This is almost everyone's way into AI, and the source of the most widespread confusion: believing a chatbot and an agent are the same thing in different clothes. They are not, and the difference decides what you can entrust to them.
Definition
A chatbot is a system that answers messages in a conversation. You ask a question, it produces an answer. Then it waits for the next one.
What defines it is not what it can do but what it does not do: it does not decide the steps, it does not use tools on its own, it does not act outside the conversation.
The line with an AI agent comes down to one word. A chatbot answers, an agent acts. As long as the output is text you read before doing something with it, you are on the chatbot side, whatever name the vendor chose.
Three generations all called chatbot
| Generation | How it works | Limit |
|---|---|---|
| Scripted | A tree of questions and buttons | Anything off the tree is lost |
| Intent-based | Recognises a request from a learned list | Every intent has to be anticipated |
| Model-based | Understands and writes freely | Can invent without supplied documentation |
The third generation has made the first two obsolete for most uses, and it is what sits behind modern Automated customer support. It brings considerable flexibility and a new risk: without a document base or a fallback instruction, it will answer everything, including what it does not know.
What a chatbot does well
Answering recurring questions. Its natural ground, provided it draws on your documents through a RAG system.
Acting as an interface. Asking a question in plain words rather than navigating six menus is a real gain, internally too.
Triaging before a human. Understanding the request, gathering context, then handing over. The chatbot does not solve everything, it prepares.
Staying under control. Since it only produces text, the risk is limited to what is written. That is what makes it the safest way in.
The most common trap is not technical. It is putting a chatbot in front of an unhappy customer. A perfect answer from a machine to someone who wants to be heard makes things worse. Always provide a visible route to a person.
When to move on
The sign is clear: if you find yourself copying the chatbot's answer into another tool to execute it, you need an agent or a Workflow.
A workflow fits if you know the steps in advance, which is the most frequent and cheapest case. An agent becomes necessary when the steps depend on what you find along the way.
Frequently asked questions
Can a chatbot know my products?
Not on its own. You have to supply your documents, either in the conversation for a one-off, or through RAG as soon as the volume exceeds a few pages.
Must you say it is a machine?
Yes, it is a transparency requirement carried by the AI Act for this type of use. It also avoids the feeling of deception when the other person works it out, which always happens.
How long to get one live?
A few hours for a first version connected to your standard replies. The real time goes elsewhere: tidying the documentation, and reading back each week what it answered to spot what it handles badly.
How do you build one that holds up?
By starting with your twenty most frequent questions rather than the ambition to cover everything. Our n8n course builds that system end to end, with traceability of the extracts used and the route out to a human.