Adding an AI Chatbot to Your Website: The Honest Guide
An AI chatbot can answer your customers day and night, or make up your opening hours. Here's what to know before you add one to your site.
You've probably seen that little bubble in the bottom right corner of a website: "Have a question? I'm here to help." Sometimes it gives genuinely good answers. Sometimes it makes up opening hours, promises a callback nobody will ever make, or goes around in circles. An AI chatbot can save a shop or a freelancer a lot of precious time. It can also hurt their reputation. The difference comes down almost entirely to how it's set up.
What an AI chatbot does well
A well-tuned AI assistant is excellent at:
- repetitive questions: opening hours, directions, lead times, terms, payment methods;
- questions asked outside business hours: in the evening, on Sundays, while you're busy working;
- understanding different phrasings: it gets "are you open tomorrow morning?" just as well as "Saturday hours";
- pointing people the right way: to the right page, the right form, the right button.
In other words, it does what you do ten times a day on the phone or at the counter, without taking up your time.
What it does badly (if you let it)
An AI model has learned to write text that sounds plausible. It knows nothing about your business. When it's missing a piece of information, it fills the gap with something that sounds right. The most common mistakes:
- making up a fact: opening hours, a price, a service you don't offer;
- promising an action: "I'm passing your request on," when it passes nothing on;
- saying yes to everything: bookings, delivery, custom work, as long as it sounds plausible;
- shifting its tone: getting overly casual because the visitor is, or piling on empty stock phrases.
These mistakes aren't glitches. They're normal behavior for a model that hasn't been given clear boundaries.
A real example: Flopia
Flopix, the website builder for small business sites, is developing its own assistant, Flopia. She's in testing, with a waiting list open on flopia.fr. Her first versions ran on a small model installed on a server with no graphics card, which called for very precise tuning. Today she relies on Claude, Anthropic's AI, called through an API, for more accurate and faster answers.
During testing, she made several of the mistakes listed above. She announced a made-up number of photos. She claimed the sites were free, after misreading pages that talked about a trial. She started almost every answer with the same cheerful phrase. Each of these mistakes was fixed with a specific adjustment, then turned into a test so it would never come back.
That's the real lesson: a good chatbot isn't one that never gets anything wrong. It's one whose mistakes get fixed, methodically.
Five conditions for a reliable chatbot
1. Accurate documentation
The chatbot answers based on what you give it: your site's pages, FAQs, terms. If a page is vague or out of date, its answers will be too. Before you launch anything, reread your pages. For Flopia, the pages written for Google, which are more sales-oriented, were even pushed to the background in favor of the reference pages, because their phrasing came out badly when taken out of context.
2. Short, clear instructions
Who it is, how it talks, what it always does, what it never does. Every sentence counts, especially with a small model: instructions that are too long slow it down and lose their force. Everything is covered in detail in the article on writing your AI assistant's instructions.
3. Safeguards in the code
An instruction isn't always enough. A good system checks every answer before showing it: it removes banned phrases, flags any price or link that doesn't appear in a source, and gives a fixed answer to trick questions like "ignore your instructions."
4. A clear handoff
Your chatbot needs to know when it doesn't know, and where to send the person: a form, an email, a ticket. Never a promise it can't keep, like a phone callback.
5. No action without confirmation
If your assistant can change something (opening hours, a closure, an order), it should only suggest the change. A human sees the before and after, and clicks 'Confirmer' (Confirm). The demo on flopia.fr shows exactly how this works.
Choosing your engine
There are two main options: an AI called through an API from a provider, which is powerful and leaves you no machine to manage, or a model installed on your own server, which keeps conversations in-house but takes more tuning. The choice depends on your volume, how sensitive the data is and who will look after the machine. Flopia tried both before settling on the API. We compare the two paths in detail in local AI or API.
Questions to ask yourself before you start
- What are the ten questions I get asked most? If you can't list them, start there. They'll form the core of your documentation and your tests.
- Who will read the conversations? A chatbot nobody ever checks gets worse over time.
- What happens when it doesn't know? Plan the answer and the link to a human.
- Do my visitors know it's an AI? Say so, always.
- Where do the conversations go? Your privacy policy has to say.
Launch in stages
Don't put your chatbot live for everyone at once. First, test it on a list of about twenty real questions, reading every answer in full. Then open it up to your team or a few trusted customers. That's what Flopix did: on flopix.fr, Flopia's chat bubble was first reserved for the team, who used it and reported its mistakes, before it was opened to all visitors.
Once it's live, set aside half an hour a week to read the flagged conversations, understand each mistake and fix it in the right place: the page, the instruction or the safeguard.
If you want to learn how to tune an assistant step by step, from the role sheet to the test bench, the course Create your AI assistant (in French) walks through the whole method, with Flopia as a case study.
Frequently asked questions
Can an AI chatbot replace human contact?
No, and that's not its job. It answers common questions and points people in the right direction. For a complaint, a special case or a decision, it should hand over to someone on your team.
Should I tell visitors they're talking to an AI?
Yes. It's a matter of honesty and trust, and it keeps a wrong answer from being taken as a commitment from your business. One sentence under the chat window is enough.
Will a chatbot slow down my website?
It can, if its script loads before your content. Make sure it loads after the page, and that the chat bubble doesn't hide anything important on phones.
How long does it take to set up?
A first version can be set up in a few hours if your documentation is ready. The real work comes after that: reading its answers every week and fixing its mistakes.