Estonia is building AI agents into the state — on purpose, and slowly

While most companies are still bolting a chatbot onto a help page, Estonia has spent the better part of a decade doing something stranger and more ambitious: wiring AI agents into the machinery of the state itself. The project is called Bürokratt, and it is less a chatbot than a blueprint for how a government might actually talk to its citizens in the age of AI.

Not a chatbot — a network

Bürokratt is best understood not as a single assistant but as what its designers call an interoperable network of public- and private-sector AI solutions which, from the user’s point of view, acts as a single channel for public services. Estonia describes it in consumer-friendly shorthand as a “Siri of public services”: a citizen uses voice or text to get what they need without navigating a dozen separate agency websites. Behind that single front door sits a growing ecosystem of assistants run by different agencies, stitched together so they can hand a request from one to another.

The name is telling. A kratt, in Estonian folklore, is a creature assembled from odds and ends and brought to life to do its master’s bidding — and one that turns dangerous if left idle without a task. Estonia chose the word deliberately, as a metaphor for both the advantages and the risks the technology carries. The reusable, open-source pieces of the system are even called kratijupid — “kratt parts”.

What it actually does

The ambition is concrete. Through Bürokratt, citizens are meant to apply for benefits, make payments, register a birth, file a tax return and renew a licence — by speaking or typing, on whatever device is to hand. The assistant is also designed to be proactive: to contact people and remind them when something needs doing, rather than waiting to be asked.

That reflects a particular philosophy of government. “People will have services offered to them and delivered in an integrated manner, around their needs,” says Ott Velsberg, Estonia’s chief data officer. The country’s former government CIO, Siim Sikkut, framed the goal as services “that are rightfully tailored to the individual”. The bureaucracy, in other words, is supposed to come to you.

Built the slow, governed way

What makes Bürokratt worth studying is less the demo than the discipline behind it. The system is open source, released under an MIT licence with its code on GitLab and access tied to Estonia’s national eID. It is assembled from shared, reusable components — speech synthesis, machine translation, speech recognition — that any agency, or private company, can pick up. It is overseen by the Ministry of Economic Affairs and Communications and built with the state’s Information System Authority and outside partners, and links to private services such as banks are designed to happen only with the citizen’s consent.

And it has been built gradually. The concept was presented alongside Estonia’s national AI strategy in 2019; pilots followed in 2020 and 2021, starting with a handful of agencies including the Police and Border Guard Board and the National Library; the first citizen-facing services arrived in 2022. Today roughly 18 government organisations are connected. The 2025 roadmap adds large language models and retrieval-augmented generation, a shared knowledge base and a speech-based assistant — work backed by tens of millions of euros in EU recovery funding — with 2026 set aside for personalised AI agents and an Estonian-adapted language model.

Agents that cross borders

The next frontier is interoperability between countries. Estonia is working so that virtual assistants deployed in different states can, in its words, “seamlessly connect and collaborate with each other”. The example officials like to give: a pregnant Estonian woman living in Finland could use Bürokratt to arrange hospital transport and manage her related needs, with Estonia’s system talking to Finland’s AuroraAI platform on her behalf. “The focus must shift from simply answering inquiries to redesigning these assistants to navigate bureaucracy,” says Velsberg.

Why it matters

Bürokratt is interesting precisely because it is not a rushed bolt-on. Estonia treats an AI agent the way it treats any other piece of public infrastructure: scoped to specific services, gated behind verified identity, connected to other data only with consent, and open to scrutiny because the code is in the open. The result is an agent people are more likely to trust — not despite the governance, but because of it.

That is the same principle that should guide AI agents inside a business. An agent becomes genuinely useful when it is bounded to a clear role, allowed to act only on data it is permitted to see, and auditable after the fact. Estonia is proving the model at the scale of a country; the lesson travels down to the scale of a single company.