not when clinics close; and detect exclusions in welfare systems as they happen, not when hunger follows.

This is not about automation. It is about correction. A state that can observe itself adjusts faster. Information does not simply improve outcomes; it improves the speed at which outcomes improve. In development, that is decisive – a country that learns twice as fast does not just grow twice as quickly; it grows differently.

Three rules to guide AI for development in the new age

If AI is to become a driver, not just a producer of effects, three principles should apply:

1. Be unforgiving about scale: If a system cannot reach millions at low cost, it does not belong at the centre of policy.
2. Treat safety as infrastructure: Privacy, accountability, and transparency are not afterthoughts – they are engineering requirements. Systems without them decay.
3. Reward influence, not intrigue: Novelty is cheap, but transformation is rare. Fund what shifts distributions, not what spruces presentations.

The most significant risk is not that AI will fail when applied to development. It is that it will succeed just enough to produce dashboards, pilots, and applause, while leaving systems unchanged – resulting in motion without momentum; accuracy without achievement. 

Learn more about data and AI

[1] The diagnostic-system analogy is based on a lecture by Asim Khwaja with Rohini Pande.