Stops motion. Your device’s reduced-motion setting always takes priority.
Municipal experiences
Local experience. Shared learning.
From pilots to day-to-day use: 9 municipal experiences with AI. Explore what is being tested, what is in operation and what others can learn.
VejleIn operationSources reviewed BerthaAI support on the intranet for finding guidance, proofreading and writing clear public-facing text.Read caseClose
Bertha: AI support where employees already work
The challenge
How can generative AI become an everyday tool beyond the group of technology enthusiasts?
The approach
Developed with InDialog, Bertha is available on the municipal intranet. Staff use it to find guidance, proofread and make public-facing text easier to understand. Intranet content informs its answers and needs ongoing maintenance. The Danish Agency for Digital Government also describes training in prompting and critical assessment, alongside legal involvement, risk assessments and a data-processing agreement during implementation.
What can you take from this?
A shared assistant also needs content ownership and staff skills. Consider who corrects outdated guidance and how errors in answers are detected and followed up.
AI Alliancen’s editorial reflections — not measured outcomes or statements from the municipality.
Before adapting the idea
Start with a bounded knowledge base, assign content owners and agree how staff receive training and report errors.
A question to take awayWho is responsible for keeping the assistant’s information accurate six months from now?
KalundborgIn operationSources reviewed Project AndreaMunicipal employees’ generative AI tools brought together in one app.Read caseClose
Project Andrea: one entry point for municipal AI tools
The challenge
How can multiple AI tools be made accessible without creating separate entry points?
The approach
Project Andrea brings generative AI tools together in one app for municipal employees. The initiative originated in Kalundborg Municipality and was developed with Promte. The solution is in operation in Kalundborg. KL's earlier case also describes risk assessments for assistants.
What can you take from this?
Access and organisation are connected: clarify who reviews new assistants and how staff find the relevant tools.
AI Alliancen’s editorial reflections — not measured outcomes or statements from the municipality.
Before adapting the idea
Identify who reviews each assistant, maintains the catalogue and keeps access and data boundaries visible.
A question to take awayHow do you maintain oversight, ownership and risk assessments as the number of tools grows?
AarhusIn operationSources reviewed AarhusAI and OS2aiGeneral and specialist assistants on a shared platform, made available to other public bodies through OS2ai.Read caseClose
AarhusAI and OS2ai: from a local platform to shared capability
The challenge
How can staff get AI assistants while other public authorities can reuse the underlying platform?
The approach
AarhusAI brings general and specialist assistants together for employees. The platform is shared through OS2 as OS2ai. Its documented capabilities include document analysis, organisational sources through RAG and shared assistants. It combines open-source and source-available components; model selection, hosting and operation require local decisions. Aarhus ITK has also described AI training and network events as part of its work on staff capability.
What can you take from this?
Shared software does not remove local operational or professional responsibility. Consider which components can be shared and which sources, assessments and workflows each authority must maintain.
AI Alliancen’s editorial reflections — not measured outcomes or statements from the municipality.
Before adapting the idea
Agree operational ownership, source access, answer testing and costs. Connect training to real tasks rather than simply providing a new tool.
A question to take awayWhat should you own and maintain locally, and what is better sustained together?
CopenhagenTrialSources reviewed TaltTesting speech recognition and language models for home-visit descriptions using synthetic citizens and records.Read caseClose
Talt: from spoken notes to draft home-visit descriptions
The challenge
Can staff document a home visit in everyday language and get help structuring the result?
The approach
Copenhagen, Aarhus, Aalborg and Systematic developed prototypes tested in 2024–2025 with staff from 11 municipalities. One use case combined speech recognition and a language model: dictation became text and then a draft visit description for staff to review and edit. The trials used synthetic citizens and records. The project also publishes code, datasets and technical reports.
What can you take from this?
Test the full workflow, not just the finished note. Errors can arise in audio, transcription and summarisation; a convincing demonstration is not a production-ready integration.
AI Alliancen’s editorial reflections — not measured outcomes or statements from the municipality.
Before adapting the idea
Agree professional quality criteria, review procedures and integration needs before transferring synthetic-data findings into a real workflow.
A question to take awayHow will staff notice when an important detail has been changed or omitted?
GladsaxeIn operationSources reviewed RoboRefConversations with citizens become draft minutes that staff read, edit and approve.Read caseClose
RoboRef: more attention during conversations with citizens
The challenge
How can social workers and care assessors document meetings without spending the conversation taking notes?
The approach
Gladsaxe describes RoboRef being used in its Family Department and Social and Health Administration. Speech-to-text and generative AI turn the meeting into draft minutes, which staff read, edit and approve. The municipality's information on AI minutes also describes informing citizens and states that individual case records are not used to train models.
