AI assistants

How to Stop Your Website AI Assistant from Making Up Answers to Customers

A chatbot that quotes wrong prices or incorrect opening hours does more damage than a simple contact form. Here is how to frame an AI assistant with your own content so it answers accurately, without hallucinating.

FlowBiz.ai11 November 2026 5 min read

Key takeaways

  • Unguarded AI invents: prices, hours, availability—everything becomes approximate
  • The framework is built before the chatbot: knowledge base, business rules, blocking thresholds
  • A useful assistant does not replace humans: it qualifies the request and hands over at the right moment
Contents
  1. Why raw AI fabricates answers that sound true
  2. Building your knowledge base: the documents that serve as guardrails
  3. Business rules: when the AI must stay silent and hand over
  4. Concrete example: the chatbot of a hypothetical property management firm
  5. What AI does poorly, and what you should not automate
  6. Where to start without getting it wrong

You have seen this scenario before. A customer lands on your site at 10 p.m., asks about pricing, and the chatbot quotes a rate that has not existed for two years. The next day, your employee spends half a day clearing up the confusion. The worst part? That customer might never have called at all had they not been reassured by an instant answer—even a false one.

The shop owners and firm partners we meet in Sion share the same concern. They want an AI assistant on their website, but not at the cost of invented answers. The challenge is not technical: it is defining a clear perimeter, with your own content, before the chatbot speaks to anyone.

Why raw AI fabricates answers that sound true

Generalist language models are trained on billions of texts. They excel at producing fluent sentences, not at verifying facts. Imagine a carpentry business in Sierre that installs a standard chatbot. A visitor asks: "Do you do aluminum windows?" The model says yes, because it is probable, even though the company specialized in wood five years ago.

This phenomenon is called hallucination. It does not disappear with a better model. It disappears when you remove the AI's ability to guess. The method is to lock it inside a knowledge base that you control, and formally forbid it to fill in the gaps.

Building your knowledge base: the documents that serve as guardrails

The foundation of a reliable assistant is a set of texts that you validate yourself. Not your entire website, which may contain outdated pages. A targeted selection:

Document typeExample for a fiduciary / accounting firmExample for a retail business
Pricing and packages2024 fee schedule, specific surchargesIndicative prices by range, delivery conditions
Hours and availabilityTelephone duty days, appointment lead timesOpening hours, click-and-collect slots
Intake proceduresDocuments to bring, standard processing timesCustom order steps, manufacturing lead times
Validated FAQsAnswers drafted by the owner, not collected at randomReturn policy, warranty, maintenance

Each document carries a last-revised date. The assistant only refers to it if the question matches a relevant excerpt. Otherwise, it hands over. This blocking threshold is what separates a gadget from a reliable tool.

Business rules: when the AI must stay silent and hand over

A useful AI customer service assistant is not the one that answers everything. It is the one that knows its own limits. Typical case: a fiduciary in Martigny receives a query about child care expense deductions. The assistant finds the general rule in its base, but detects that the case mentions cross-border work. It blocks immediately: "For this situation, Ms. Dupont will call you back within 24 hours."

Business rules translate into explicit instructions. Never give personalized tax advice. Never confirm availability without checking the connected calendar. Never quote a precise price for a complex estimate—only a range or a base package. These barriers are coded, not left to the model's discretion.

To go further on the automations that complement this setup, see our automation examples designed for French-speaking Swiss SMEs.

Concrete example: the chatbot of a hypothetical property management firm

Imagine a Sion-based agency managing 120 apartments. Its AI assistant is fed by: the list of available properties (updated weekly), tenant selection criteria (no online credit score), viewing hours, garage and parking space rates. A visitor asks: "I have a dog, is that allowed?" The excerpt found states: "Pets subject to written landlord approval, category dogs prohibited." The assistant quotes this text, without adding that "it should be fine" or "it is often refused."

Another question arrives: "My rent is CHF 1,850, when is the next indexation?" The base does not contain the specific contract. The assistant replies: "I cannot access your rental file. Contact Mr. Müller at 027 xxx xx xx or leave your details." No approximate calculation, no invented date. The visitor is not satisfied, but they are not misled.

This type of scenario requires upstream structuring work. Our 2-minute AI quiz lets you assess whether your situation is suitable.

What AI does poorly, and what you should not automate

Let us be direct. A poorly configured small business website chatbot damages your credibility. Here are the cases where full automation is counterproductive:

  • Price negotiations. Even a range can be interpreted as a commitment. A human must take over.
  • Disputed situations. A dissatisfied customer describing a complex problem deserves listening, not a standardized response.
  • Sensitive personal data. The assistant may collect a phone number for a callback. Not an AHV number, not marital status, not income amount. These thresholds must be blocked technically, not merely discouraged.
  • Rapidly shifting regulated topics. If your sector faces frequent changes (think labor law adaptations or the 8.1% VAT rate), the knowledge base ages quickly. You need a revision process, not neglect.

Transparency also matters under Swiss legislation. The Swiss nFADP (nLPD) framework requires documenting automated processing. For precise details, consult the Federal Data Protection and Information Commissioner (FDPIC) or your legal advisor.

Where to start without getting it wrong

  1. Audit your existing content. Which pages on your site are up to date? Which internal documents are authoritative? Discard or archive the rest before letting an AI read them.
  2. List your ten most frequent questions. Not the ones you imagine—the ones your employee on the phone actually notes. Draft the answers yourself, one sentence per idea.
  3. Define three blocking thresholds. From what point does the assistant refuse to answer and offer human contact? Custom pricing, unconfirmed specific deadline, sensitive legal situation?
  4. Test with scenarios from your trade. Have colleagues or friends play the difficult customer. Note every approximate or invented response. Fix the knowledge base, not just the prompt.
  5. Plan a monthly review for the first three months. Real usage always diverges from predictions. A frozen assistant quickly becomes useless or dangerous.

At FlowBiz.ai, we build AI assistants for French-speaking Swiss SMEs, hosted in Switzerland, delivered in stages, with your content and your rules. If this concerns you, contact us to discuss it concretely.

Frequently asked questions

My site already has many pages, can I feed them all to the assistant?+

No, that is risky. Outdated or contradictory pages fuel hallucinations. It is better to select validated, dated documents and revise them regularly. The assistant is not a search engine: you must give it a single source of truth per topic.

Can the assistant book appointments on its own?+

Only if you connect its calendar with strict rules. Otherwise, it risks offering slots that are already taken or ignoring your constraints. The safest approach is to qualify the request, then propose a verified slot or a callback.

Do I need to disclose that an AI is responding?+

Yes, transparency is recommended by the FDPIC and reassures visitors. A discreet notice suffices: "I am a digital assistant, my responses are based on the official information of [your company]." This also limits unrealistic expectations.

What if the assistant still gives a wrong answer?+

Build a direct reporting mechanism into the chat, and analyze every case. Fix the knowledge base, not just the message. An isolated wrong answer is a symptom: the cause lies in missing documents or rules.

And in your business, what could we automate?

Answer a few questions and get a free, personalised AI pre-diagnosis with three concrete leads in two minutes.

Share this article

How to Stop Your Website AI Assistant from Making Up Answers to Customers | FlowBiz.ai