Name in original language
AI-apuri
Initiative overview
AI-apuri was introduced to address a practical limitation of generic large language models: they generate answers based on their general training data and cannot reliably rely on curated, task-specific information. Many work processes at the Authority require up-to-date internal guidance, legal interpretations and structured external sources. The initiative leverages retrieval-augmented generation (RAG) architecture, where chatbots retrieve answers from defined internal and external data sources. This allows more accurate, transparent and context-specific responses while maintaining human oversight. The objective is to improve efficiency, consistency and access to knowledge in daily work.
Civil servants can build chatbots tailored to their team’s needs, configuring them with curated datasets and, if required, instructing them to answer only on the basis of those sources. Current use cases are internal: the chatbot supports the civil servant, for example by retrieving information or drafting a reply, but the final communication is always reviewed and sent by a human (human in the loop). The platform architecture allows restricted and sensitive personal data to be processed securely, as inference residency is controlled and training data capture is not permitted.
A distinctive feature of the initiative is its “citizen development” model. Instead of centralised development only, a network of civil servants designs and maintains tailored chatbots for their own teams with support from the ICT unit. This approach has enabled the scaling of use cases across the organisation, and currently 12 chatbots are in production.
In the future, the initiative is expected to expand both in scope and maturity. The current informal developer community will be developed into a more structured virtual group to improve prioritisation, shared learning and quality. While current use cases are internal, selected chatbots may later be made available externally, provided that their datasets rely solely on appropriate external sources and that their use complies with Finnish legislation.
Results, outcomes and impacts
The initiative has led to 12 chatbots in production and over 60 explored use cases, tracked through a structured funnel from idea to deployment. Usage is monitored per chatbot. While some tools handle relatively few questions, a single interaction can save significant time: drafting a reply may save about 30 minutes per case, and translating and formalising a consumer complaint up to 15 minutes. Impact is measured through usage data, funnel progression and case-level time saving estimates. Although individual use cases are limited in scope, the low threshold for creating new agents and the Authority’s case volumes create substantial cumulative potential for efficiency, consistency and improved access to information.
