The most interesting part about Julia is not that Roma Capitale built a chatbot. In 2026, a chatbot alone is no longer news. The interesting part is the ambition the project has taken on over time: turning artificial intelligence into a new gateway to the city and, progressively, to the administrative machinery.

Julia was created with a relatively defined objective: helping those who live in, visit, or pass through Roma navigate mobility, culture, events, tourism, and essential services. In just over a year, however, the project expanded its scale. From the initial March 2025 release, it advanced to version 2.0, and then to 2.5 in June 2026, integrating directly into Roma Capitale’s official portal. This is where Julia stops being just a digital guide and becomes a more significant experiment: determining whether natural language can replace at least part of traditional navigation in the Pubblica Amministrazione.

It is a shift that warrants attention. Roma Capitale’s portal logs approximately 18 million visits a year, hosts over 160,000 pieces of informational content, and provides around 90 online services. For citizens, the issue is rarely a lack of information: it is finding it, figuring out which page is up to date, which office is responsible, which documents are required, and what the next step is. Julia aims to tackle precisely this friction.

From tourist guide to city interface

The project was announced in October 2024, during Microsoft’s AI Tour in Roma, as a virtual tourist guide developed with contributions from Microsoft and OpenAI technologies. The initial idea was straightforward: provide tourists and residents with a multilingual assistant capable of querying data from local operators and accredited sources, suggesting itineraries, and providing information on transport, museums, restaurants, events, and lesser-known sights.

The first public release arrived on March 7, 2025. Roma Capitale unveiled Julia as its first urban generative artificial intelligence: an assistant available on WhatsApp, Telegram, Messenger, and web chat, capable of communicating in over 80 languages. The technology behind the initial version used GPT-4o, engineered into the project with Microsoft’s support, while NTT DATA worked on developing the first release alongside other partners.

Even in version 1.0, the architecture’s most compelling trait emerged: Julia was not designed to freely search the web for just any information. Instead, responses are built from a certified knowledge base composed of official data from Roma Capitale and other validated sources. It is a less flashy choice than promising an assistant that “knows everything,” but far more sensible for a public service.

The real value: not knowing everything, but knowing where the answer comes from

Generative artificial intelligence has a structural issue: it is very good at producing linguistically persuasive answers even when the information is incomplete or inaccurate. In a creative engine, this flaw may be tolerable. In an official city assistant, it is much less so.

If a municipal system gets an exhibition's opening hours wrong, the damage is limited. If it makes a mistake regarding an administrative procedure, a deadline, a transit rule, or information tied to a public service, the response carries a different weight precisely because it originates from an institutional entity. Users naturally tend to place greater trust in it.

This is why the concept of a certified knowledge base is probably Julia’s best starting point. The goal should not be turning the municipality into a local version of ChatGPT, but building a system that thoroughly knows a defined set of information, cites its sources, and recognizes its own limitations.

The New York case shows just how real this issue is. In 2024, the MyCity chatbot, launched to help businesses navigate regulations and services, was found providing erroneous guidance even on labor laws, housing, and commercial obligations. In some tests, it went so far as to suggest unlawful conduct. The experiment became a textbook case: when a generative response bears the seal of a public administration, a hallucination is not merely a technical glitch, but a matter of institutional trust.

Julia starts from a more cautious philosophy. That does not mean it is immune to errors: Roma Capitale itself warns that, like any AI-powered chatbot, it can make mistakes. It does mean, however, that the project attempts to reduce risk through the quality and validation of sources rather than relying solely on the model's generalized capabilities.

Julia 2.0: memory, geolocation, and the first agents

With version 2.0, presented on July 31, 2025, the project picked up the pace. The model shifted from the GPT-4o family to GPT-4.1, the system prompt was revised, and new specialized agents emerged. Roma Capitale announced three additional agents integrated into a new multi-agent architecture.

The concept is significant. Instead of entrusting every query to a single general-purpose system, the architecture distributes workloads across domain-specialized components. It is an increasingly common logic in agentic AI: an orchestrator interprets the request, identifies the required competencies, and coordinates different modules before delivering a single response.

