| Governments cannot turn AI adoption into better performance through technology alone. Their ability to absorb AI depends on four institutional capacities: strategic decision-making, implementation, clear rules, and enforcement. |
I have met with government officials across Latin America, and one question comes up in almost every conversation: how do we use AI? It’s understandable. We are all wondering about the same thing. But this is not the best place to start. The more important question is: what is your government actually capable of absorbing?
Where should governments start with AI?
Generative AI is not just another digital tool. It appears to have the characteristics of a general-purpose technology. It can be used across many sectors, it improves over time, and it can generate new waves of innovation. That means it could fundamentally change the way we work in the public sector, just as earlier general-purpose technologies changed how societies and organizations operated.
That is exactly why governments need to think strategically about how they respond to such a structural transformation. Put differently, what institutional foundation does a government need to turn AI’s potential into reality?
Think of it like constructing a building. The higher you want to build, the stronger foundations you need to support the work. The same is true with AI. Governments will only be able to build as far as their institutional foundation allows.
In government, that foundation is the rules, processes, capabilities and people that turn a decision into action.
Adopting AI is not the same as absorbing it.
Adoption means announcing, acquiring, testing, or piloting an AI tool. Probably every government has done some version of this by now.
Absorption means incorporating AI into the way an organization works on a daily basis. That implies changing its processes, responsibilities and workflows, which will also change the way decisions are made, policies implemented and services delivered.
I haven’t seen that in the public sector at scale yet. Neither has most of the private sector. The World Economic Forum found that 84 percent of companies have not redesigned jobs around what AI can do, and 75 percent say AI is not having a transformative effect on their business, because it has simply been layered on top of the way they already worked.
Microsoft’s 2026 Work Trend Index calls this the Transformation Paradox. People are moving faster than their organizations. This is a systems problem, and to fix systems, you need to redesign them. According to Microsoft’s report, organizational factors have more than twice the impact on effective AI use as individual ones.
Adopting technology is not enough
This is not a hypothetical risk. This example from Latin America and the Caribbean is telling.
According to the World Bank, by 2022 every country in the region had public financial management and tax management information systems. Ninety-one percent had systems for managing human resources, and 84 percent had electronic procurement systems. This was remarkable progress, built gradually over two decades.
Many of these systems were designed primarily for administration and compliance, not for decision-making. The same report finds that 96 percent of the data produced by government management information systems in the region is used for descriptive analysis. Only about half is also used for diagnostic or predictive purposes.
In other words, a government can have platforms, databases and fancy dashboards without having the routines, incentives, analytical capability and decision rights required to act on what the data shows.
I like to think of it as having a Ferrari dashboard attached to an engine that cannot move.
The same could happen with generative AI. Perhaps faster, and with more money spent, because the tools are more impressive and the pressure to announce something is higher. The biggest risk is that governments adopt the appearance of AI capacity without the institutional capacity to use it.
Where does your system stop?
My work at the IDB has helped me think systematically about what “state capacity” means in practice. Looking at the literature and experience of Latin America and the Caribbean, but also lessons from around the world, we identified four key capacities an effective state needs. I believe these capacities can also help governments diagnose their readiness for AI.
For me, an effective state is one that is able to implement policies, deliver services, and create an enabling environment for private sector development, under clear rules of the game that apply impartially to all.
The four capacities are connected. Governments will have different weaknesses and depending on the particular problem, one could be the one limiting AI’s potential.
1. Capacity for strategic decision-making
AI adoption, like digital transformation more broadly, cuts across government. It affects procurement, HR, budgeting and many other areas. It cannot be reduced to scattered agency-level experiments.
Governments need to decide where AI matters most, which problems should receive attention first, and what needs to change around the technology to leverage its potential. Doing this requires a whole-of-government approach, led from the center and informed by executing agencies. The center’s role is to set priorities, allocate scarce resources, sequence investments, establish common standards, resolve coordination failures, and create the conditions for agencies to adopt AI.
2. Capacity for implementation
Technology alone will not change anything. The decision to use AI has to be backed by operational systems that support it. That is data, infrastructure, procurement, processes, and, most importantly, people. The binding constraint will not be the same everywhere, and each element has its own challenges.
- If data is incomplete, fragmented or inconsistent, it’s not fit for decision-making.
- Digital systems in general, and generative AI in particular, pose procurement challenges. On average, IT projects have experienced substantial costs overruns, an average of 73%, showing the complexity of this type of projects. This only magnifies with AI.
- AI adoption will also require more than a small group of specialists. It will require managers who understand how to redesign work, public servants who know when and how to use these tools, technical teams who can support implementation, legal teams who can interpret risks, and leaders who can make decisions under uncertainty. Having people with the right skills in the right positions is what will make the difference.
- Experience from the private sector suggests that serious AI adoption is not just about giving people access to tools. It involves rethinking processes.
3. Capacity to set clear rules
Governments have a double role; they are the regulator but also users. In this post, I focus on the latter.
AI adoption inside government will require clear rules for responsible use, particularly considering the impact public decisions have on society. People are already using free AI tools in many organizations, often without clear guidance. That creates risks around sensitive information, quality control, accountability, and consistency. It also limits the ability to scale the positive uses of AI because everything remains informal and fragmented.
4. Capacity to enforce those rules
Rules are only as strong as the state’s ability to enforce them. Inside government that starts with a simple question: does anyone know how AI is being used? In most organizations, the honest answer is no. Few organizations track which tools people use or check outputs before they are final. Tools are changing faster than guidance documents.
Enforcement here does not mean policing employees. Research from Microsoft and the World Economic Forum shows that as AI takes on task execution, the risk of bad outputs rises sharply. Therefore, organizations need to build automated auditability, permission boundaries, and quality checks directly into their workflows.
That is why enforcement capacity has to be built alongside adoption, not added later. By the time you notice a problem, the practice is already everywhere.
Leapfrogging is not skipping
It’s easy to feel behind with AI. Being behind can also be a chance to build differently, faster, and smarter. But leapfrogging is not skipping. Governments still need to understand problems, map processes, redesign workflows, and strengthen operations. AI may help build some foundations faster. But it does not remove the need for foundations and for adapting the systems you already have in order to enable AI transformation.
I believe AI can help governments make better decisions, implement policies more effectively, and improve the services people receive. That is precisely why we need to look beyond the excitement of a pilot or the launch of a new tool.
The models will continue to improve, whether governments are ready or not. But governments will not rise automatically to the level of the latest AI model. They will only be able to build as far as their institutions allow.
The place to start is not with the tool.
Choose a problem that matters to your organization. Map the process to understand what would have to change for AI to improve the outcome. Then, find the first point at which the institution cannot make that change. That is where your AI constraint is, not in the model, in the foundation.
