When we talk about adopting artificial intelligence in the public sector, we tend to look far ahead. Perhaps too far: autonomous agents, process automation, decision support, service personalisation, predictive analytics. Then come infrastructure, technological sovereignty, proprietary versus open-source models, governance and security.
These are all important issues. Some are essential to our very ability to govern this transformation, yet remain systematically underestimated — technological sovereignty being one obvious example. Equally important is the need to start from a shared vision that can guide the adoption process. And yet I have the feeling that, while imagining how AI will profoundly transform public administrations over the coming years, we risk overlooking what we can already do with it today.
This does not mean that long-term vision is any less important. Quite the opposite: without a clear direction, the risk is to accumulate disconnected initiatives and technology without a strategy. But building that vision does not prevent us from acting in the present. There are concrete applications, already available and immediately actionable, that can create value while the deeper transformation takes shape. More importantly, they can already deliver a very tangible benefit to citizens.
From this perspective, I believe there are two areas that are very simple to describe, yet enormously important: helping public administrations make themselves easier to understand, and helping them understand people better. They may sound like almost trivial goals. In reality, they touch one of the deepest problems in the relationship between citizens and institutions: complexity, and the sense of distance that complexity creates.
Complexity Is Not Just a Design Problem
Anyone working on public service design has been trying to reduce this complexity for years. We simplify processes, redesign services, improve websites and applications. We have built design systems, reorganised information architectures, rewritten content and tested all of this with real people.
All of this matters. But part of the complexity does not originate in the interface, and therefore cannot simply be solved by designing a better interface.
A large public institution applies laws, regulations and procedures that have been built up over decades. New provisions are layered on top of previous ones: some amend them, others introduce exceptions, transitional rules, new requirements, thresholds or incompatibilities. Complexity accumulates and grows year after year.
On top of this, responsibilities may be distributed across different institutions, individual circumstances can affect access to benefits, and procedures may vary depending on the specific situation. We can and should continue simplifying everything that can be simplified, but realistically we will not be able to remove decades of regulatory and procedural layering in the short term, nor rapidly transform the way legislation itself is drafted — a process that will gradually need to move closer to the principles of legal design. That is anything but a simple task.
So the question can be reframed: if we cannot immediately eliminate all this complexity, can we at least stop passing it on to citizens?
This is where, in my view, AI can already help.
Today, We Do the Translating
To navigate public services, people still need to understand at least part of the logic of the administrative system. They need to know which benefit or service to look for, where their problem fits, which term to type into a search engine, perhaps even the official administrative name of the service or, worse still, remember an often completely opaque acronym. Then they have to find the right page, read it and understand it.
In other words, today it is often the citizen who has to do the translation work.
People start from a need expressed in their own language — “I lost my job”, “my child was born”, “my mother is no longer self-sufficient” — and have to translate it into the language, categories and structures of public administration. Artificial intelligence gives us the opportunity to reverse this relationship: rather than asking people to learn how the administration thinks, we can build systems capable of understanding how people think and speak.
To me, this is the real promise of an advanced virtual assistant. Not a more sophisticated version of old-style FAQs, and not a box where users type a question hoping they have guessed the right words, but a translator between citizens and bureaucratic language.
People should be able to describe what is happening to them in ordinary language. The system should then interpret the need, map it onto the underlying administrative structure, identify the relevant information and services, and return an understandable answer.
Put simply: today we translate ourselves into the language of public administration. Tomorrow, public administration can translate itself into ours.
But First, We Need Something Clear to Say
This immediately raises another problem. An assistant may understand a question perfectly well and produce an apparently flawless answer, but in a public administration that answer must not merely sound convincing: it must be correct. And to be correct, it has to rely on trustworthy sources.
This is far less straightforward than it may seem, because in a large public institution the source of truth is often not a single document.
The same service may be described across dozens of different sources produced at different times: a law may have established the original framework, a later decree may have changed some eligibility requirements, a circular may have clarified a procedure, and a subsequent notice may have updated a deadline or introduced a new operational rule. Meanwhile, everything that has not been changed may still be governed by a document written many years earlier.
None of these sources, taken in isolation, necessarily contains a complete and up-to-date description of the service. The correct answer therefore has to be reconstructed by combining information distributed over time, distinguishing what is still valid from what has been amended, supplemented or superseded.
This is where retrieval becomes much more difficult than simple document search. It is not enough to find the documents that are semantically closest to the question. The system must identify the sources that are actually relevant, determine which parts of each source apply to the specific case, establish which information is still current, and recombine it into a single coherent answer.
Simply connecting a language model to thousands of documents does not solve the problem. It gives the model the same complexity that we are trying to avoid passing on to citizens.
Something is needed in between: a knowledge base designed for people — reliable, controlled, up to date, coherent and written in understandable language. And this is where the second immediate use of AI comes into play.
First, Let AI Help Us Write Better
For several years, at INPS we have been working systematically on simplifying content. We have defined writing guidelines, standardised information structures, adopted metrics such as the Gulpease readability index, and introduced user testing to verify whether rewritten texts are actually easier to understand.
The obvious problem is scale. A large public institution produces and manages thousands of web pages, letters, forms and communications. Rewriting everything manually is possible, but it requires an enormous amount of time and expertise.
This is where generative AI can become a very powerful accelerator. Not by replacing human oversight, but by producing a first transformation of the content: simplifying vocabulary, reducing syntactic complexity, improving structure and preserving meaning.
We tested this approach on dozens of pieces of INPS content and then validated the results with users. Across 50 service pages, with a panel of more than 1,600 people and around 5,000 evaluations, simplified versions were preferred to the originals in more than 70% of cases. Versions produced with AI support achieved levels of quality and readability comparable to — and in some cases higher than — those achieved through specialist human rewriting.
The interesting point, however, is not to prove that “AI can write well”. We already know that. The real point is that, for the first time, we have a tool that can help us build at industrial scale a body of public knowledge designed to be understood.
And that same knowledge can become the foundation on which virtual assistants operate.
The Two Things Become One
At this point, the two workstreams converge. On one side, we use AI to transform complex content into information that is simpler, more structured and easier to understand. That content is then checked, measured and validated, and becomes the source on which the assistant relies.
The assistant can then do what modern language models are particularly good at: understand a freely formulated question and identify, within a controlled body of knowledge, what is needed to answer it.
The process can also continue to improve itself. User questions reveal where an explanation does not work, where information is missing, or where a piece of content creates ambiguity. That evidence can be used to improve the source, and a better source in turn produces better answers.
A cycle emerges:
At that point, the chatbot stops being an isolated object sitting on top of a website and becomes the interface to an information system built around people.
We Do Not Need to Wait for the Public Administration of the Future
This is perhaps the part I find most interesting. To do these two things, we do not need to wait until the entire public sector has been redesigned around artificial intelligence, nor until processes have been fully automated. We do not even need to wait until every institution has built its definitive AI infrastructure.
We can start now.
We can use artificial intelligence to make public knowledge simpler and easier to understand. And we can use that same knowledge to build assistants capable of understanding people and taking on, on their behalf, part of the complexity of public administration.
These are two relatively focused levers, but they act exactly where citizens meet public institutions. More importantly, their impact can be measured: we can find out whether people understand a text better, whether they find what they are looking for more easily, whether they identify the right service, take less time, make fewer mistakes or complete a task with less effort.
And perhaps this is exactly where we should start. The debate about the profound transformation that AI will bring to public administration is necessary and will continue for years. In the meantime, however, we can do something much simpler: