European cities are turning artificial intelligence from a laboratory experiment into working public infrastructure in 2026, with municipalities including Espoo, The Hague, Riga, Oulu, Amsterdam, Manchester, Leipzig, Bordeaux, Ghent and Rotterdam testing AI in administration, mobility, education, policing support and civic services. The important part is not that city hall has discovered chatbots. It is that Europe’s local governments are beginning to treat AI as a governance problem before they treat it as a productivity feature. That choice may decide whether the next wave of urban software strengthens public service or quietly outsources democratic judgment to systems nobody can properly inspect.
The Eurocities Digital Forum in Sofia, held from April 15 to 17, arrived at a useful moment. The AI hype cycle has moved past the stage where every institution needed a “strategy” slide and into the less glamorous stage where procurement rules, staff training, data quality, audit trails and accountability matter more than demos. For cities, that shift is unavoidable. Unlike consumer AI apps, municipal AI does not live in a sandbox; it touches housing queues, traffic fines, public employment, classrooms, social services and the digital tools residents must use to get through daily life.
The most revealing examples from European cities are not the flashiest ones. Espoo’s intranet AI assistant, designed to help roughly 12,000 city employees find internal information more quickly, sounds modest beside visions of autonomous urban operating systems. But that modesty is the point. A civil servant who can find the right policy, form, procedure or case note faster may deliver a better service without turning a resident into a test subject.
Espoo is combining GPT-based tools and Microsoft Copilot-style workflows with a familiar public-sector problem: internal knowledge is scattered, outdated, duplicated and hidden behind organisational boundaries. AI does not magically fix that. If anything, it exposes the mess. A model trained or connected to inconsistent internal material will reproduce the confusion faster and with more confidence.
That is why Espoo’s lesson matters for every municipality now eyeing Copilot, Azure OpenAI, Google Gemini, local large language models or sovereign cloud alternatives. The first question is not “Which model should we buy?” It is “What information do we actually have, who owns it, how current is it, and who is allowed to use it?” In public administration, the AI project often begins as an IT deployment and quickly becomes an information-governance audit.
For Windows-heavy municipal environments, this is not theoretical. Microsoft’s ecosystem is already embedded in the daily work of many European city administrations: Teams meetings, SharePoint sites, Outlook inboxes, Entra ID identities, Intune-managed devices and Office documents stretching back years. AI assistants that sit inside that productivity stack can be powerful precisely because they are close to the work. They are also risky for the same reason.
The convenience layer can become a leakage layer. A poorly configured AI assistant may surface documents that employees technically had access to but never should have been able to discover at scale. A summarisation tool may blur the line between official policy and draft material. A chatbot may confidently answer a staff question using obsolete guidance. In a city administration, these are not mere annoyances; they can become unequal service delivery.
Recruitment is a classic trap for public-sector AI. It looks administratively attractive because there are many applications, repeated processes and measurable outcomes. But the very data used to train or tune these systems often reflects past discrimination, institutional preference and unequal access to opportunity. If a city government deploys AI to rank candidates, it risks encoding yesterday’s workforce into tomorrow’s hiring pipeline.
The Hague’s approach also suggests that an AI strategy is useful only if it enables refusal. Too many institutional AI policies read like permission structures: here is how we will innovate, here is how we will pilot, here is how we will modernise. A serious public AI strategy must also say no, not later after harm appears, but early enough to prevent the harm from becoming bureaucratic routine.
That is especially important under Europe’s AI Act, which is moving from legal text into operational reality. The law entered into force in 2024, with key provisions on prohibited practices and AI literacy already applicable from February 2025 and broader obligations arriving in stages. For cities, this means AI compliance is not a future legal department project. It is now part of procurement, HR, education, policing support, citizen services and vendor management.
The AI Act’s risk-based logic aligns with what competent local governments already know: not all deployments are equal. An internal search assistant is not the same as an automated welfare-risk score. A traffic-flow model is not the same as a facial recognition dragnet. A translation tool for public notices is not the same as a system making eligibility decisions. The work of city government is to keep those distinctions visible when vendors would rather describe everything as digital transformation.
