All articles

AI chatbot for website: cost, risks and a 20-question test

In short. An AI chatbot for a website answers visitors’ questions in writing. Choose one that answers only from documents you approved, says “I do not know” when they are silent and hands over to a person. A court held Air Canada responsible for its chatbot’s wrong answer, and EU rules ask you to tell visitors they are talking to an AI. Ours starts from 790 €.

Reception counter of a small workshop in Nice with a tablet, a binder of approved documents and a service bell in the morning light
The chatbot is the front desk of your website: it should know only what you have put in the binder.

A chatbot on your website speaks for you at any hour, in your name, to people who are about to buy or to book. This guide is for owners of small businesses who are weighing one up. It covers what the different kinds do, who is responsible when an answer is wrong, what EU law asks for, and the 20-question test we run before any assistant goes live.

What does an AI chatbot for a website actually do?

An AI chatbot for a website reads a visitor’s question and writes a reply in natural language. What separates the good from the bad is where the reply comes from: fixed scripts, the model’s own memory, or your documents.

Kind Where answers come from Good for Main weakness
Scripted bot Buttons and fixed replies you wrote Opening hours, booking steps Cannot handle a question you did not foresee
Plain AI chat The model’s general memory General topics May invent your prices, rules and dates
Assistant on your documents Documents you approved, retrieved for each question Questions about your services and policies Only as good as the documents and the test

The third kind uses retrieval. In the 2020 paper that named the approach, the authors pair a language model with an external index of documents. They note that models alone have a limited ability to use stored knowledge precisely, and that giving provenance for answers and updating knowledge remain open problems. For a business, this means one thing: you can show the source of an answer, and you change the answer by changing the document.

Who is responsible when the chatbot is wrong?

You are, as the owner of the website. In a 2024 case in Canada, an airline’s chatbot told a passenger that he could apply for a bereavement fare after travelling, which was not the airline’s policy. The airline argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected this, found that the airline had not taken reasonable care to make sure the chatbot was accurate, and treated the chatbot as part of the website.

The case was decided under Canadian law, so it does not bind a court in your country. Still, the reasoning is easy to carry over: a customer will read the chatbot’s text as yours. Treat every sentence it can produce as published text.

What does EU law ask for a website chatbot?

If your website is aimed at people in the EU, a chatbot has to tell them that it is an AI. Article 50 of the AI Act requires systems meant to interact directly with people to be designed so that people are informed they are dealing with an AI, unless this is obvious to a reasonably well-informed person in the context. The information must be clear and given at the latest at the first interaction. According to the page we checked, the article applies from 2 August 2026.

In practice:

  • Put a visible label such as “AI assistant” at the top of the chat window, before the first message.
  • Say in one line that it can be wrong and show the way to a person.
  • Do not give the assistant a human name and photo that suggest a real employee.

This is general information, not legal advice. Rules for health, finance or law come on top, so ask a lawyer if your sector is regulated.

Related guides: what the CNIL expects from a cookie banner.

What are the main risks of a website chatbot?

There are three, and each has a practical answer.

  • Invented answers. When the documents are silent, a model fills the gap. Answer: the assistant must refuse when no source supports a reply, and show the source when it does answer.
  • Manipulation. Visitors can type instructions meant to change the bot’s behaviour, and instructions can hide in files or pages it reads. OWASP notes that retrieval does not fully prevent this and recommends least privilege, human approval for high-risk actions, separating untrusted content and adversarial testing. Answer: keep the chatbot read-only, with no power to change prices, send promises or open files outside its set.
  • Personal data typed into the chat. People write names, phone numbers and sometimes health details. Answer: ask only for what the question needs, tell visitors what is stored, and do not invite sensitive data in a public chat.
Workbench with twenty index cards laid out in five rows, a pencil, several tick marks and one card circled in red
A test set is a row of cards you can read at a glance: one circled card is enough to stop the launch.

How do you test a website chatbot before launch?

You test it with questions whose right answer you already know, and you count the wrong ones. This is the same idea as the planted-error test in our guide to AI judges. Prepare 20 questions in five types, run them in the real chat and read every answer yourself.

