AI in public administration now has rules. It still needs a method
AI should free up human time in public administration, without requiring frontier models. The AI Act and Italian law handle the easy part; the method is still missing.
In 2023 I wrote that public administration should take advantage of artificial intelligence, “always bearing its limitations in mind”. It was a cautious piece, written while I was standing in the Ivrea local elections. Looking back, it had a flaw: it treated the issue as a technical question. Risks on one side, opportunities on the other; just balance them. That is not how it works. It is a political choice, and always has been.
Where I stand: I want it inside, not outside
Let me clear up a misunderstanding. I am not among those who would keep AI away from public offices. Quite the opposite. I believe it should become part of their processes, without too much hesitation. Not because it is fashionable, but for a practical reason. Public administration is drowning in repetitive work: logging correspondence, routing requests, finding documents, translating forms, answering the same question for the hundredth time. Every hour spent on that is an hour taken away from what public administration exists to do: meet people and handle real cases, the ones a machine cannot and should not manage.
That is the point, rather than cutting costs. If AI takes over the mechanical work, human time goes back where it matters: direct contact with citizens and decisions with real consequences. Efficiency is not the goal; it frees people for work of greater value.
You do not need the world’s most powerful model
Something rarely heard at conferences needs saying: none of this requires a frontier model, the latest release at an astronomical price. Almost never. A municipality’s tasks are narrow and dull: understand and route a request, summarise a resolution, translate a page, answer recurring questions. A smaller model, perhaps an open one hosted in-house, can do these things very well at a fraction of the cost.
Chasing the biggest model is the same mistake as buying the most expensive management software because “it must be the best”: you pay for power you will never use, which is even harder to justify with public money. The job is to find the right balance between what is actually needed and what it costs. It is an engineering decision before it is a political one, and almost nobody makes it when writing procurement specifications.
Two years later, the context has changed. Since 2 August 2026, the European AI Act has been fully in force. Since 10 October 2025, roughly a year ago, Italy has had its first AI law, Law 132. We are no longer asking “is this allowed?”. The rules exist. That is a good thing.
The problem is that the rules have done the easy part.
What they have resolved
Three things in particular deserve recognition.
First, responsibility. Italian law says AI in public administration is a supporting tool, and decisions remain with a person who is accountable for them. No administrative decisions signed by an algorithm. That sounds obvious, but it was not: for years, the temptation was to hide behind “the system decided”.
Second, prohibitions. Some uses are simply illegal. Social scoring of citizens by a public body, for example, is not something to debate case by case. It is prohibited. Good.
Third, the obligation to look before acting. For systems considered high-risk, of which public administration has many, from welfare to access to essential services, the authority must assess the impact on people’s rights before switching the system on. Not after the damage is done. You can no longer experiment on citizens and see what happens.
For all its complexity, the European framework has one merit here: it makes you think in terms of risk rather than enthusiasm.
What they have not resolved
Everything else. In other words, what really matters when you run a public authority.
A law tells you what you cannot do. It does not tell you how to do things well. All the difficult questions are still there.
Who inside the municipality has the skills to understand what they are buying? Almost nobody. So you rely on a supplier, who sells you its solution, with its data, servers and renewals. Two years later, you are locked in and leaving costs more than staying. Italian law has even removed the requirement to keep data on national servers: debatable, but at least an explicit decision. No regulation will solve supplier dependence for you. You solve it in the procurement specifications, if you know what to write.
Then there is real transparency. Not the little “you are talking to a virtual assistant” notice, which is the bare minimum required by the rules. I mean telling citizens which system you use to decide what, with which data, and how they can challenge an answer. The rules do not spell that out in detail. It is a choice, and almost nobody makes it.
The risk is using the rules as an excuse
There is a guaranteed way to get this wrong, and it is the easiest: treating the AI Act as a rubber stamp. “We are compliant”, written in the report, and off you go. Compliance as a fig leaf.
We will see it in the coming months. Municipalities proudly announcing a chatbot “with artificial intelligence” on their official website because it looks modern, while their correspondence office still takes fifteen days to answer a certified email. Technology as a press release. Complying with a law does not give you a method. It merely keeps you from being fined.
One thing I have not changed my mind about since 2023, speaking more as someone who did not get elected than as an engineer: transparency does not eliminate disagreement; it makes it more mature. That applies to judging Carnival teams and, far more seriously, to how public administration makes decisions about people’s lives. Explicit criteria, declared choices, responsibility resting with someone who has a name. The rest is marketing.
If I stood for election again
It could happen sooner or later, perhaps somewhere other than Ivrea. I already know how I would approach this, because it concerns how to govern rather than technology.
I would put AI to work where it saves real time, rather than where it makes headlines. No frontier models to answer four recurring questions. Skills inside the municipal organisation first, tools afterwards, never the other way round. Procurement specifications written by people who understand what they are buying. And one simple rule, stricter than any article of law: for every system that affects a citizen, that citizen must be able to know it exists, what it does and whom to complain to.
The rules are finally here. The people governing must provide the method. That does not come from Brussels.
Note — I wrote this article with help from AI, using it to research the legislation, structure the text and refine its wording. The approach, argument and opinions are mine, as they should be given the subject. It seemed right to say so.