Version 2.0 also introduces geolocation on WhatsApp, Telegram, and web chat. Users can share their location and receive contextual information about what is nearby. New informational domains have been added, including public parks, places of worship, sports and leisure, as well as news and updates on the city and mobility. Existing information has also been improved, covering sharing mobility, ZTL, study rooms, Roma Pass, and MIC Card.

This is where Julia begins to look less like a “tourism bot” and more like a conversational urban infrastructure. A question such as “where can I study near me late at night?” combines location, hours, services, and context. It is precisely the kind of request that a traditional website handles poorly and that a conversational interface can make far simpler.

The 2.5 leap: Julia enters Roma Capitale

Version 2.5, unveiled in June 2026, is arguably the most significant milestone to date. Julia is integrated directly into the institutional portal. The widget remains accessible while browsing, maintains conversational context, and can return direct links to official pages, documents, and sources.

The point is not cosmetic. A portal with 160,000 pieces of content cannot be made truly simple through better menus alone. Its complexity mirrors that of the administration itself: departments, municipalities, regulations, procedures, services, tenders, forms, and contact details. A conversational layer can instead allow users to start from the problem rather than the bureaucratic structure.

The difference is substantial. Citizens should not necessarily have to know whether a procedure falls under a department, a municipality, or a specific office. They should be able to ask: “I need to do this, what documents do I need, and where do I have to go?”. If Julia can translate organizational complexity into an understandable journey, its value is not technological: it is administrative.

In 2.5, the architecture evolves further toward specialized agents coordinated by an orchestrating Super Agent. Official sources also indicate new informational domains, from cinema listings to post offices, public water fountains, waste bins, Punti Roma Facile, and anti-violence centers. The project is developed with an ecosystem of partners that includes, across the various releases, Microsoft, Intellera, NTT DATA, Digitalia, BIP, Fastweb, and 01 Sistemi.

This is where Public Administration can truly change

To understand Julia’s potential, one must take the term “artificial intelligence” off the table for a moment. The problem it attempts to solve has existed for decades: administrations hold large amounts of information but present it according to internal logic, whereas citizens think in terms of needs.

“I got a fine,” “I need to change my address,” “I want to know if the ZTL is active,” “I’m looking for an anti-violence center,” “where can I play sports with a voucher?”, “which documents do I need for this procedure?”. These are questions that cut vertically across different databases and organizational structures.

A well-designed conversational system can become a translator between these two worlds. It does not make decisions on behalf of the administration and does not replace the civil servant; it reduces the cognitive cost required to reach the right information.

This is perhaps the most important distinction to maintain. Public AI performs best when it helps citizens understand, find, and prepare. It becomes far more delicate when it shifts from informational support to decision-making: granting or denying a benefit, classifying a person, setting priorities, or producing administrative effects. Today, Julia operates primarily on the former ground, and that is an advantage.

Privacy: a solid approach, but communication must be crystal clear

Julia’s privacy policy includes several significant choices. Roma Capitale states that user data is not used to train the chatbot and that, under normal use, personal data is retained for a maximum of 24 hours, strictly limited to the time needed to retrieve the conversation. Conversely, opening an inquiry with 060606 or 060608 operators triggers a separate step leading into the CRM, with user consent.

There is also an important distinction between the official web chat and third-party channels. If Julia is used via WhatsApp, Messenger, or Telegram, the policies of the respective platforms also come into play. This is a point that should be communicated with great clarity, especially because one of Julia’s strengths is precisely its availability on channels users already rely on.

One element remains that deserves a clearer explanation. The presentation of Julia 2.0 mentioned a conversation history extended to seven days; the current privacy policy, however, sets a 24-hour maximum for retaining personal data during standard chatbot use. The two statements may be technically compatible—for instance, through a distinction between conversational state and personal data—but for a non-technical user, this distinction is not straightforward. A public project should provide full transparency at this level as well.

The question of costs: metrics are needed, not just releases

A public AI infrastructure cannot be evaluated solely on the quality of its demo. It must also be assessed on its cost and, above all, on the outcomes it delivers.

The PNRR pages of Roma Capitale indicate, within the broader digital tourism transformation ecosystem, a 1.7 million euro intervention for the “The Roman Land” app, designed to integrate with the Tourist Virtual Assistant platform. It is not accurate to use that figure as the “cost of Julia”: it refers to a broader, connected project, not a declared total for the assistant. Yet this very distinction demonstrates why publishing an easily readable financial breakdown of the service over time would be useful.

