AI x Development: tools, programs or trajectories?

For most of us, AI first shows up as a productivity tool. This is the least contested and most immediately useful way AI is being taken up across the international development community.

In practice, this means using AI to save time drafting donor reports, generating a monitoring, evaluation and learning framework with a reasonable number of indicators and using an AI note taker in a donor coordination meeting so you can actively listen to the tour-de-table of each donor’s grand achievements. These uses slot easily into existing workflows, require little organisational change, and deliver visible gains quickly – critical in the increasingly resource-constrained practice of development cooperation.  

However, we know these efficiency gains are unevenly distributed. OECD data shows that even where AI access is widespread, usage and benefits remain skewed toward higher-income, better-educated groups, reinforcing existing inequalities rather than offsetting them. Who gains time, and who does not, is being shaped by existing inequalities.

There are also risks in equating efficiency with progress. Automating weak processes can entrench poor incentives rather than fix them. Faster reporting does not necessarily mean better programs, and flashier analysis does not guarantee better judgment.

This suggests two clear implications for how we think about AI for efficiency in development practice. First, efficiency is only valuable if it improves development outcomes. AI should not be judged by speed alone, but by whether the time it frees up is reinvested in tackling the effectiveness challenges organisations too often lack the space to address. Second, inclusive adoption matters. If AI primarily saves time for already privileged individuals or organisations, then reducing inequality in access, skills, and institutional permission to use AI should itself become a deliberate objective of international development.

Where efficiency is about how development cooperation operates, effectiveness is about what it achieves. AI enters the chat in two ways: embedded within existing programs, and increasingly as a focus of programs in its own right.

In all development sectors (for example health, agriculture, and public administration) AI is already being incorporated to support service delivery, decision-making, and learning when embedded in capable systems. And it is playing a role in development cooperation processes, like informing program design, supporting implementation, and strengthening monitoring, evaluation, and learning.  

Additionally, AI is increasingly becoming an issue of development cooperation in itself or is being tagged on to programs focused on digital transformation. This includes efforts focused on equitable access to AI capabilities, support for AI governance and regulation, investment in digital public infrastructure, and the development of skills, data, and institutions needed to use AI responsibly. Institutions like the United Nations Development Programme, the Centre for Global Development and the Gates Foundation are steadily picking up this type of work, but for most, treating AI as a sector itself is still on the horizon.  

Consistent findings across public sector research remind us that AI does not fix broken systems. What it can do, however, is make capable systems more effective when introduced with discipline. A report by the International Growth Centre highlights a small set of principles important for low- and middle-income countries. These include prioritising low-risk, internal use cases, investing in data quality and skills before scaling, and measuring value rigorously rather than assuming impact. Evidence from Oxford Insights reinforces this point, showing that countries that perform well are not those with the most advanced pilots, but those with stronger policy capacity, governance arrangements, procurement systems, and institutional coordination.

AI use also changes the politics of locally led development because it makes the traditional case for highly paid international advisors harder to justify. If a national ministry can now draft a policy brief with three well-placed prompts, the economics of flying someone in from Washington or Canberra start to look shaky – and their daily rates even shakier.

But there’s a catch. Most frontier models are trained on overwhelmingly non-local data, which means they will carry cultural, linguistic, and political biases, and can subtly re-import foreign assumptions back into national decision-making. That risk, or at least perceived risk, is driving a growing ‘sovereign AI’ push, where countries and regions build or adapt their own models, trained on local languages, norms, and institutions, so the productivity gains of AI don’t come bundled with a new dependency loop.  

The task for development actors is therefore to support faster, responsible adoption. The sooner countries and communities can experiment, adapt, and build local capability with safeguards in place, the less likely they are to fall further behind. But there is a genuine tension here that is worth naming. Faster adoption means more reliance on the foreign-trained frontier models that carry the very dependency risks described above. Slower adoption means falling further behind. There is no clean resolution – but a tension worth making explicit for countries to navigate on their own terms.

While the previous two lenses focus on how AI affects development cooperation as a set of capabilities and interventions, this third lens looks beyond tools and programs to ask how AI is shaping development trajectories themselves. As AI becomes embedded across economies, institutions, and societies, it will increasingly determine patterns of poverty and prosperity, the distribution of economic opportunity, the nature of governance, and the dynamics of conflict and stability. Whether it narrows or widens those gaps depends on choices being made right now. As UNDP's The Next Great Divergence argues, the stakes are generational.

The starting point is a distribution problem. AI capabilities are highly concentrated in a small number of countries and firms that dominate frontier models, compute, and talent. Economic gains will likely accrue disproportionately to economies already at the technological frontier, and even Anthropic has warned that differentiated AI use risks deepening inequality. The numbers are stark: Singapore's AI readiness score is over 80 out of 100, while Timor-Leste scores more than 50 points lower. In the Pacific, mobile broadband covers 86% of the population, yet only 27% actually use mobile internet. IMF analysis finds around 50% of jobs in advanced Asian economies are exposed to AI, compared to just 25% in emerging economies, meaning richer nations will integrate AI faster and productivity gaps are likely to widen.  

Concentration also creates risks beyond inequality and has implications for conflict and stability. Just one example is analysis by Forethought, which warns that sufficiently advanced AI could enable a small group, or even a single person, to seize power – including in established democracies – precisely because frontier capabilities are already in very few hands.

Whether AI deepens these divides or helps close them comes down to an interdependent set of factors: digital infrastructure, institutional readiness, skills, governance frameworks, and political systems. Democratic systems may be better positioned to build the public trust and accountability mechanisms that responsible AI requires but face slower regulatory cycles. Authoritarian states can move faster on adoption and infrastructure but are more likely to deploy AI for surveillance and control rather than human development.  

In Southeast Asia, these dynamics are already playing out: Indonesia's 2024 election saw political actors deploy AI for covert micro-targeting and image manipulation, including a deepfake of former President Suharto endorsing a party. Indonesia, the Philippines, and Singapore have each since moved to regulate AI in elections differently. For lower-capacity states more broadly, UNDP evidence shows AI is already interacting with existing inequalities in state capacity to produce path-dependent outcomes, narrowing the window to scale promising innovations before the gap widens further.

The honest answer on what the development community controls is: not much. Most forces shaping AI's global trajectory sit in technology markets, geopolitics, and politics well beyond the reach of development cooperation. But development actors are not irrelevant. They can shape the debates and agendas where AI policy is formed, insert development values into those conversations, and help countries build the skills, institutions, and resources to make their own strategic choices. Which I believe could add up to a lot.