What does a city's Chief Digital Officer actually do now that AI has arrived?

In Munich, city staff draft emails, summarise long documents and brainstorm with MUCGPT, an assistant the city built itself. It went live in March 2024; more than 15,000 of the city’s roughly 40,000 staff now use it regularly, and it handles over 10,000 chats a day, the city says.

Berlin’s Senate Chancellery switched on BärGPT on 25 November 2025. CityLAB Berlin built it as open source, testing it with administrative staff, and it can search Parla, the archive of the city parliament’s written questions. Its makers count over 17,000 users since launch, and the tool is still in trial operation. Hamburg developed LLMoin with the public IT provider Dataport from mid-2023 and, after a staged rollout, made it available to about 40,000 staff, an administration of about Munich’s size; we found no published count of how many use it. Amsterdam’s ChatAmsterdam started as a pilot with 800 officials. Each city counts something different, access, users or regular users, so the figures cannot be ranked against each other.

Seen in order, the wave took less than two years, from Munich’s launch to Hamburg’s whole administration.

Hamburg’s CDO now describes the assistant as office equipment.

Behind the numbers sit different choices. Munich and Berlin built on open-source code and publish it; Hamburg built with its public IT provider; in Sweden, Sundsvall and Ånge built Eneo, an open platform anyone can run. Amsterdam built its own assistant but hosts it on Microsoft’s Azure cloud, where staff choose between OpenAI’s GPT 5.5 and Mistral Large 3. Alexander Scholtes, then the alderman for ICT, said in May 2025 that by developing and managing its own AI tools, the city keeps control over its data. An early study of 150 pilot users judged roughly one prompt in fourteen risky, because it held personal data or a statement about a decision, so the city added a filter of its own on top of Microsoft’s.

Other cities bought or wrote rules. Barcelona announced in March 2025 that some 500 staff would get Microsoft’s Copilot that year, and that the council’s AI infrastructure would move to the Microsoft cloud by mid-year; we could not confirm how the rollout went. Vienna chose rules over a tool: its staff compass allows public generative AI tools for work, but only with public data; restricted and confidential material may go only into the city’s internal platform, cleared for now after a data-protection assessment.

From our library

Kompass für den dienstlichen Umgang mit generativer Künstlicher Intelligenz (KI-Kompass, Version 3.0)

Vienna’s staff guide details public‑genAI usage, DPIA‑cleared confidential data handling, and 2026 updates.

Read more →

We found no independent evaluation, published by any of these cities, of what its assistant changed: time saved, errors made, or what residents noticed. Hamburg’s best evidence is an October 2025 poll in which over 80% of the users asked were satisfied or positive about it; the city calls the sample representative but has not published its size.

The most thorough public test of a city chatbot we found is American. New York launched its MyCity chatbot in September 2023 to answer business owners’ questions about city rules. In 2024 The Markup and THE CITY found it telling employers they could take a cut of workers’ tips and that businesses could refuse cash, both against city law.

In December 2025 the city’s Comptroller, Brad Lander, published an audit of MyCity. His auditors asked the bot the same question twice and got two different answers. More than 2,200 questions were put to it in July and August 2025; of the 70 users who left feedback, 50, or 71.4%, were negative. Officials at the city’s Office of Technology and Innovation replied that, as a “generative AI”, the information produced “may not always be consistent”.

From our library

Audit Report on the New York City Office of Technology and Innovation's MyCity System (IT25PAR20028)

Audit reveals MyCity’s $100 M rollout faltered: no unified form, 46% childcare apps ineligible, chatbot missed half queries.

Read more →

The office also said the bot was over 95% accurate week to week. The auditors found it had counted prompts recorded rather than questions asked.

And the office’s rate left out answers it had filed under “Room for Improvement”. In January 2026, while looking for budget savings, Mayor Zohran Mamdani called the bot “functionally unusable”, and the city took it down on 4 February.

Europe’s own lesson came from a city that followed the guidebook. Amsterdam’s Smart Check, a model to flag welfare applications for investigation, had bias tests and a human-rights assessment. In its 2023 pilot, nearly 1,600 applications went through it. According to an investigation by MIT Technology Review, Lighthouse Reports and Trouw, bias reappeared and the model was no better than caseworkers at spotting the cases that deserved scrutiny. In late November 2023 the city stopped it.

From our library

Inside Amsterdam’s high-stakes experiment to create fair welfare AI

Amsterdam’s AI welfare pilot flagged excess cases, revealed hidden bias, and was abruptly halted.

