The political decision first
Before the first line of code, the city’s elected leadership must decide: what is the assistant’s role toward the resident? Information, orientation, processing of requests?
The answer commits the project. An assistant that only routes residents to the right department is easy to scope. An assistant that claims to process a request directly (registration, application, complaint) requires a full chain that goes beyond the usual scope of an AI assistant.
The assistant doesn’t replace public service. It makes it more accessible. If the promise drifts toward replacement, the project derails.
Defining the scope of answerable questions
Five families of questions a municipal assistant can handle without risk:
- Practical information: hours, addresses, department contacts.
- Administrative procedures: required documents, processing times, available online services.
- Local life: calendar, events, markets, school-related services.
- Benefits and programs: eligibility, conditions, forms.
- Local regulations: basic urban planning, parking, cleanliness.
Three families to exclude: complex individual situations (family cases, disputes), emergencies (immediate redirect), and anything that touches the mayor’s office and political action.
Choosing accessible sources, without loading everything
The classic mistake: “let’s load the entire website and all our council resolutions.” The result: the assistant returns answers based on ten-year-old archives, including outdated information.
Sources to prioritize:
- The official website (current pages only, not archives).
- Procedure sheets validated by the departments.
- The city calendar (with API connection if one exists).
- Council resolutions from the last three years, clearly tagged as such.
Every loaded document must have a clear status: in force or archived. Otherwise the assistant returns obsolete information.
Organizing human escalation
Any assistant open to the public must have a clear exit to a human:
- A “speak to an agent” button always visible.
- Phone and email of the relevant department displayed next to every sensitive answer.
- Automatic detection of sensitive topics (emergency, minors, violence) with immediate redirect.
Out-of-scope questions are never “improvised” by the assistant. It says it doesn’t know, and routes the user.
The question of language: level, tone, accessibility
Three principles:
- A2 level by default (accessible to non-native speakers and to readers with literacy difficulties).
- Consistent formal address, without administrative jargon.
- Short sentences (15-20 words), one idea per sentence.
A municipal assistant that reads like a city ordinance fails. It should speak like the most available and patient front-desk agent in the building.
Regular audit of conversations
All conversations are logged (anonymously). A monthly review by the communications team identifies:
- The most frequent questions (to better document).
- Questions the assistant refused to answer (to decide: extend scope or hold the line?).
- Sensitive questions detected and redirected (to monitor).
Without this feedback loop, the assistant freezes and loses relevance within six months.
Results in numbers
These figures are medians across three pilot municipalities (15,000 to 50,000 residents). Satisfaction hinges on two variables: response speed, and clarity of escalation to a human when the assistant doesn’t know.
See also Five use cases in public services and our local government page.
Twenty minutes to identify the minimum viable scope for your resident-facing assistant. We look at the website, the organization, and the existing sources.
Book a demo→