Since the arrival of large language models such as Claude, ChatGPT and Gemini, the terms “chatbot” and “AI agent” are often used interchangeably. The difference is concrete, though: it changes what the tool can do for your business, what it costs to set up, and the precautions it requires.
A chatbot answers
A chatbot receives a question and generates a text answer, from what it already knows or from documents it has been given. It is useful for answering frequent questions (opening hours, delivery times, return policy) or pointing a visitor in the right direction, but it cannot do anything more than talk.
A chatbot can also draw on your own documents: with each question, it is given the relevant passages from your FAQ, product sheets or terms of sale. It then answers with your information rather than generalities, but it is still limited to talking.
An agent acts
An AI agent can, on top of answering, use tools: check a product’s real stock, check an order’s status, create a ticket, offer an appointment slot, send an email. Technically, this relies on tool use (function calling):
- the customer asks a question, for example “Where is my order A-1042?”;
- the model understands it needs a piece of information and requests a specific tool, with its parameters;
- the application checks that this action is allowed, then runs it (here, a read from the orders database);
- the result is sent back to the model, which writes the answer, or requests another tool if needed.
Each tool is described to the model with a name, a description and its parameters. For example, in the Claude API format:
{
"name": "order_status",
"description": "Returns the status and carrier of an order from its number",
"input_schema": {
"type": "object",
"properties": { "number": { "type": "string" } },
"required": ["number"]
}
}The key point: the model only requests the action. Your application, written by a developer, decides whether to run it.
Concrete examples
- Online store: an agent on WhatsApp or on the site handles order tracking, checks whether a size is in stock, and passes return requests to a person.
- Service provider: the agent checks the calendar, offers slots and books the appointment.
- Quote requests: the agent asks the right questions, summarizes the need and creates a record in your tool, ready to be handled.
An agent stays under control
A well-designed agent only has access to the tools it is explicitly given, and nothing else. A few basic rules:
- read-only tools whenever possible;
- human approval for sensitive actions (a refund, a significant order change): the agent prepares, the person approves;
- a log of every action requested and executed;
- an easy hand-off to a real person when the agent does not know.
The risks and countermeasures are covered in Integrating AI into your site without putting your data at risk.
Side by side
| Chatbot | AI agent | |
|---|---|---|
| What it does | Answers from its knowledge or your documents | Answers and acts in your tools |
| Setup | Quick: texts and documents to provide | Longer: each tool is built and secured |
| Running cost | Low | Higher: several exchanges with the model per request |
| Gain | Fewer simple questions to handle | Whole tasks handled without intervention |
| Precautions | Accuracy of answers | Tool permissions, human approval, logging |
Where to start
If the need is limited to answering questions, a chatbot is enough and costs less. If you want to save time on concrete, repetitive tasks, start with an agent that has a single tool, the one that handles the most frequent request (often order tracking). Measure for a few weeks how many requests it handles on its own, then add the next tools one by one.
Going further: my expertise: AI agent developer · read also: Claude, ChatGPT, Gemini: how to choose the right model for your project
