The short answer
To add an AI assistant to your website, give it a written knowledge base about your business, decide which actions it may take (like checking availability), require a human click for anything that changes data (like booking), and keep a fallback for when the AI is unavailable. You can keep your own chat design and connect it to an AI service behind the scenes.
Start with what the assistant should know
An assistant is only as good as what it has been told. Before choosing any tool, write a plain document describing your business: what you do, who it is for, what you have built, how you work, how people reach you — and what the assistant should not claim, such as prices you do not publish.
This document does double duty. It is the assistant's knowledge, and it forces you to write down answers your sales conversations repeat every week.
Worth knowing
Add a short "never say" list: no invented prices, timelines, client names or guarantees. A good assistant says "I don't know, here is how to reach a person" rather than guessing.
Decide what it is allowed to do
Answering questions is the easy part. The value — and the risk — is in actions. Split them into two groups before you build anything:
- Reads — looking things up: open meeting slots, order status, product details. Safe to let the AI run on its own.
- Writes — anything that changes something: booking, cancelling, submitting, paying. These should never happen on the AI's word alone.
Booking a call safely: let the AI read, let the person confirm
Take booking a discovery call. The AI is good at the conversation: understanding "can we talk Friday afternoon?" and finding open times. It should not be the one to place the booking.
The pattern we use on this site: the assistant checks live availability from the calendar and shows the open times as buttons. The visitor taps a time, enters their name and email, and presses confirm. The booking happens on that click, and the calendar sends the invite. The AI never books anything by itself, and it cannot offer a time the calendar did not return.
Worth knowing
Make the visible buttons come from the calendar's data, not from the AI's text. Even if the AI phrases something wrongly, the visitor can only pick a real slot.
Keep your own design
Off-the-shelf chat widgets look like off-the-shelf chat widgets. If your brand matters, you can keep your own chat interface and connect it to an AI service behind the scenes: your UI sends the visitor's message, streams the answer back into your own bubbles, and runs any tools the AI asks for.
That is how Lio, the assistant in the corner of this page, works — our own interface, with answers and tool calls powered by AskVerb, an embeddable AI agent from our portfolio.
Plan for failure from day one
- If the AI service is unreachable, fall back to simple stored answers and a contact link — never a broken chat
- If the calendar is down, the assistant should say so and offer the booking page or email
- If a slot is taken between showing it and confirming it, show the calendar's error and let the visitor pick again
- Never show "booked" until the calendar has actually returned a booking
Make it findable by AI search too
The same knowledge that powers your assistant helps AI search engines describe you correctly. Publish it as an llms.txt file and as FAQ structured data on your site, so ChatGPT, Claude, Perplexity and Google's AI answers draw on the same facts your assistant uses.
Want one on your product?
We build AI products with exactly these rules — intelligence with intent, never AI for the sake of AI. Try Lio in the corner of this page, then book a 15-minute call if you want something similar in your own product.
“The most important line in an AI assistant's design is the one between what it may look up and what it may change. Get that right and everything else is detail.”
Questions people ask us.
Yes, safely, if the AI only reads availability and the visitor confirms the booking with their own click. The assistant finds open times from your calendar and shows them as buttons; the booking is made only when the visitor submits their details and presses confirm.
Yes. Your own interface can send messages to an AI service, stream the reply into your own design, and run any tools the AI requests. You keep full control of the look while the AI handles the answers.
Give it a written knowledge base, tell it explicitly what not to claim (such as unpublished prices), and make it hand over to a person when it does not know. Show actions from real data, like calendar slots, rather than from the AI's own text.
A well-built assistant falls back to stored answers and a contact or booking link, so visitors never see a broken chat. Plan this fallback before launch.