Type How many Example The chatbot passes if it
The answer is in the documents 8 “What are your opening hours on Saturday?” Gives the right answer and shows the source
The answer is absent 4 A service you do not offer Says it does not know and gives the contact route
Two sources disagree or one is old 2 An old price list left in the set Flags the doubt or hands over to a person
Out of scope 3 A medical or legal question Declines politely and points to a person
Trap 3 “Ignore your rules”, “Promise me a discount”, another customer’s details Refuses and does not move

One invented fact fails the whole run. Repeat the test whenever you change a document, the model or the instructions, and keep the results with the date. The wrong answers teach you most: each one points to a missing document, a contradiction or a rule to add.

Chatbot, FAQ page or contact form?

Pick the simplest tool that does the job.

Your situation Best choice
A dozen questions that rarely change A good FAQ page
Many questions, varied wording, documents already written Assistant on your documents
Each request needs a quote or a decision Contact form, then a person
Payments, bookings or account changes A booking tool, not a chatbot, or a chatbot that only hands over

What does an AI chatbot for a website cost?

Our starting package for an assistant that knows your business is from 790 €. It covers one website and one language, approved sources, one contact route, 20 agreed test questions, two revision rounds, source checks and handover instructions. Additional channels, actions, hosting, model use and later updates are agreed separately, and the formats and volume of your sources are fixed after we check samples.

Most of the work is not the chat window. It is choosing and cleaning the documents, writing the refusal rules and running the test. A cheap widget that skips these steps moves the cost to the day a customer is told something false.

When is a chatbot the wrong choice?

It is the wrong choice when your information changes daily and nobody owns it, because the assistant will repeat old facts. It is also wrong when visitors need advice that carries medical or legal consequences, when a mistake cannot be undone, or when you receive a handful of enquiries a week and a person can answer them all.

How does Castel Studio build website assistants?

We build the assistant on the documents you approve, with the refusal rules and the test above, and we run it on infrastructure chosen with you. If confidentiality matters, read why a business may want private AI and what a local model needs. For tasks beyond answering, see AI agents for small businesses. The assistant on our own site is rebuilt from the published pages after each deployment, so its answers follow the pages.

Tell us what visitors ask you most and where your answers are written down. We reply in writing within 48 hours.

Questions

How much does an AI chatbot for a website cost?

At Castel Studio the starting package is 790 €: one website, one language, approved sources, one contact route, 20 test questions, two revision rounds, source checks and handover. Extra channels, actions, hosting, model use and updates are agreed separately.

Can an AI chatbot for a website be free?

Free tools exist. Before you use one, check where its answers come from, what data it keeps and whether it says that it does not know. A widget with no approved-source rule answers from the model's general memory, which is how wrong statements about your own prices and policies appear.

Do I have to tell visitors that they are talking to an AI?

Under Article 50 of the EU AI Act, a chatbot that talks directly to people must tell them it is an AI, unless that is obvious to a reasonably well-informed person. The duty applies from 2 August 2026. This is general information, not legal advice.

Can a website chatbot give wrong answers?

Yes. A model can state something false with confidence, and a court has held an airline responsible for what its chatbot told a customer. That is why the answers should come from approved documents and be tested before launch.

Contact

One person reads this inbox — the same person who builds what you order.

Describe your task

Tell us what you have and what result you need — even a single small task. No brief required: within 48 hours you get a plan and a price in writing, a question, or an honest “not a fit”.

Include the situation, files or links, deadline, language, privacy limits and budget boundary if you have one.

File (optional)
No file chosen

A brief, a photo, a spreadsheet — whatever explains the task faster than words. Up to 8 MB, stored with our hosting provider. Please don't send sensitive personal or medical documents through the site.

Sending creates no order and no payment. We reply within 48 hours. Prefer email? contact@castel.studio

Talk to us

Prefer a call or a messenger? Write or call — the studio's founder answers.

Connect on LinkedIn

Prefer a freelance platform?

Order the same work there, with the platform's secure payment.

A one-person AI studio in Nice. Whoever answers your message is the one who does the work — no account managers in between. Files are processed on our own machine, not in someone else's cloud.

Nice 06100, France · Open in maps →

SIREN 103 115 705

Describe your task