In short. A chat assistant on a site that explains insurance, health or law should say less than the model is able to. guide-assurance.ch, a site about insurance in French-speaking Switzerland, has put one online. It answers from the published guides, cites its sources, declines personal advice, collects no health data, does not store conversations and hands over to a call-back form where consent is given outside the chat. Here is how those rules become technical choices.
The risk: an assistant that says more than the site
A serious information site explains the rules, cites its sources and lets the reader decide. A language model has none of these reflexes by default. Ask it which health insurer to pick and it will produce a fluent, personal answer, sometimes right, sometimes invented, always confident. In insurance, advising a person with a view to a contract is a regulated activity (art. 45 ISA sets the intermediary's duty to inform). The assistant therefore has to stay on the ground of general information.
What the assistant in this example does
General answers, always sourced
The assistant answers only from the site's guides, retrieved by searching the published pages. Every answer ends with a "Sources" line linking to the site page and to the official text (Fedlex, Federal Office of Public Health). The server strips any link that is neither a page of the site nor a source cited in a guide. When the guides do not cover the question, the assistant says so and points to the closest page or official source.
Personal advice declined
Asked "which insurer is cheapest?" or "which one do you recommend?", it ranks no insurer, recommends none and computes no individual premium. It explains the criteria that matter (deductible, model, cover), recalls that basic benefits are the same everywhere and points to the official premium calculator. A fixed sentence, appended by the server rather than the model, closes every answer: general information, not personal advice.
No health data
The assistant never asks about health. If a visitor mentions an illness, it does not repeat it, says it does not need that information and gives the general rule. Health data is sensitive personal data (art. 5 let. c FADP): the safest option is to keep it out of the pipeline.
No stored conversations
The server does not keep conversations and links them to no request. Logs hold only content-free measures: response time, length, whether a call-back was offered. The conversation stays in the browser tab and disappears when it is closed. The model runs with a provider hosted in Europe, with a content filter that blocks attempts to subvert it.
Hand-over to the form, consent outside the chat
When the visitor shows an intention (switching insurer, needing help for a business, moving to Switzerland), the assistant offers once to arrange a call back from an adviser at a partner broker. It asks for no name, phone number or e-mail. The button opens the site's call-back form, prefilled only with criteria given in the conversation (area, postcode, year of birth, household size), each checked by the server. The visitor reviews everything and gives consent in the form. A "yes" typed in the chat passes nothing on.
Turning the rules into architecture
- Instructions and data kept apart. Behaviour rules live in a versioned file; guide excerpts are added to each question with their address and sources.
- Server-side checks. Filtered links, prefill criteria validated one by one, closing notice appended after the answer: what matters does not depend on the model's goodwill.
- Usage limits. Bot check on the first message, session token, messages per conversation and a daily cap.
- Content-free analytics. Opening the assistant, number of messages, moving on to the form: nothing the visitor writes.
Tests to pass before going live
A set of about twenty reference questions is replayed whenever the instructions change: cancellation deadlines, cross-border questions, ranking requests, disguised advice, a mentioned illness, a phone number typed in the chat, an off-topic request. A single answer that recommends a contract, invents a figure or repeats health data blocks the release. The set also serves as evidence the day someone asks how the assistant was checked.
Frequently asked questions
Can the assistant at least point to a product? It can explain the selection criteria and point to official tools. The choice stays with the user or a licensed professional.
Must the model be hosted in Switzerland? Not necessarily. You need to know where the data goes, what the provider does with it, and be able to document it. For a site that receives health-related questions, hosting in Switzerland or Europe makes the discussion easier.
Is a disclaimer enough? No. It informs the user; it replaces neither server-side checks nor tested refusals.
Going further
houle.ai's AI consulting helps frame this kind of project: scope, choice of model and hosting, test sets. Contact houle.ai to discuss it.