How can LMIC governments govern with AI?

We left the conversation, held under the Chatham House Rule, with the following five takeaways, informed by a series of surveys of delegates throughout the day-long conference:

Start with the problem, not the technology

We started the conference by asking participants - drawn heavily from policy and economics backgrounds, including several senior government officials – when they believed most citizens would primarily interact with the government through an AI system. A few chose 2028. At 2030, the balance began to shift, with the majority believing that they would interact via AI. By 2035, almost everyone expected AI to mediate much of the relationship between citizens and the state.

But agreement disappeared when we asked what AI should actually do. Around two-thirds did not think AI would ever fully replace doctors. Yet almost everyone thought AI could handle basic diagnostics and complex treatment plans by 2035. Only around a third thought it could perform complex surgical procedures, and around a quarter thought it could replace counselling or interpersonal care.

The distinction is interesting. We found the question “should governments adopt AI?” too broad, and that both delegates and speakers approached sectors with great levels of heterogeneity (doctors not only treat patients but also provide human empathy).

Instead, governments need to break public services into tasks and ask where AI improves the outcome, where it should assist a human, and where automation would make the service worse (or less trustworthy). That will mean measuring success against public problems informed by consultation with local communities, not just technical benchmarks. Ask: does the system reduce processing time, improve targeting outcomes for people, or help a nurse or tax official make a better decision? Do people trust the system? One useful approach is to break public services into a bundle of tasks – the individual decisions and actions that make up the delivery of a service – and then identify where AI can improve them.

The hard part is often everything around the technology

Instead of discussing the technology itself, we spent most of the time discussing data, institutions, and people needed to make technology work.

Almost everyone agreed that for AI adoption to succeed, we need more work on improving data quality, interoperable systems, rules for sharing information, and procurement capacity. These are the harder and less glamorous parts of AI adoption, where far less money and focus are on right now.

We also heard concerns that current investments are too often concentrated on pilots and non-government social enterprises, rather than on the capabilities governments need to responsibly adopt AI at scale. One view favoured starting with specific, discrete applications: find important problems, test whether AI can solve them, and scale what works. Social enterprises are playing (and can play more so) an important role here by demonstrating new approaches to solving such policy problems.

But these demos will not automatically translate into government-wide adoption. One government participant made the case that while funding social enterprises can prove what is possible, it can also crowd out investment in the public institutions responsible for taking these solutions to scale. The same problem also applies to public-sector pilots: governments can accumulate a long list of (often donor-supported) promising experiments without ever being able to scale them, as the latter requires broader state capacity that stands apart from philanthropy.

One particular form of capacity governments need is the ability to procure technology effectively without hollowing out the state. Reliance on external consultants, contractors, and vendors can be useful (and, frankly, is needed), but if it becomes the default, governments risk never building the internal capability to assess technologies and adapt their systems over time. This also creates risks of procurement and vendor lock-in, something many governments experienced in earlier waves of technology adoption.

The alternative is to find a new balance: concrete applications (both by governments and social enterprises) to demonstrate value, while investing in the common data and governance infrastructure that makes application and procurement easier to scale. The UK's Incubator for AI, who presented their model at the conference, is a useful case study on building project-agnostic state capacity.

Governments want sovereignty, but sovereignty may have a price

The strongest survey results from the room concerned sovereignty.¹ Almost everyone preferred greater government capacity to exercise agency and control over the AI systems used, even if this meant accepting a less capable model.

The concern is reasonable. Governments do not want critical public systems permanently tied to one company, one foreign jurisdiction, or one technical architecture. They want control over sensitive data and the ability to adapt, retraining systems for local languages and changing suppliers whenever appropriate.

But this raises an important trade-off that we did not delve into at this conference: what if sovereignty and capability pull in opposite directions? If we suppose that the best proprietary models will continue to perform much better than an open weight model that a government can host or modify, how much capability should the government sacrifice for greater control?

There are ways the government can preserve some options even under proprietary models. They can demand interoperability, avoid unnecessary vendor lock-in, and maintain the technical capacity to switch providers. But open-weight models provide far more local control, including over where models are hosted, how they are adapted, and how data is handled. These all (again) require greater state capacity, and we need more discussions to articulate this sovereignty-capability trade-off.

One interesting area for immediate action can be helping LMIC governments coordinate regional dialogue and investments that improve their bargaining power and build shared infrastructure. We learned about some ambitions to this end, and it can be a valuable area for philanthropists to support, to inform and empower regional consortia to negotiate with companies.

Human rights and trust are part of the infrastructure

Participants were slightly more worried about governments moving too quickly without adequate safeguards than about moving too slowly to capture productivity gains from AI adoption.

This concern was strongest in areas such as health, where decisions can create large-scale harm that cannot easily be reversed, especially in life-or-death situations. Many argued for considering human rights and agency throughout the process of building and deploying AI, instead of being an afterthought. For this to happen, governments need process-based tools (such as due diligence, impact assessment, and risk management frameworks) to identify and address human rights harms throughout the AI lifecycle, from data collection to system design, deployment, and maintenance. Law and regulatory institutions play a critical role to this end by aligning incentives, creating accountability for technology providers and users, and preventing “ethics-washing”.

Any assessment of AI’s human rights impacts in government must also account for local political economy and institutional incentives. Several speakers argued that practitioners need to keep human rights at the centre of decision-making even as governments face tighter fiscal space, pressure to deliver services more efficiently, limited capacity for due diligence and regulation, and negotiations with private vendors whose incentives may not align with the public interest.

Related to this, there were a range of questions: Who is missing from the training data? Who bears the cost when the system makes a mistake? Can a citizen challenge an automated decision? When must a human intervene? Who is accountable when something goes wrong? Governments should also measure social performance, not simply model performance. It is not unreasonable to expect that a system can become faster while making a public service harder to access, less legitimate, or less trusted.

Governments need to prepare for AI beyond today’s use cases

The final lesson for us came partly from what we did not discuss. We spent substantial time on governance, privacy, procurement, data systems, and useful applications. We didn’t spend much time on the scenarios that occupy the frontier AI community: highly autonomous agents, large-scale cyber risks, and rapid labour-market disruption.

We understand why. Governments have immediate problems to solve, and AI offers great potential to that end. But it creates a risk that we build governance for the AI of 2026 while failing to prepare for the AI of 2030. When we asked whether AI would ultimately make government institutions stronger or weaker, the room was roughly divided.

This uncertainty calls for two kinds of capacity in the governments of LMICs. First, they need the capacity to improve government outcomes today. Second, they need the capacity to understand how more powerful AI can reshape firms, public finances, service delivery, and the state itself. If, by 2035, most citizens navigate government through AI, we will need to ask more fundamental questions about what kind of society and economy we want to build with these capabilities – and what kind of governance it requires.

This blog reflects the authors’ personal reflections on those discussions and does not represent the views of their institutions, or those of the host institutions. The conversations were held under the Chatham House Rule, hence no speaker is directly attributed.

1. Sovereignty is an increasingly common theme in the policy debate on AI, but is often used to refer to many different objectives and problems. In this blog, we use “sovereignty” to mean a government’s capacity to make independent decisions around AI development and use.