Domain Atlas / Benefits navigation & public-facing chat
Burokratt
Burokratt is Estonia's national network of public-sector chatbots operated by the Information System Authority: each participating institution runs its own assistant, a central classifier routes a citizen's query between them and oversees the handover, and from 2025 a shared knowledge module built from the eesti.ee state portal feeds cross-domain answers. RIA's page lists 20 participating organisations and trade press reports 18 integrated; an independent 2025 ethnography drawing on twelve insider interviews (conducted in late 2023, when the system spanned ten institutions) found it marketed as advanced AI while functioning much like an FAQ list, with use differing considerably by institution and low in some. No published session volumes, escalation-to-human rates, or answer-accuracy figures, and no dedicated algorithmic-oversight body, published evaluation framework, or national-audit report on the network, were located in the public record.[3]
What happened
Burokratt is Estonia's national network of public-sector chatbots, operated by the Information System Authority (RIA) under the Ministry of Justice and Digital Affairs. It grew out of the country's 2019 national AI (the "kratt") strategy: a vision and concept paper in 2020, an alpha tested in 2021 in three agencies (the Police and Border Guard Board, the Consumer Protection and Technical Regulatory Authority, and the National Library), and a first full rollout in 2022 beginning with the Consumer Protection agency. The design is deliberately federated. Each participating institution runs its own virtual assistant with its own knowledge base, and a central component described as a "global classifier" routes a citizen's query to the right institutional bot and oversees the handover between assistants, so one chat window fronts a background message exchange among many institutional nodes. The country's chief data officer has framed the network as interoperable in the manner of Estonia's X-Road data-exchange layer and has offered the open-source stack for free reuse, with Belgium and Finland named as negotiation partners (Luxembourg is softer, conference-stage interest) and cross-border interoperability with Finland's AuroraAI explored. The software is free and open source under MIT licensing — the public code organisation holds 89 repositories with active development through mid-2026 — and institutions pay roughly EUR 150 per month for State Cloud hosting plus usage fees.
How big the network is depends on how you count. RIA's own page lists 20 participating organisations, including the Tax and Customs Board, the Police and Border Guard Board, Statistics Estonia, the Health Insurance Fund, the state portal eesti.ee, the National Library, and the municipality of Rae Parish; some of those entries are portals or sites rather than distinct institutions, which helps explain why trade-press coverage quoting officials reports 18 organisations integrated. The independent ethnographic study by Kaun and Manniste, drawing on twelve insider interviews conducted from September to November 2023, records the system as first piloted in 2021 in two institutions and implemented near the end of 2023 in eight additional institutions including one municipality — ten institutions in total at that time. The programme's technology has moved through generations under an unchanged governance network. The first generation (2021 to 2024) was rule- and intent-based, built on the open-source RASA natural-language framework, deployed in the national State Cloud with the state authentication service; the European Commission's Public Sector Tech Watch entry, last updated in June 2025, still lists the project as "in development" and records RIA's own statement that establishing organisational partnerships is difficult because the expertise needed to jointly develop this kind of AI solution is very high. A 2025 milestone programme layers retrieval-augmented generation over institution-specific databases plus a shared general knowledge module built by ingesting public information from the eesti.ee portal — covering healthcare, social benefits, pensions, transport, employment, education, and taxation — so the central assistant can answer cross-domain questions from a single window. A further step, planned from 2026, would give each institution its own AI agent in a cooperative agent network on an Estonian-adapted large language model; this is a stated plan, not a deployed capability, and is treated here as prospective.