That sounds reassuring until one asks what “final authority” actually means. Does the inspector meaningfully review evidence or merely approve a pre-filled recommendation? Can the inspector see why the system flagged a case? Is there a reliable way to detect systematic errors? Can residents challenge not only the fine but also the technical pipeline that produced it?
These questions matter because automation bias is not a theoretical flaw. People tend to trust machine output when it arrives packaged as objective, consistent and official. In a busy public office, a recommendation system can become the decision-maker in practice even when the formal policy says otherwise. The interface becomes the institution.
For cities, the standard should be stronger than “a person clicked approve.” Human oversight should mean trained staff, documented review, error sampling, appeal rights, clear public explanations and the ability to suspend a system when it behaves badly. It should also mean that the human reviewer has enough time and institutional authority to disagree with the machine.
This is where local government can either improve on the private sector or repeat its mistakes. The private AI economy often treats user friction as a failure. Public administration should sometimes treat friction as a safeguard. A pause, a second review, a written explanation or a refusal to automate a sensitive step can be the difference between efficiency and arbitrary power.
Riga’s work on mobility data, sensors and digital twins points to the practical need. Urban planning depends on information from transport systems, energy networks, climate models, road works, emergency services and citizen-facing platforms. If these systems cannot talk to one another, AI becomes a decorative layer over disconnected databases. If they can talk without rules, cities risk building surveillance infrastructure by accident.
The idea behind European data spaces is not simply to pool everything into one giant database. The stronger model is federated: data owners retain control, access is governed, standards allow interoperability, and use cases are built around defined public purposes. That approach fits the European instinct to make digital infrastructure accountable rather than merely scalable.
Oulu’s emphasis on trust and distributed control is significant here. Trust is often invoked as if it were a communications challenge: publish a policy, hold a workshop, reassure the public. But in digital government, trust is an engineering and governance property. It depends on who can access data, how access is logged, whether systems are auditable, how long data is retained, and whether residents have meaningful rights.
For IT professionals, this is the real architecture conversation. AI models will change rapidly; today’s frontier system becomes tomorrow’s commodity feature. Data architecture, identity management, security classification, retention policies and interoperability standards are more durable. Cities that neglect those foundations will find themselves repeatedly buying new AI tools to compensate for old data failures.
Local governments are not hyperscale cloud companies. They do not have unlimited engineering teams, spare data centres or the luxury of rebuilding every platform from scratch. They depend on vendors because vendors provide working systems, security updates, integration support and economies of scale. The question is not whether cities can avoid dependency altogether. They cannot. The question is whether they understand and govern the dependencies they accept.
This is where the Microsoft angle is unavoidable for WindowsForum readers. Many municipalities already sit deep inside the Microsoft stack, and Microsoft is aggressively positioning Copilot, Azure AI, Fabric, Purview and security tooling as a coherent enterprise AI platform. That integration is attractive because it can reduce deployment complexity. It is also strategically sticky.
A city that builds AI workflows around one vendor’s identity model, document graph, compliance tooling and assistant interface may gain speed but lose bargaining power. Exporting documents is easy; exporting institutional habits is not. Once employees learn to ask an assistant embedded in their office suite for policy answers, meeting summaries and case preparation, the assistant becomes part of the administrative fabric.
European cities are therefore right to ask hard questions before convenience becomes dependency. Sovereignty does not require rejecting commercial platforms. It does require exit plans, open standards where possible, contractual audit rights, clear data-processing terms, portability requirements and internal expertise strong enough to challenge vendor claims. A city that cannot understand the system it uses cannot credibly govern it.
That gap matters because unprepared institutions do not necessarily avoid AI. They often adopt it badly. Staff use public chatbots with sensitive material. Departments buy overlapping tools. Managers approve pilots without evaluation plans. Procurement teams accept vendor language they cannot test. Legal and security teams are brought in after the system is already politically committed.
AI literacy obligations under the European framework make this more than a best-practice issue. Public bodies deploying AI need staff who understand capabilities, limitations, risks and appropriate use. But the deeper point is cultural. A city cannot govern AI if only the IT department understands it, and it cannot use AI well if IT understands the technology but not the service context.