Read more →

The quietest frontier is the note-taker. In social care in England and Scotland, AI tools transcribe and summarise conversations with the people social workers support; by early 2025 one such tool was in use in 85 local authorities. Swindon’s own pilot, run with the tool’s maker, Beam, with 19 social workers over 184 meetings, reported a 63% cut in the time spent on assessments and case notes. The Ada Lovelace Institute held 39 interviews in 17 local authorities for Scribe and prejudice?. Almost all the social workers said the tools helped, and one council found its records became fuller. But councils judged the tools mainly on efficiency: only two had extended their pilots to check the quality of the records. One of those found the AI often did not raise quality and may have let unchecked hallucinations into official care records. One social worker described a summary that recorded suicidal ideation the client had never mentioned.

Who answers when it is wrong? The individual employee. Amsterdam’s register entry for its assistant says the user “remains responsible for the final content” and any next steps, and the Ada Lovelace Institute found that social workers assume full responsibility for what the transcription tools produce.

So the CDO now owns a new question: how do we know it works? Of the assistants and chatbots in this read, only New York’s has been through an independent audit, and the auditor found it unreliable.

The Netherlands turned the city register into a national one. When the Algoritmeregister opened in December 2022 it held 109 algorithms; by mid-February 2024 it held 291, with Amsterdam, Utrecht and The Hague contributing the most. It has grown in every month the register’s dashboard shows.

By October 2026 the register held 1,573 descriptions from 353 publishing organisations; municipalities supplied 987 of them. Listing is still voluntary. Of 342 municipalities, 208 are listed as publishing, 86 have joined without publishing anything and 48 have not joined at all; the median publishing municipality lists three algorithms. Amsterdam lists more than any other municipality. Elsewhere, Barcelona’s public register of algorithms was, in December 2024, planned but not yet fully implemented, according to a CIDOB case study.

How much a register can miss is clearest one level up. When the Netherlands Court of Audit asked 70 central-government bodies in 2024 what AI they used, they reported 433 systems. Only 5% were in the public register, and for 35% of those in use the organisation did not know whether they did what was expected. That is national government, not cities; we found no comparable count for a city.

Contracts travelled further. The clauses Amsterdam wrote in 2018 for buying algorithmic systems became the basis of the EU’s model contractual clauses for public buyers of AI, published in September 2023 and updated in March 2025 in a full version for high-risk systems and a light one for the rest. Not every buyer asks the question. Among English councils answering the LGA’s 2025 survey, 27% require suppliers to declare whether they use AI, and 25% did not know their own policy.

From our library

State of the Sector: AI – Update (Research report, 2025)

Survey of 104 English councils reveals 33% response, highlighting adoption gaps, supplier reliance, and skill‑funding challenges.

Read more →

Then there is the law. The EU AI Act’s literacy duty has applied since 2 February 2025. The AI Omnibus rewrote it: organisations that provide or use AI must now take measures to support the AI literacy of their staff, and the text says this does not require them to guarantee any specific level. AI that public authorities use to evaluate people’s eligibility for essential public benefits and services is classed as high-risk, and before using such a system a public body must assess its impact on fundamental rights and register it in an EU database. Those duties were due in August 2026. The timeline shows where they stand now.

Until December 2027, then, there is no record of how cities meet the high-risk duties.

A register entry is not an audit. In October 2023, Amsterdam’s own audit office found that the city had “properly examined whether an algorithm is fair” for only one algorithm. Of the five high-risk systems whose documents went into its report, Smart Check was the only one that had also been through a human-rights assessment. A month later, after its own pilot showed bias, the city stopped it.

From our library

Algoritmen – Hoe Amsterdam algoritmen beter kan toepassen (Rekenkamer Amsterdam, onderzoeksrapport, 20 oktober 2023)

Audit reveals only one bias‑checked algorithm among 29 registered, exposing a hidden accountability gap.

Read more →

Every Swedish municipality that answered a survey in autumn 2025 said it used AI in some way. Only 7% said all their staff had received AI training, less than a year after the EU literacy duty began. The duty does not require every employee to be trained, but it does cover everyone who works with an AI system, and in every responding municipality someone did. The survey was run by Akavia, a trade union for graduates; 170 of the 290 municipalities replied. Fewer than one in four had an AI strategy, and a quarter of the smaller ones had offered no AI training at all.

