Beyond AI Pilots: How Governments Can Build the Capacity to Scale

By Sandra Naranjo Bautista

An AI tool can help identify a procurement bottleneck, but whether someone will act on that finding is a different story.

I keep thinking about that distinction after attending Google’s Public Sector Summit. I went interested in what AI can do for government, and with a question that keeps coming up in my work: how do we build the institutional capacity to put that potential to use?

The event reinforced an argument I’ve made before. Governments’ ability to benefit from AI will be limited by their institutional capacity. Technological advances alone won’t overcome those limits. That capacity has to grow while government keeps delivering on its mission. Government can’t stop operations to prepare for AI. It has to deliver while it transforms.

Here are five reflections from the event.

1. AI’s potential is enormous, but it requires more than the technology.

One demonstration I particularly liked focused on public procurement. It showed how AI agents could help officials detect bottlenecks and analyze supplier relationships.

I could see the possibilities. At the same time, I was thinking about what would need to happen for an official to act on that analysis. Could they validate the findings? Would they have the authority to address the bottleneck? Would another department need to agree?

Putting AI to work requires three conditions to come together:

• Technical foundations: usable data, infrastructure, and connections to existing systems.
• Operational capacity: staff, skills, resources, procurement arrangements, and routines for using and checking the tool.
• Authority and support: permission to change how work gets done, cooperation across departments, and incentives to act.

A tool can work well while the decision it supports stays stuck.

The summit was an invitation to build and to have the boldness to try. I agree; unless we put these capabilities to use, nothing will happen. But implementation also means creating the conditions that let people act on what the technology makes possible. That is institutional capacity.  

2. What does trust mean in government AI?

Trust was a recurring theme among government officials at the summit. I hear it from our counterparts across Latin America and the Caribbean too. Several different concerns sit behind that one word.

The first is protecting information: what you can share, where it will be stored, and how confidentiality will be maintained.

The second is confidence in the results. Let me give you an example of what I mean. I had to process some data I had collected. Uploading the database to an AI tool was the easiest option. But I had two concerns. The first was confidentiality, which required appropriate safeguards. The second was whether I could verify how the tool reached the answer and validate the results. I wasn’t worried the result would be wrong, but I didn’t feel comfortable sharing something I couldn’t defend.

In the public sector, that matters. Your decisions can have legal, financial, or political consequences. As we delegate more work to AI agents, we need to decide who checks that work, how, and when someone intervenes. You delegate the task, but not your responsibility. You are still accountable.

A third concern is citizens’ trust in decisions made with or by AI. That raises questions about transparency, fairness, legitimacy and the chance to challenge a decision. It deserves a separate discussion.

3. Start with the problem and be willing to change the process.

There is always the temptation (and honestly, the pressure) to show that you are doing something with AI. But introducing a tool does not necessarily improve a service.

That is why you have to start with a problem that matters to the organization. Think about pain points you have now. Where is work getting stuck? What is taking too long? Which decisions could improve if officials had better information?

Then follow that problem through the process. If AI makes one task faster, what happens next?

A working paper covering more than 500,000 developers shows this difference. It estimated cumulative gains across successive generations of AI coding tools of 240% in coding activity measured by commits, but 30% in software releases. The authors interpret the gap as consistent with human bottlenecks elsewhere in the production chain.

The implications for government are worth considering. Faster document processing can send more cases to an already overloaded review team. The constraint moves without the service improving.

Choose a manageable first step and judge it by what changes across the whole process. A faster task matters when it helps produce a better result. To get the benefit, you have to be willing to let go of the “usual way of doing things”.

4. Leadership and grassroots adoption go hand in hand.

The cases presented at the summit highlighted the importance of leadership. Leaders set direction, choose priorities, provide resources, and authorize implementation. But leaders also need the knowledge of the people doing the work.

Public servants know the day-to-day hurdles of an organization: the exceptions, the informal routines, and why a seemingly simple change can be so difficult. That expertise is essential to deciding where AI can help.

It requires space to experiment and feedback loops that carry what people learn into real workflows. Microsoft’s 2026 Work Trend Index highlights the importance of organizational conditions. Its survey analysis found that culture, manager support, and talent practices were more strongly associated with reported AI benefits than individual mindset and behavior (These are self-reported benefits, not demonstrated causal effects).

A team may be encouraged to experiment while carrying the same workload. A manager may be accountable for results but unable to authorize the changes required to achieve them. One department may carry the costs of implementation while another receives most of the benefits.

Under those conditions, slow adoption may be a predictable response. Some of these problems can be solved inside a team. Others need agency leadership or political authorization. We can’t expect enthusiasm to overcome constraints people do not control.

5. Why do AI pilots stall? Think about what happens next.

A pilot may work because it has a dedicated team, temporary funding, or close support from a vendor. Before expanding it, we need to understand which of those conditions can continue. Who will own the solution when the pilot ends? Who will maintain and finance it? Who will evaluate whether it works?

Ask these questions at the start, while there is time to answer them.

I would also ask what the organization will learn that it can use again. Will staff be better able to evaluate another tool? Will there be a clearer process for validating results? Will the next initiative face fewer obstacles because of what this one resolved?

A successful pilot shows that something can work. Preparing to scale means establishing how it will keep working under ordinary operating conditions.

My preferred starting point is still to think big, but choose a manageable first step. Have a vision for transforming the process without trying to do everything at once.

Build capacity by turning learning into changes in how government works

Governments learn by doing. They build the capacity to solve problems by solving them. Pilots can contribute to that process, but experimentation alone is not enough.

A pilot can strengthen a team’s skills. For that learning to change how the organization operates, it must inform decisions about resources, responsibilities, and routines. That requires connecting what teams discover with people who have the authority to act.

The answer is not to wait until government has perfect data, systems, and skills. It won’t happen. We need to build those capabilities while putting AI to work, starting with improvements that are feasible under current conditions.

Each initiative should solve a real problem today and leave the organization better prepared for the next one. That is the opportunity I see: deliberately using AI implementation to build institutional capacity.

Moving beyond pilots means making useful innovations part of how government operates, while leaving the organization more capable with each implementation.