Bordeaux Metropole’s distinction between digital skills and digital culture is one of the more important ideas in the Eurocities discussion. Skills are knowing how to use a tool. Culture is knowing what the tool does to institutions, relationships and choices. A clerk can learn to prompt a chatbot in an afternoon; understanding when not to use one takes experience, ethics and organisational confidence.
For residents, the same distinction applies. Digital inclusion is not just broadband access, a smartphone and a password reset service. It is the ability to understand when an automated system is involved, what rights exist, how to challenge an outcome and how to participate in decisions about technology. A city that digitises services without building civic digital culture may improve efficiency while widening the gap between those who can navigate systems and those who are governed by them.
The educational challenge is not simply whether pupils should use AI. They already will. The question is whether schools can teach students to understand AI as a tool, a source of error, a labour-saving device, a persuasion machine and a social force. That is a much harder curriculum than “write better prompts.”
Amsterdam’s work on digital citizenship and online harms adds another layer. Cities are seeing offline consequences from online life: mental health pressure, cyberbullying, conflicts that begin in digital spaces and spill into schools, neighbourhoods and public services. AI will intensify that problem by making synthetic content cheaper, impersonation easier and harassment more scalable.
Manchester’s framing of digital citizenship around inclusion, infrastructure and representation is important because AI policy can become elite policy very quickly. The people most affected by automated systems are often least represented in their design. Residents without reliable connectivity, language confidence, institutional trust or spare time are unlikely to attend consultation workshops about algorithmic governance. Yet they may be the first to feel the consequences.
Leipzig’s support for NGOs, clubs and associations recognises that civil society is part of the digital public-service ecosystem. Local organisations often mediate between residents and government, especially for vulnerable groups. If those organisations lack digital capacity, then AI-enabled public services may become harder to access precisely for the people who need support most.
For city governments, this raises the stakes. A chatbot that answers an employee’s question can be wrong. An agent that updates a case record, sends a message, schedules an inspection or triggers a procurement step can be wrong and consequential. The transition from advice to action is where governance must become sharper.
Most municipal processes were not designed for autonomous software actors. They rely on tacit knowledge, informal escalation, legacy systems, handwritten exceptions and human judgment about when a case is unusual. AI agents promise to automate the connective tissue between systems, but that connective tissue is often where accountability lives.
This is also where Windows and enterprise endpoint management become relevant again. If agents operate through browsers, office suites, line-of-business applications and cloud APIs, then identity, permissions and device security become central controls. A compromised account with an AI agent attached may be more dangerous than a compromised account alone. The agent can move faster, summarise more, attempt more actions and potentially obscure the operator’s intent.
Security teams will need to think about AI not only as data risk but as operational risk. Which agents can act on behalf of which users? What logs are retained? Can actions be replayed and explained? Are agents blocked from sensitive systems by default? Can a city disable an AI workflow during an incident? These questions belong in the same room as procurement and policy, not in a post-breach review.
A resident trying to access housing support may not need an AI assistant. They may need a simpler form, a translated letter, a caseworker with time, or a policy that does not require the same information three times. A transport department may not need predictive analytics if it lacks basic maintenance data. A school may not need generative content if teachers are already overwhelmed by platform churn.
This does not mean AI is a distraction. It means AI must be subordinated to service design. The best deployments will be boring in the right way: a faster internal search process, a better translation workflow, a safer evidence-review pipeline, a more reliable data exchange between departments, a tool that helps staff prepare for a resident meeting without making the decision for them.
The worst deployments will be exciting in the wrong way. They will produce dashboards for leaders, pilots for press releases and opaque decisions for residents. They will call themselves smart while making the city harder to question. They will save time in one department by exporting confusion to another.
Public-sector AI succeeds when it makes responsibility clearer, not when it hides responsibility behind automation. That is a high bar, but it is the correct one. Cities are not start-ups seeking product-market fit. They are democratic institutions with duties to people who cannot simply choose a competing provider.
