Crescendo LabCAAC Go

Chatbot vs AI agent: what taking action means on a website

A chatbot answers. An AI agent answers and then does something with the answer: it saves the visitor's details, records a booking request, brings in a person, or sends the details to another system. The difference is not the model behind the chat. It is whether the software is allowed to act, what it is allowed to do, and how carefully it does it.

Definition

A website chatbot replies to visitor messages with text. A website AI agent replies and can take a limited set of actions on the business's behalf, such as saving a lead, recording a booking request, handing the conversation to a person or sending data to another system, within rules the business sets.

The difference in one table

ChatbotAI agent
Answers from your contentYesYes
A visitor leaves their detailsShows a form, at bestSaves the lead from the conversation and notifies you
A visitor asks for a timeLinks to a booking pageRecords the request with the time they asked for
A question it should not answer"Please email us"Hands the thread to a person, with the context attached
Other systemsNoCalls your API or webhook with the conversation's details
What can go wrongA wrong answerA wrong answer, or a wrong action

The last row is why agents need stricter testing than chatbots. An invented price is bad; an invented booking is worse.

Five ways to tell which one you are looking at

  1. Ask it to book. An agent asks for a time and tells you what it saved. A chatbot sends you to a page.
  2. Give one contact channel only. An agent saves with an email or a phone number. Many tools keep asking for the other one and lose the visitor.
  3. Ask something it cannot know, such as "Is Saturday 10am free?". An agent that is not connected to a calendar should say the team will confirm. If it says yes, it is guessing.
  4. Ask for a person. An agent opens the conversation to your team with the history. A chatbot gives you an email address.
  5. Ask what it just did. "Did you save that?" An agent can say what it recorded, because it recorded it.

What "acting" should never mean

An agent acts within rules you set. It should not:

Those four are the things we test every month in the CAAC Go Agent Benchmark, alongside plain answering, in seven languages, across the main AI models.

Where CAAC Go sits

CAAC Go is a website AI agent. It answers from your own content, and it takes four actions:

The Shopify app adds order-status answers. What it does not do: write to calendars by itself, take payments, browse the web, or act outside the actions you turn on. Setup for the actions is in the leads and booking requests guide; the handoff rules are in the human handoff guide.

Common questions

Is an AI agent just a chatbot with a better model?

No. The model decides how well it answers; the actions decide what it can do. The same model can run as a plain chatbot or as an agent. What changes is the set of actions it is given and the rules around them.

Do I need an agent, or is a chatbot enough?

If visitors only need answers, a chatbot is enough. If the point of the conversation is a booking, a quote or a callback, you want an agent, because the answer alone leaves the visitor to find a form.

Is an agent riskier?

It can be, which is why the actions are limited, logged and tested. A booking request that a person confirms is low risk. An agent confirming slots on its own, without a calendar connection, is the risk to avoid.

Can an agent work on WhatsApp or email as well as the website?

Yes. In CAAC Go the same agent, with the same actions, runs on the website widget, a connected WhatsApp number and email.

Try it on your website

Paste your link and ask your own AI agent anything. Free, no signup.

Keep reading

Last updated 2026-11-08