From our library

AI i kommunens tjänst – Så använder Sveriges kommuner AI

Survey of 170 Swedish municipalities reveals universal AI use, few strategies, and minimal staff training.

Read more →

England’s figures come from a different group: heads of IT or senior responsible officers. In February and March 2024, 74 councils answered the Local Government Association. For its 2025 update, 58% of them revised their answers and 30 more councils joined over the winter of 2024–25; the other original answers, about 31, were carried over unchanged, giving 104 responses, a third of councils. Between the two rounds, use rose while the barriers stayed where they were.

Funding still tops the list, but the next two barriers are about people.

Germany’s Zukunftsradar asked people in leadership roles, such as mayors and heads of department, and says itself that its results are not representative: 77% thought AI and automated systems sensible in local government, yet 13% used any, and 71% named data protection as a main obstacle. A vendor-backed global panel of 250 cities, run by ThoughtLab with ServiceNow, Deloitte and NVIDIA, found 18% using generative AI widely or selectively in 2024, and shows the AI job sometimes leaving the CDO’s desk: six in ten of its “AI leaders” had appointed a senior head of AI, with a chief AI officer or AI director in New York and Singapore; in Barcelona it names the chief data officer, in Amsterdam the chief technology officer. These surveys asked different people different questions; their numbers cannot be added together.

Some cities are changing who they employ. New York’s new administration announced it would hire public-sector technologists to build its own tools and rely less on outside vendors.

Hamburg, Berlin and Amsterdam paired their assistants with teaching. Hamburg ran training courses on both the tool and basic AI skills alongside LLMoin’s rollout, BärGPT comes with a help centre of video tutorials meant to build staff’s confidence with AI, and Amsterdam says it runs courses and e-learning alongside ChatAmsterdam. We found no published account of what the training changed.

A budget can buy an assistant. On its own it does not supply what England’s next two barriers name: people in the town hall who can judge what an AI system is doing.

The pattern we found best documented in Europe is a city that builds once and lets others reuse it. Hamburg offers LLMoin to other administrations through Dataport, and Lower Saxony, Bremen, Saxony-Anhalt and Rhineland-Palatinate have decided to use it. Stuttgart runs Munich’s MUCGPT under the name ChatSTR, and Munich publishes the code and invites other administrations to deploy it. In the Dutch register, the province of Limburg and the municipality of Meierijstad both list GovChat-NL, an open-source platform for Dutch public bodies that each runs under its own name.

Sweden shows why the terms of sharing matter. Sundsvall had made the case more than a year before its split: Sweden’s 290 municipalities have essentially the same mission and needs, it wrote in May 2024, yet each finds its own solutions, which leaves digital welfare unequal; generative AI was a blank sheet on which they could work together. Sundsvall and Ånge had built an AI platform called Intric with private partners. In September 2025 they left it, because a single company controlled the code and kept parts of it closed, and carried on as Eneo under an open licence that any public body or supplier can use and build on.

Follow the tools rather than the cities, and the pattern is easier to see.

In each case the second user started from working code or tested wording, not from a blank page. Bigger bets on shared language models are still plans. GPT-NL, a Dutch model built by TNO, SURF and the Netherlands Forensic Institute on a €13.5m starting budget, was being tested in early 2026 by five launching customers, from ministries to financial institutions; the report names no municipality. Copenhagen’s finance administration said in May 2024 that it planned to join the Danish Language Model Consortium, in which the municipal association KL and Aarhus already take part. At EU level, the LDT CitiVERSE EDIC, a consortium for local digital twins, held its first general assembly in València in December 2024.

That changes what a CDO is for. The useful question is less “what should we build?” than “who has already built it, and can we share the work of keeping it safe?”

Three questions follow from this read, and every city’s answers will differ.

1. Which tool another city already runs could we reuse, instead of building our own?
2. What do we publish about the AI we use, and who checks that it works as described?
3. How are our staff being trained now, and will that be enough when the high-risk duties arrive in December 2027?

The Smart Cities Hub is built to make those questions quicker to answer. It aims to map who in Europe’s city halls is building, buying or sharing AI, so that a city choosing a tool can find one that already runs it, and it collects evidence, including reports in German, Dutch and Swedish that rarely travel.

Some of the projects and organisations behind this read are already under Data Platforms, AI & Digital Twins; the assistants named here are not all on the map yet. If your city runs an assistant, keeps a register or has written a procurement template, add it to the Hub so the next city choosing one can find yours.