The European city may turn out to be the right scale for AI’s hardest public questions: close enough to residents to see consequences, large enough to build capacity, and constrained enough by law and politics to resist the fantasy of frictionless automation. If AI is to work for people rather than merely on them, city governments will need to keep making the unfashionable choices: clean the data, train the staff, reject the risky use case, demand accountable vendors and preserve human judgment where public power is at stake. The technology will keep accelerating; the cities worth watching are the ones making sure democracy does not have to run behind it.
The Eurocities Digital Forum in Sofia, held from April 15 to 17, arrived at a useful moment. The AI hype cycle has moved past the stage where every institution needed a “strategy” slide and into the less glamorous stage where procurement rules, staff training, data quality, audit trails and accountability matter more than demos. For cities, that shift is unavoidable. Unlike consumer AI apps, municipal AI does not live in a sandbox; it touches housing queues, traffic fines, public employment, classrooms, social services and the digital tools residents must use to get through daily life.
City Hall Discovers That AI Is Not a Product Category
The most revealing examples from European cities are not the flashiest ones. Espoo’s intranet AI assistant, designed to help roughly 12,000 city employees find internal information more quickly, sounds modest beside visions of autonomous urban operating systems. But that modesty is the point. A civil servant who can find the right policy, form, procedure or case note faster may deliver a better service without turning a resident into a test subject.Espoo is combining GPT-based tools and Microsoft Copilot-style workflows with a familiar public-sector problem: internal knowledge is scattered, outdated, duplicated and hidden behind organisational boundaries. AI does not magically fix that. If anything, it exposes the mess. A model trained or connected to inconsistent internal material will reproduce the confusion faster and with more confidence.
That is why Espoo’s lesson matters for every municipality now eyeing Copilot, Azure OpenAI, Google Gemini, local large language models or sovereign cloud alternatives. The first question is not “Which model should we buy?” It is “What information do we actually have, who owns it, how current is it, and who is allowed to use it?” In public administration, the AI project often begins as an IT deployment and quickly becomes an information-governance audit.
For Windows-heavy municipal environments, this is not theoretical. Microsoft’s ecosystem is already embedded in the daily work of many European city administrations: Teams meetings, SharePoint sites, Outlook inboxes, Entra ID identities, Intune-managed devices and Office documents stretching back years. AI assistants that sit inside that productivity stack can be powerful precisely because they are close to the work. They are also risky for the same reason.
The convenience layer can become a leakage layer. A poorly configured AI assistant may surface documents that employees technically had access to but never should have been able to discover at scale. A summarisation tool may blur the line between official policy and draft material. A chatbot may confidently answer a staff question using obsolete guidance. In a city administration, these are not mere annoyances; they can become unequal service delivery.
The Hague Draws a Line Where the Algorithm Wants to Rank People
The Hague’s reported refusal to use AI to classify job applicants is the kind of decision that deserves more attention than another chatbot pilot. It shows a public authority recognising that some uses of AI are not made safe merely by adding a human reviewer at the end. If a system has already sorted, scored or filtered people using biased data or opaque proxies, the “human in the loop” can become a rubber stamp with plausible deniability.Recruitment is a classic trap for public-sector AI. It looks administratively attractive because there are many applications, repeated processes and measurable outcomes. But the very data used to train or tune these systems often reflects past discrimination, institutional preference and unequal access to opportunity. If a city government deploys AI to rank candidates, it risks encoding yesterday’s workforce into tomorrow’s hiring pipeline.
The Hague’s approach also suggests that an AI strategy is useful only if it enables refusal. Too many institutional AI policies read like permission structures: here is how we will innovate, here is how we will pilot, here is how we will modernise. A serious public AI strategy must also say no, not later after harm appears, but early enough to prevent the harm from becoming bureaucratic routine.
That is especially important under Europe’s AI Act, which is moving from legal text into operational reality. The law entered into force in 2024, with key provisions on prohibited practices and AI literacy already applicable from February 2025 and broader obligations arriving in stages. For cities, this means AI compliance is not a future legal department project. It is now part of procurement, HR, education, policing support, citizen services and vendor management.
