The Agency for Administrative Modernisation (AMA) runs the official portal through which citizens reach the Portuguese state. Behind it sit more than 2,000 public services, and the people asking about them range from lawyers to foreign residents to citizens who struggle with written text.

Thousands of services, no single place to ask

Finding the right service meant navigating phone queues and portals that were hard to follow. There was no one place where a citizen could simply ask.

The same questions, again and again

Human operators carried most of the administrative workload, answering recurring questions that did not need a person.

Access that depended on language and literacy

How easily a citizen could get an answer depended on the language they spoke and how comfortable they were reading and writing. Public information had to be reachable by everyone.

The assistant sits on Portugal’s official government portal, where citizens already look for public services. They ask a question in their own words, in text or by voice, and get an answer drawn from the government’s own content on more than 2,000 services.

Answers about public services cannot be approximated, so the assistant uses retrieval anchored in official government documentation rather than open sources. Each answer is built from the relevant official content, which keeps it accurate and consistent with what the agency itself publishes.

Language was part of the design from the start. The assistant works in twelve or more languages, including Portuguese, English, French and Spanish, so foreign residents can ask in the language they are most comfortable with. The same answer also has to work for a lawyer and for a citizen with low literacy.

For citizens who cannot or will not type, an avatar offers a voice interface alongside the text chat. A person can speak a question and hear the answer, which opens the portal to people a text box alone would leave out.

The project was measured from day one against three targets: conversations handled per day, answer accuracy and escalation rate to human operators. Those same measures now show the result: more than a million conversations since January 2025, 94% accuracy and 6% escalation.