How much does it cost to serve a thousand conversations? What is the impact of the model, cloud infrastructure, database maintenance, testing, and integrations? How much work is saved for the contact center? How many people manage to complete a procedure without calling or physically visiting a service desk? How many answers are deemed incorrect? How many are escalated to a human operator?

These are the metrics that separate a showcase project from genuine public digital infrastructure.

The benchmark shouldn't be ChatGPT, but Singapore

To compare Julia with other public systems, looking at commercial chatbots is not enough. A compelling benchmark is Singapore, where GovTech manages VICA, a conversational platform shared by over 60 government agencies and used for more than 100 chatbots. GovTech reports an average of over 800,000 monthly queries.

This figure is not meant to establish a ranking between cities. It serves to illustrate how evaluation changes when AI becomes infrastructure: number of entities served, query volume, operational continuity, user satisfaction, resolution rate, and reduction of manual workloads. Singapore measures the system as a public digital service, not as a tech showcase.

This is the transition Rome needs to make with Julia. After the rollout phase, the phase of public and comparable results must follow.

A smart choice: no advertising, no opaque commercial ranking

Another characteristic of Julia deserves protection. Roma Capitale has introduced the service as free from advertising, whether explicit or covert. Businesses can register their activities free of charge on the 060608 portal to have them included in the assistant’s knowledge base.

This is an important decision, because an urban assistant can very quickly turn into an economic intermediary. When asking “where can I eat nearby?” or “which hotel do you recommend?”, the system determines visibility and, potentially, revenue. The neutrality of the response criteria must therefore be treated as part of the service’s governance.

In the future, it would be beneficial to make public the principles by which Julia selects or ranks equivalent businesses: distance, relevance, opening hours, accessibility, data completeness, and result rotation. The more the assistant is used, the more economic significance these rules will carry.

The biggest risk is not that Julia makes mistakes. It is that we stop verifying it

Every generative system will make mistakes sooner or later. Demanding zero errors is unrealistic. The quality of a public project is measured by how errors are prevented, identified, and corrected.

For Julia, this means establishing an ongoing evaluation framework: sets of critical test questions, regression testing with every release, source monitoring, user feedback, audits on sensitive domains, and a straightforward mechanism to hand off to a human when an answer lacks sufficient certainty.

The inclusion of official links and sources in responses is already a positive step, as it enables users to verify information. But the next step could be making the confidence level explicit, or at least clearly distinguishing between verified facts, suggestions, and content that requires confirmation from an administrative office.

The next challenge: from an assistant that informs to one that guides

The real transformation will happen if Julia manages to guide citizens through a process without becoming the decision-maker. Today, it can explain which documents are required. Tomorrow, it could confirm that the user has understood the steps, directly open the appropriate service, pre-fill non-sensitive details, remind them of deadlines, or pass the context on to an operator.

This is the compelling frontier of AI in the PA: not replacing the administration, but rendering part of its complexity invisible.

For a city like Rome, the potential is particularly significant. The capital brings together nearly three million residents, millions of city users, a global influx of tourists, immense cultural heritage, and a highly complex administrative machine. It is precisely the kind of environment where a conversational interface can generate real value if governed effectively.

Why Julia is a project to watch

Julia is still a young project. It would be premature to describe it as a completed revolution of the Pubblica Amministrazione. But it would be just as reductive to dismiss it as merely another institutional chatbot.

Its trajectory is compelling: certified sources instead of indiscriminate web scraping, progressive specialization of agents, integration with channels citizens already use, geolocation, direct integration into the institutional portal, and a gradual expansion from tourism-focused content to administrative services.

Rome is achieving something that many administrations announce but few successfully push into production: using a generative model as the operational interface of a real city, with real data and real citizens.

Now comes the hardest part. Julia’s next release will matter less than the figures Rome will be able to publish regarding its usefulness. Accuracy, resolution rate, time saved, accessibility, escalation to operators, satisfaction, and costs should become as much a part of the narrative as the model used or the number of agents.

If this transition occurs, Julia could become more than just a good urban innovation project. It could become a replicable model for how artificial intelligence enters Public Administration without turning governance into a black box: technology up front, sources and public accountability behind it.

Sources