The AI Act’s risk-based logic aligns with what competent local governments already know: not all deployments are equal. An internal search assistant is not the same as an automated welfare-risk score. A traffic-flow model is not the same as a facial recognition dragnet. A translation tool for public notices is not the same as a system making eligibility decisions. The work of city government is to keep those distinctions visible when vendors would rather describe everything as digital transformation.
Riga Shows Why “Human Oversight” Must Mean More Than a Checkbox
Riga’s use of AI in traffic control, where systems help prepare evidence while a police inspector retains final authority over penalties, captures the practical compromise many cities will try to strike. AI can process images, detect patterns, organise case material and reduce administrative load. But the state’s coercive power still requires a human decision-maker who can be questioned, appealed and held responsible.That sounds reassuring until one asks what “final authority” actually means. Does the inspector meaningfully review evidence or merely approve a pre-filled recommendation? Can the inspector see why the system flagged a case? Is there a reliable way to detect systematic errors? Can residents challenge not only the fine but also the technical pipeline that produced it?
These questions matter because automation bias is not a theoretical flaw. People tend to trust machine output when it arrives packaged as objective, consistent and official. In a busy public office, a recommendation system can become the decision-maker in practice even when the formal policy says otherwise. The interface becomes the institution.
For cities, the standard should be stronger than “a person clicked approve.” Human oversight should mean trained staff, documented review, error sampling, appeal rights, clear public explanations and the ability to suspend a system when it behaves badly. It should also mean that the human reviewer has enough time and institutional authority to disagree with the machine.
This is where local government can either improve on the private sector or repeat its mistakes. The private AI economy often treats user friction as a failure. Public administration should sometimes treat friction as a safeguard. A pause, a second review, a written explanation or a refusal to automate a sensitive step can be the difference between efficiency and arbitrary power.
Data Spaces Are Europe’s Answer to the Smart-City Hangover
The smart-city boom of the 2010s promised dashboards, sensors and real-time optimisation. It also left many cities with fragmented platforms, vendor lock-in, abandoned pilots and data silos dressed up as innovation. The new European language of data spaces is, in part, a response to that hangover. It is an attempt to make data sharing less ad hoc, less proprietary and more governable across sectors.Riga’s work on mobility data, sensors and digital twins points to the practical need. Urban planning depends on information from transport systems, energy networks, climate models, road works, emergency services and citizen-facing platforms. If these systems cannot talk to one another, AI becomes a decorative layer over disconnected databases. If they can talk without rules, cities risk building surveillance infrastructure by accident.
The idea behind European data spaces is not simply to pool everything into one giant database. The stronger model is federated: data owners retain control, access is governed, standards allow interoperability, and use cases are built around defined public purposes. That approach fits the European instinct to make digital infrastructure accountable rather than merely scalable.
Oulu’s emphasis on trust and distributed control is significant here. Trust is often invoked as if it were a communications challenge: publish a policy, hold a workshop, reassure the public. But in digital government, trust is an engineering and governance property. It depends on who can access data, how access is logged, whether systems are auditable, how long data is retained, and whether residents have meaningful rights.
For IT professionals, this is the real architecture conversation. AI models will change rapidly; today’s frontier system becomes tomorrow’s commodity feature. Data architecture, identity management, security classification, retention policies and interoperability standards are more durable. Cities that neglect those foundations will find themselves repeatedly buying new AI tools to compensate for old data failures.
Digital Sovereignty Becomes a Budget Line, Not a Slogan
Ghent’s concern about who controls AI systems reflects a wider European anxiety that has moved from policy circles into procurement offices. Digital sovereignty can sound abstract, but for a city it becomes concrete very quickly. Where is the data stored? Which law applies? Can the city audit the model? Can it switch vendors? What happens if a service changes price, terms or technical behaviour?Local governments are not hyperscale cloud companies. They do not have unlimited engineering teams, spare data centres or the luxury of rebuilding every platform from scratch. They depend on vendors because vendors provide working systems, security updates, integration support and economies of scale. The question is not whether cities can avoid dependency altogether. They cannot. The question is whether they understand and govern the dependencies they accept.
This is where the Microsoft angle is unavoidable for WindowsForum readers. Many municipalities already sit deep inside the Microsoft stack, and Microsoft is aggressively positioning Copilot, Azure AI, Fabric, Purview and security tooling as a coherent enterprise AI platform. That integration is attractive because it can reduce deployment complexity. It is also strategically sticky.
A city that builds AI workflows around one vendor’s identity model, document graph, compliance tooling and assistant interface may gain speed but lose bargaining power. Exporting documents is easy; exporting institutional habits is not. Once employees learn to ask an assistant embedded in their office suite for policy answers, meeting summaries and case preparation, the assistant becomes part of the administrative fabric.
European cities are therefore right to ask hard questions before convenience becomes dependency. Sovereignty does not require rejecting commercial platforms. It does require exit plans, open standards where possible, contractual audit rights, clear data-processing terms, portability requirements and internal expertise strong enough to challenge vendor claims. A city that cannot understand the system it uses cannot credibly govern it.
The Public Sector’s AI Skills Gap Is a Democratic Problem
Bulgaria’s AI Readiness Index, with participating public administrations averaging 49 out of 100, is a useful corrective to the idea that AI adoption is merely blocked by cautious bureaucracy. Many public institutions are not resisting AI because they lack imagination. They are hesitating because they lack trained staff, internal policies, leadership engagement and confidence about legal obligations.That gap matters because unprepared institutions do not necessarily avoid AI. They often adopt it badly. Staff use public chatbots with sensitive material. Departments buy overlapping tools. Managers approve pilots without evaluation plans. Procurement teams accept vendor language they cannot test. Legal and security teams are brought in after the system is already politically committed.
AI literacy obligations under the European framework make this more than a best-practice issue. Public bodies deploying AI need staff who understand capabilities, limitations, risks and appropriate use. But the deeper point is cultural. A city cannot govern AI if only the IT department understands it, and it cannot use AI well if IT understands the technology but not the service context.
Bordeaux Metropole’s distinction between digital skills and digital culture is one of the more important ideas in the Eurocities discussion. Skills are knowing how to use a tool. Culture is knowing what the tool does to institutions, relationships and choices. A clerk can learn to prompt a chatbot in an afternoon; understanding when not to use one takes experience, ethics and organisational confidence.
For residents, the same distinction applies. Digital inclusion is not just broadband access, a smartphone and a password reset service. It is the ability to understand when an automated system is involved, what rights exist, how to challenge an outcome and how to participate in decisions about technology. A city that digitises services without building civic digital culture may improve efficiency while widening the gap between those who can navigate systems and those who are governed by them.
Schools and Civil Society Are Where AI Leaves the Office
Linkoping’s use of AI in education, with teachers creating learning materials, quizzes and interactive assignments, shows how quickly municipal AI moves beyond city hall. Education is one of the most promising and contested domains for generative AI. Used carefully, it can help teachers adapt materials, support multilingual classrooms and reduce preparation time. Used carelessly, it can flood classrooms with low-quality content and weaken trust in student work.The educational challenge is not simply whether pupils should use AI. They already will. The question is whether schools can teach students to understand AI as a tool, a source of error, a labour-saving device, a persuasion machine and a social force. That is a much harder curriculum than “write better prompts.”
Amsterdam’s work on digital citizenship and online harms adds another layer. Cities are seeing offline consequences from online life: mental health pressure, cyberbullying, conflicts that begin in digital spaces and spill into schools, neighbourhoods and public services. AI will intensify that problem by making synthetic content cheaper, impersonation easier and harassment more scalable.
Manchester’s framing of digital citizenship around inclusion, infrastructure and representation is important because AI policy can become elite policy very quickly. The people most affected by automated systems are often least represented in their design. Residents without reliable connectivity, language confidence, institutional trust or spare time are unlikely to attend consultation workshops about algorithmic governance. Yet they may be the first to feel the consequences.
Leipzig’s support for NGOs, clubs and associations recognises that civil society is part of the digital public-service ecosystem. Local organisations often mediate between residents and government, especially for vulnerable groups. If those organisations lack digital capacity, then AI-enabled public services may become harder to access precisely for the people who need support most.