What can you take from this?
A clear review process is part of the solution. Examine whether drafts capture what matters and whether reduced writing time outweighs the work of checking them.
AI Alliancen’s editorial reflections — not measured outcomes or statements from the municipality.
Before adapting the idea
Agree citizen information, recording handling, approval responsibilities and error follow-up with the professional teams using the tool.
A question to take awayHow will you assess both the quality of the record and attention during the conversation?
GentofteIn operationSources reviewed Specialist AI assistantsInternal guidance on occupational injuries, environmental issues and IT support.Read caseClose
Specialist assistants for municipal questions
The challenge
How can specialised internal guidance become easier to find and use across professional teams?
The approach
Gentofte describes internal Promte assistants covering occupational injuries, insurance, environmental issues, contaminated land and IT support. The supplier's user-group account describes an occupational-injury assistant from Gentofte and its feedback on a KLE municipal filing-classification assistant. These are bounded professional uses within a shared platform, rather than one chatbot for every task.
What can you take from this?
A narrow purpose can make answers easier to test and ownership clearer. Consider who maintains the guidance and when a question must be passed to a specialist.
AI Alliancen’s editorial reflections — not measured outcomes or statements from the municipality.
Before adapting the idea
Assign a professional owner, gather authoritative texts and write realistic test questions, including questions the assistant should not answer on its own.
A question to take awayWhich bounded area of guidance would make the most useful first assistant?
OdenseIn operationSources reviewed SkoleGPTSkoleGPT for municipal school pupils and teaching staff, supported by guidance and staff workshops.Read caseClose
SkoleGPT: shared access and critical understanding of AI
The challenge
How can pupils explore AI with professional guidance rather than simply using whichever services they find?
The approach
In October 2025, Odense opened SkoleGPT to all pupils and teaching staff in its municipal schools. Schools received guidance on AI and digital literacy, while staff were offered subject-focused workshops. Pupils can examine answers, work on tasks and explore creative expression. The platform's documentation describes a language model, an input filter and separate interface and model servers.
What can you take from this?
Providing access and developing critical understanding are different tasks. Consider how teachers make errors, source criticism and limitations part of learning, and how experience is shared between schools.
AI Alliancen’s editorial reflections — not measured outcomes or statements from the municipality.
Before adapting the idea
Agree the teacher's role, age-appropriate tasks, input boundaries and a straightforward process for following up unsuitable answers.
A question to take awayHow can pupils learn to challenge an AI answer rather than simply accept it?
SvendborgBetaSources reviewed SpinnaTeachers’ choices of subject, topic and teaching approach become a proposed lesson sequence.Read caseClose
Spinna: lesson planning developed with teachers
The challenge
How can AI support preparation without taking over a teacher's educational judgement?
The approach
Tåsingeskolen is developing Spinna with Svendborg Municipality. The project reports that 30 teachers helped test and shape it. Teachers select a subject, topic and teaching approach; these choices form a prompt that generates a proposed lesson sequence. The published privacy policy describes both use without an account and Unilogin access with saved lesson plans.
What can you take from this?
User involvement can go beyond trying a finished prototype. Examine how teachers' choices and criticism change the product, and which parts still need their own professional work.
AI Alliancen’s editorial reflections — not measured outcomes or statements from the municipality.
Before adapting the idea
Test across subjects and pupil groups, agree permissible inputs and assess the subject-specific and educational quality of plans before using them.
A question to take awayWhich choices should AI suggest, and which must remain the teacher's own?
ViborgIn operationSources reviewed AI in energy managementENTO analyses building energy consumption and recommends areas for action, which technicians investigate and follow up.Read caseClose
AI-based energy management: from consumption data to action
The challenge
How can technicians identify buildings where a practical intervention could reduce unnecessary energy use?
The approach
Viborg uses ENTO to analyse consumption data automatically and recommend areas for action in municipal buildings. Data comes from utility companies. Energy technicians investigate anomalies and follow up with local staff, for example on faulty valves or ventilation running unnecessarily. The municipality's climate evaluation describes further meters being connected to the system.
What can you take from this?
A recommendation only creates value when someone can act on it. Examine how alerts are prioritised, who investigates them and how the effect of the resulting change is followed up.
AI Alliancen’s editorial reflections — not measured outcomes or statements from the municipality.
Before adapting the idea
Clarify data quality, meter coverage and follow-up ownership. Consider building use and indoor conditions before judging a consumption change to be an improvement.
A question to take awayIs there a clear path from a data anomaly to a completed operational task?
Status reflects the cited sources or the dated confirmation stated for each case, not ongoing verification of operations. Learning points are editorial reflections, not measured outcomes.