Estonia's Bürokratt: Pioneering an 'Agentic State' for Seamless Government Services with AI

“Should citizens have to know which ministry does what in order to receive government services?”

That was the question that stayed with me as I spoke with Rasmus Eimla at Estonia’s Information System Authority, or RIA. Eimla leads Bürokratt, Estonia’s national AI assistant and digital administrative service platform.

The ultimate vision of Bürokratt is surprisingly simple. Citizens should be able to ask the government questions and receive services through a single point of contact, even if they do not know which ministry or agency is responsible for which task. But the technology needed to implement this simple goal has changed fundamentally over the past few years.

When Bürokratt first began, generative AI as we now know it did not exist. The early system had to anticipate what questions citizens might ask, classify the user’s intent, and then provide prepared answers. The problem was that this required a great deal of manual work. Questions had to be predicted, intents had to be classified and trained, and answers had to be managed. For public agencies, this demanded considerable time and effort, while the utility gained was limited.

Then LLMs arrived.

Bürokratt began to change as well. Institutions can now use their websites and official materials as knowledge bases, and retrieval-augmented generation, or RAG, can help answer citizens’ natural-language questions. What matters is that AI does not use just any knowledge. It is designed to answer based on controlled sources provided and managed by each institution. When it does not know the answer, it should say that it does not know rather than inventing one, and guide the citizen to another way of making an inquiry.

Interestingly, this process produced an unexpected effect. As institutions placed their websites and materials into the AI knowledge base and tested them, they began discovering errors and contradictions in their existing information. The introduction of AI revealed not an AI problem, but a quality problem in existing government information. In the end, AI began to act as a mirror reflecting the government’s information systems.

What is even more interesting about Bürokratt is its structure.

Estonia is not trying to place all government information into one giant AI knowledge base. Each institution operates its own Bürokratt and remains responsible for its own knowledge. This resembles the decentralized philosophy that Estonia’s digital government has maintained for decades. RIA explains that Bürokratt is already being used by several public-sector organizations, and that individual institutional assistants can be connected and developed into a network in which AI agents may cooperate in the future.

For example, a citizen may ask a question to the Bürokratt of Agency A, but the issue may actually fall under Agency B. In that case, the system can identify the appropriate responsibility and connect the citizen to Agency B’s Bürokratt. The citizen does not need to understand the government’s organizational chart. Behind the scenes, however, each institution’s responsibility and expertise remain intact.

In other words, the state appears as one government to the citizen, but the data and responsibility inside government are not centralized into one place.

I see the essence of Estonia’s new vision of the agentic state as a shift from answering to acting. Today’s generative AI mostly answers questions. An AI agent, by contrast, can receive a goal, find the necessary information, communicate with other systems and carry out a sequence of tasks. When applied to government, the implications become much larger.

Until now, citizens have had to know what they need to do, visit government websites, fill out applications and submit the required documents. In an agentic state, this relationship can be reversed. A citizen may state only the goal, and an AI agent can find the relevant government agencies, check the necessary data and carry out the required procedures. Estonia’s Aruait project, known in English as Reason Reserve, is an experiment preparing for precisely this future.

RIA describes the purpose of Aruait as building the legal, technical and governance foundations that will allow the public sector to use autonomous AI agents safely. What is interesting is that Estonia’s current information systems can already exchange data, but they do not yet provide a complete framework for safely delegating tasks to AI agents, granting them authority, enabling trusted communication among agents, and allowing them to perform legally meaningful actions on behalf of citizens.

That is why Aruait addresses a new set of questions.

Who is an AI agent?

Who gave that agent authority?

How far may it act?

Can that authority be revoked at any time?

If it acts incorrectly, who is responsible?

To address these questions, RIA is studying a framework called “Identity 2.0 for Machines,” which would verify the identities of AI agents and allow citizens or businesses to grant them limited, traceable and revocable authority to act on their behalf. It is also developing an AI Agent Trust Registry, a public registration and verification system for public- and private-sector AI agents, as well as interoperability standards that allow different agents to communicate and cooperate securely.

The most difficult problem in the agentic state is not AI itself.

During the interview, I asked Rasmus: “Then when does Estonia become an agentic state? Around 2030?” He smiled and replied that no one is yet brave enough to present such a specific roadmap.

That answer was more impressive than a confident prediction would have been.

Just because something is technically possible does not mean a state can immediately implement it. There are legal issues. There are cybersecurity issues. There is the question of how far citizens’ authority should be delegated to AI. There is the question of responsibility. And above all, there are people.

I also heard that some people are already willing to give their AI agents authentication credentials or even PIN codes. Some people do not trust AI at all. Others trust it too much. In the end, the problem the agentic state must solve is not simply a question of AI performance. It begins with what AI can do, but ultimately moves toward who permits that action, who is responsible for it and how much citizens can trust it.

Estonia’s experience allows us to see the evolution of digital government as a sequence.

A digital government that provides public services online has developed into a proactive government that, based on data connections, can identify and offer the services citizens need without forcing them to search and apply one by one. With the recent combination of LLMs and RAG, it is evolving into an AI-enabled government that understands citizens’ natural-language questions and needs and connects them to appropriate information and services. In the next stage, the agentic state, AI agents authorized by citizens will interact with multiple government systems and carry out real administrative tasks on citizens’ behalf.

But the higher the stage, the more important governance becomes.

When AI merely provides information, the risks are different. When AI acts on behalf of citizens, much stronger systems of identity, authority, accountability and security are required. That is why Estonia’s experiment with the agentic state is so interesting. It is not because Estonia is developing the world’s most powerful LLM. In fact, the opposite is true. Estonia is experimenting with what new governance is needed for AI to act inside the state system, based on the foundation it has built over more than two decades: digital identity, data interoperability, distributed responsibility, cybersecurity and institutional trust.

This is not only Estonia’s problem.

As AI agents become widespread, every country will eventually face the same questions. If AI files taxes on my behalf, applies for permits, books healthcare services and communicates with government, how should the state trust that AI? If the AI makes a wrong decision, is the responsibility on the citizen, the AI developer or the government agency that provided the service? And what is the final domain in which the state should never allow AI to act?

What I saw in Tallinn was that the agentic state is not merely a distant, abstract idea.

At the same time, it is not a future that can be reached simply by developing a better AI model. The more advanced AI becomes, the more the task of the state shifts from securing better technology to deciding what authority that technology should receive and how it should be held accountable.

Estonia’s experiments with Bürokratt and Aruait show that this transition has already begun.