AI in public administration: the new AgID guidelines call for a focus on service value

AI in public administration: the new AgID guidelines call for a focus on service value

From virtual assistants to better public services: why AI adoption should start with the needs of citizens and public organisations.

A citizen finds an answer in seconds. The wording is clear, the link works and the conversation feels natural. But after reading it, do they actually know what to do next? We believe this question belongs at the centre of artificial intelligence projects in public administration: which obstacle have we removed from the person’s journey and from the work of public offices?

AgID’s new guidelines offer an opportunity to address that question systematically. Our view is that a project’s value should be assessed through the service it improves, with a clear definition of what meaningful results would look like before implementation begins.

AgID’s guidelines: a structured adoption process

Through Determination No.169 of 9 November 2026, AgID published its guidelines for AI adoption in Italian public administration. The final document addresses governance, risks, skills, data quality, personal data protection and security. Section 4.5.2 requires performance indicators for each use case and their monitoring throughout its life cycle.

This focus on results matters, but it needs careful interpretation: AgID does not require every assistant to manage entire administrative procedures. Regulatory compliance and the usefulness of a design are distinct questions. The adoption guidelines are final; development and procurement guidelines follow a separate adoption process that is still ongoing, according to AgID’s AI information page.

When an answer already makes a difference

Explaining a difficult page, helping someone find a form or guiding people unfamiliar with administrative language can create practical value. A conversational assistant can be a sound choice when it removes a real obstacle while keeping sources and alternative contact channels accessible.

The risk of superficial innovation arises when the presence of AI itself becomes the evidence of success. Conversation counts alone do not tell us whether people understood the steps or whether offices receive more complete requests. A system can consistently provide correct answers and still be of little use. This happens when it measures the quality of the information it provides without checking whether that information actually helps citizens achieve their goals.

An example: understanding how to start a business

Consider a hypothetical example, rather than a customer project: someone wants to open a business and asks where to start. An assistant directs them to the SUAP page, Italy’s one-stop shop for business activities. That is useful guidance, but it may not be enough to understand which information to prepare.

We could envisage more focused support: clarifying the intended activity and municipality, finding the relevant official information, explaining unfamiliar terms and suggesting points to check. The journey should distinguish established information from matters requiring confirmation by the responsible office.

The difference is not necessarily the amount of text generated. It lies in recognising the difference between a generic enquiry and a specific situation, asking the necessary questions and building a verifiable path based on the official information available.

This does not mean inventing a universal procedure or automatically deciding whether a person meets the requirements. Nor does it imply submitting applications or making decisions on their behalf. A useful system makes its boundaries explicit and helps people reach the appropriate next step, preserving responsibility for checks and decisions.

Five questions before choosing a solution

  1. Which problem are we solving? Identify an observable obstacle: scattered information, repeated enquiries, missing documents or steps people struggle to understand. “Adding AI to the portal” describes a technology, rather than a need.
  2. What changes for the user? Define the intended outcome from their perspective. For instance, recognising the relevant service and knowing what to prepare. Include people who prefer conventional channels or use assistive technologies.
  3. How will we measure the result? Compare the starting situation with a limited pilot. Possible indicators include understanding the steps, time spent finding information and errors detected, with declared criteria and samples.
  4. How reliable is the information? Establish authorised sources, ownership of updates and a process for outdated or conflicting information. Test ambiguous questions and situations where the system should stop and refer someone to a person.
  5. How does it connect with office workflows? Check existing systems, permissions and handover arrangements. An integration makes sense when it removes an unnecessary step without introducing new risks or duplication.

Measuring the service and interpreting the signals

Fewer telephone calls might indicate clearer information, but could also mean people cannot find a contact number. More conversations might show interest or reveal confusing navigation. Interpretation requires several sources of evidence: reviewing journeys, testing with users and discussing findings with the people who handle enquiries.

In the SUAP example, we could check whether people can identify the steps they need to investigate after receiving guidance. We could also assess how often answers refer to relevant, current sources. These are proposed evaluation criteria to agree on, rather than achieved results or promises to shorten administrative processing times.

What a software company contributes

Translating a need into requirements, connecting systems and checking behaviour before release are design and engineering tasks. This is why we combine AI opportunity assessment with software development and systems integration. The appropriate choice might be to configure an existing tool, build a dedicated component or improve data and content first.

In some cases, the best result could come from better organised content, a simpler procedure or an integration between existing systems, without introducing new artificial intelligence features.

A credible project also defines responsibilities, maintenance and the conditions for moving forward. If a pilot does not meet the agreed criteria, we can improve it or stop. Discovering that within a limited scope provides useful evidence for an investment decision.

Start with the problem and make value testable

Before asking which artificial intelligence solution to adopt, a public organisation should ask which problem it intends to solve. The success of AI in public administration should be measured by its ability to make services more understandable, accessible and effective, rather than by the number of virtual assistants deployed.

Real innovation is not adding another interface between citizens and public administration. It is narrowing the gap between a need and its solution.

Through our AI Opportunity Assessment, we can examine processes, data and alternatives, defining an initial experiment and criteria for evaluating it. Request a discussion about AI opportunities: we start with the service to improve and the people who use it.

Detailed reference: AgID guidelines, final version dated 9 November 2026, particularly sections 4.5 and 4.12 and chapters 9–11. The proposed design criteria and example express our operational interpretation.

Do you have a digital process or project to evaluate?