Two registers describe the same system, and they do not agree. The promotional register — official pages, an Enterprise Estonia briefing, and trade press — calls Burokratt a "Siri of public services," reports a UNESCO/IRCAI top-100 listing, and frames it as rewriting how people interact with the state. The independent academic register is more sober. Kaun and Manniste found the system marketed as advanced AI while functionally resembling an FAQ list, found that use "differs considerably" by institution — minimal in one, reaching about one-third of daily requests in another — and documented procurement-driven vendor churn (major components were built by Net Group, Texta, Stacc, Solita, and Microsoft), value conflicts with the universalism of the welfare state, and a striking under-use of data: chat interactions are converted into analysable data, but that datafication potential went largely unexploited beyond basic usage statistics. (A concrete phone-versus-email volume comparison that is sometimes read into Burokratt actually belongs to a separate Swedish municipal chatbot studied in the same paper, and is not a Burokratt figure.) The survey-vignette experiment by Alishani and Homburg, published online in December 2025, parameterises the citizen side: intended use is driven by perceived usefulness and trust in the technology, privacy concerns matter for service-provision uses but not for information-provision uses, and trust in government, explainability, and the amount of information provided were unrelated to intended use.
What is missing from the record is as important as what is in it. No published session volumes, escalation-to-human rates, or answer-accuracy figures were located in any public source; no dedicated algorithmic-oversight body, published evaluation framework, or national-audit report on Burokratt was found; and no harm, error, or discrimination incident is on record. The absence of documented incidents here reflects the absence of published evaluation, not evidence of absence. The funding record needs the same care: the European Commission's Recovery and Resilience Facility project record lists a EUR 53 million EU contribution but describes it as being "for the two projects... altogether," bundling Burokratt with core digital-infrastructure and cloud-transition components, so it should not be read as tool-only spend; the nearer tool-scale figures are the roughly EUR 1.5 million spent through early 2022 and the roughly EUR 13 million budgeted over the following four years. This case is therefore a topology and governance exemplar rather than a calibration case — its value is the clearest documented real instance of a federated, indexed multi-entity structure, not a measured performance record.
The sociotechnical reading
Most navigation tools in this atlas are a single node: one assistant, one knowledge store, one caseworker or citizen at the other end, and the whole safety question is whether the human verifies before relying. Burokratt is the atlas's clearest example of the other geometry — a federation, a chatbot of chatbots. Read on the system map, it is many institutional error-source nodes, each with its own memory store, coupled through routing links and, from 2025, one shared knowledge module that every node reads. That structure changes where the risk lives, and the distinct lesson of this case is a structural one: in a federation the failure surface is the shared link, not the node. A single stale or wrong entry in the module every assistant reads is not one institution's mistake; it is retrieved into every institution's answers. A classifier that routes a query to the wrong bot is not a local slip; it is a network-wide misdirection. Per-institution stores do the opposite work — they quarantine an institution's own errors to its own answers — so the map's blunt reading is that error propagation is a property of the link structure, and the two shared links (the classifier and the shared module) are a monoculture whose failures correlate across the whole network while the private stores localise the rest.
That reframes what governing this system means. It is not, first, "is each bot accurate?" It is three link-structure questions the record answers only partly. First, validate the shared links: check the shared module and the classifier's routing against the institutions' own authoritative content, because that is the single write that reaches everyone — on the map it is the central team's ingestion of portal content into the shared module, the most consequential edge on the diagram. Second, reconcile the copies: when the same public fact lives in a shared module and in an institution's own store, drift between them is exactly where a citizen gets two different answers depending on which bot the classifier picked. Third, stand up a federation-level evaluation, because the most concrete finding in the independent record is an absence — Estonia built the interoperability, the classifier, and the shared module faster than it built anything to watch them, and the academics are currently the main external check. There is a quieter fourth reading the map keeps visible: adoption itself is the binding variable in places. Where a bot is barely used, the human take-over that is meant to catch it gets little practice and the chat logs that could retrain it go unexploited, so the low-traffic nodes are both the least valuable and the least watched. The honest boundary is that none of this measures the citizens on the other side of the window — the system makes no eligibility or benefit determination, and what a shared-link failure moves here is the quality of the state's own answers, not any person's outcome, in a case where those answers have, so far, gone almost entirely unmeasured.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.