The AI Agent Is Coming for the Workflow, Not the Press Release
Rotterdam’s warning that the future digital city will not be defined by AI alone is well timed because the technology itself is changing. The current wave of generative AI began in public consciousness as text generation and chat. The next wave is increasingly about agents: systems that can take instructions, call tools, interact with software, move through workflows and act with partial autonomy.For city governments, this raises the stakes. A chatbot that answers an employee’s question can be wrong. An agent that updates a case record, sends a message, schedules an inspection or triggers a procurement step can be wrong and consequential. The transition from advice to action is where governance must become sharper.
Most municipal processes were not designed for autonomous software actors. They rely on tacit knowledge, informal escalation, legacy systems, handwritten exceptions and human judgment about when a case is unusual. AI agents promise to automate the connective tissue between systems, but that connective tissue is often where accountability lives.
This is also where Windows and enterprise endpoint management become relevant again. If agents operate through browsers, office suites, line-of-business applications and cloud APIs, then identity, permissions and device security become central controls. A compromised account with an AI agent attached may be more dangerous than a compromised account alone. The agent can move faster, summarise more, attempt more actions and potentially obscure the operator’s intent.
Security teams will need to think about AI not only as data risk but as operational risk. Which agents can act on behalf of which users? What logs are retained? Can actions be replayed and explained? Are agents blocked from sensitive systems by default? Can a city disable an AI workflow during an incident? These questions belong in the same room as procurement and policy, not in a post-breach review.
The Vendor Demo Is the Easiest Part of the Job
The Eurocities examples share a useful scepticism: AI is powerful, but it is not needed for everything. That sentence should be printed above every municipal innovation lab. The public sector has spent years being told that every problem is a technology problem if described at the right conference. Cities know better because they inherit the messy human remainder after the demo ends.A resident trying to access housing support may not need an AI assistant. They may need a simpler form, a translated letter, a caseworker with time, or a policy that does not require the same information three times. A transport department may not need predictive analytics if it lacks basic maintenance data. A school may not need generative content if teachers are already overwhelmed by platform churn.
This does not mean AI is a distraction. It means AI must be subordinated to service design. The best deployments will be boring in the right way: a faster internal search process, a better translation workflow, a safer evidence-review pipeline, a more reliable data exchange between departments, a tool that helps staff prepare for a resident meeting without making the decision for them.
The worst deployments will be exciting in the wrong way. They will produce dashboards for leaders, pilots for press releases and opaque decisions for residents. They will call themselves smart while making the city harder to question. They will save time in one department by exporting confusion to another.
Public-sector AI succeeds when it makes responsibility clearer, not when it hides responsibility behind automation. That is a high bar, but it is the correct one. Cities are not start-ups seeking product-market fit. They are democratic institutions with duties to people who cannot simply choose a competing provider.
Europe’s Cities Are Writing the Practical Manual in Real Time
The useful lesson from Sofia is not that European cities have solved AI governance. They have not. The useful lesson is that the serious cities are asking operational questions early enough to matter.- AI pilots are moving into real municipal workflows, but the strongest examples start with a defined public-service problem rather than a desire to deploy the newest model.
- Human oversight is meaningful only when staff have time, training, authority and audit trails sufficient to challenge automated outputs.
- Data quality, interoperability and governance are now core urban infrastructure, not back-office technical concerns.
- Digital sovereignty is becoming a procurement and risk-management issue because cities increasingly depend on global platforms for essential administrative functions.
- AI readiness depends on public servants, residents and civil-society organisations understanding the technology well enough to use it, question it and refuse it.
- The next phase of AI agents will force cities to govern software that does not merely answer questions but acts across workflows.
The European city may turn out to be the right scale for AI’s hardest public questions: close enough to residents to see consequences, large enough to build capacity, and constrained enough by law and politics to resist the fantasy of frictionless automation. If AI is to work for people rather than merely on them, city governments will need to keep making the unfashionable choices: clean the data, train the staff, reject the risky use case, demand accountable vendors and preserve human judgment where public power is at stake. The technology will keep accelerating; the cities worth watching are the ones making sure democracy does not have to run behind it.