Key takeaways in 30 seconds
- A classic chatbot follows a script: buttons, decision tree, pre-written answers. It does not understand, it routes.
- A conversational AI agent understands free-form phrasing and can execute an action: book, look up an order, move an appointment.
- What the customer sees: the chatbot makes you click "Other request"; the agent answers the request.
- Chatbots remain valid for very short, tightly framed journeys. Beyond that they create more frustration than they save time.
Three generations, routinely confused
The button chatbot. You click "Hours", "Prices", "Contact us". It is a menu dressed as a conversation.
The keyword chatbot. It scans your sentence for words and fires the matching answer. "I'd like to cancel Thursday's booking" triggers the cancellation reply, but it does not know which booking.
The conversational AI agent. It interprets the whole sentence, keeps context across messages, queries your real data and performs the action.
The one question that separates them
Ask a compound question: "Are you open Saturday and do you deliver to Lyon?"
- The button chatbot cannot receive it.
- The keyword chatbot catches one topic and ignores the other.
- The AI agent answers both, in order.
That is exactly how real customers write, because nobody asks one thing at a time.
What it changes commercially
The cost of a badly placed chatbot never shows on an invoice — it shows in customers who give up. After three clicks without an answer, a visitor does not complain. They close the tab and call your competitor.
When a simple chatbot is enough
- fewer than five recurring, purely factual questions;
- your goal is to filter before a human contact, not to resolve;
- low volume and a tight budget.
When you need an AI agent
- customers write compound requests with dates, names, quantities;
- the answer depends on your data: availability, stock, history;
- you receive messages on several channels and want equal quality on all;
- you lose enquiries at night and at weekends.
The deciding criterion: inform or finish
Ask one question: do you want the system to inform, or to finish the job?
Informing displays an answer. Finishing writes a row in your database. The second needs an agent connected to your tools, not a chat widget.