Using AI to talk with residents: ground rules

A good starting rule: automate what repeats and has a single answer, such as opening hours, the requirements of a procedure or the status of an appointment. Anything that affects a right, decides whether someone gets a benefit or communicates a ruling is decided and signed by a person at the body.

Write that boundary down before configuring anything. It helps you write the system's instructions and also tells residents plainly what to expect from each channel.

Residents have a right to know who they are talking to. The European Union requires it for AI systems that interact with people, unless it is obvious (Article 50 of the AI Act), and Ontario's directive in Canada asks its ministries and provincial agencies to disclose publicly when they use AI in a service. In Argentina, the Undersecretariat for Information Technologies published recommendations for trustworthy AI in the public sector in 2023, which are not mandatory.

Put the notice in the first message and in the channel's introduction: "I am the Municipality's automated assistant. If you would rather talk to a person, write HELP". Textra does not add that notice by itself: you write it in the agent's instructions or in your messages.

Decide in advance when to hand over: when the resident asks, when the question touches a personal case or when the system does not know the answer. Also decide who receives the conversation, during which hours and what happens outside them, so nobody is left without a reply.

The agent you build in Textra can hand the conversation over: it sends a courtesy message if you set one up, notifies a web address of your body and stops answering. Someone from your team can also step into the conversation whenever they want.

Not every resident has a smartphone or mobile data all month. A text message reaches any mobile phone, and a short message in plain language is easier to understand than regulatory text pasted as it is.

In Textra, notices can go out on WhatsApp, SMS or email, and automatic replies with an AI agent work on WhatsApp only. Textra does not translate messages or detect the person's language: if you need to communicate in other languages, you have to write those versions yourself.

Sending each group what applies to it avoids flooding everyone with notices they do not care about. The risk is at the other extreme: an opaque criterion can leave a neighbourhood, an age group or a community without information they needed.

Use simple criteria that you can say out loud (neighbourhood, type of query, month of the appointment) and check which groups do not receive each notice. In Textra you build audiences from a CSV with custom attributes and, before sending, you see the count and a sample of who will receive it.

If a resident asks why they got a given reply, someone at the body has to be able to answer. That is easier when frequent questions follow fixed steps and AI is kept for open questions, and when the system can only do what you authorised.

In Textra you can define a fixed-step flow, without AI, for questions with a single answer; flows are written in JSON, with no visual editor. Each agent can only use the tools from your systems that you enabled in its list. And each agent run leaves a record of which tools were used and when, without the text of the messages, which your team can check from the conversation.

In Australia, between 2015 and 2019, the scheme known as Robodebt claimed welfare debts calculated automatically from an average of annual income. It affected hundreds of thousands of people, and the Royal Commission that investigated it delivered its report in 2023 and concluded that the scheme was neither fair nor legal.

In the Netherlands there were two separate cases. In February 2020, a court in The Hague ruled that the law behind the SyRI system, which cross-referenced data to detect fraud, violated the right to private life, among other reasons because of its lack of transparency. Separately, the parliamentary inquiry concluded in December 2020 that basic rule-of-law principles had been violated in the childcare benefits case, and the government resigned in January 2021.

These are not chatbot cases, but they leave a design lesson: when a system can harm a person, there must be human review and a clear way to complain. That is what is worth settling before the first send.

Pick a single topic, a small group of residents and a short period. Measure which questions were resolved on their own, which were handed over and what the team corrected, and adjust the instructions with that information before expanding.

Check with your body's legal team which rules apply to the personal data you will handle. In Argentina, Law 25.326 on personal data protection applies, overseen by the Agency for Access to Public Information. Textra does not certify compliance with any rule: your body assesses that.