Chatbot vs AI agent: what is the difference, and which do you need?

A chatbot answers questions. An AI agent answers questions and also does things: it looks up an order, books an appointment, creates a ticket or updates a record, by calling your systems through tools and deciding which steps to take. The output of a chatbot is text a person reads. The output of an agent can change something in your business. That one difference drives the build cost, the running cost, the risk and the amount of testing each one needs.

Most businesses should start with a chatbot that answers well and add actions one at a time. This post explains the difference in practice, how an agent decides what to do, and how to tell which one your problem needs. It draws on the bots and agents we build in our AI agents and chatbots work; the wider picture of what agents can do today is in AI agents for business.

What does each one actually do?

There are three kinds of bot in use, and the names get mixed up. The table separates them.

The cost figures come from our breakdown of what an AI chatbot costs to build and run. An agent costs two to four times as much because each action is an integration with one of your systems, plus rules about when it may act, plus tests that prove it acts correctly.

How does an agent decide what to do?

An agent is a language model inside a loop that your code runs. On each turn:

1. Your code sends the model the conversation, your instructions and a list of tools. Each tool has a name, a plain description and the inputs it takes, such as lookup_order(order_number) or book_slot(date, time, service).
2. The model decides whether it can reply directly or needs a tool. If it needs one, it returns a request to call that tool with specific inputs.
3. Your code checks the request, runs the tool against your system and sends the result back to the model.
4. The model reads the result and either calls another tool or writes the reply.

The model never touches your database. It asks, and your code decides whether to carry out the request. That is where all the safety lives: which tools exist, what each one is allowed to change, and what your code checks before it runs one. Anthropic's tool use documentation describes the mechanism; OpenAI and Google work the same way under different names.

A chatbot is the same loop without step 3. It retrieves your content, reads it and replies.

Do you need an agent, or a workflow with AI in it?

Here is where we often argue against the bigger build. If the steps are always the same, do not let a model choose them. "When an enquiry arrives, extract the name, phone number and service, create a CRM lead, send a confirmation" is a fixed sequence. Code should run it, with a model doing only the one step that needs reading: pulling the details out of a free-text message. That is cheaper, faster and easier to test than an agent deciding the order of operations every time. Our guide to AI workflow automation covers that pattern.

An agent earns its place when the path depends on the conversation. A customer who asks "can I move my Thursday appointment to next week, and is Dr Rao in then?" needs the bot to check a calendar, then a staff rota, then book, in an order that depends on what each step returns. A fixed workflow cannot handle that without becoming a tangle of branches.

A simple test for which one you need:

• The job ends with an answer. You need a chatbot that retrieves your content.
• The job ends with something changed, by the same steps every time. You need a workflow, with AI where something must be read or written.
• The job ends with something changed, by steps that depend on the conversation. You need an agent.

What changes when a bot can act?

Answering wrongly is embarrassing. Acting wrongly costs money. So an agent needs four things a chatbot can mostly skip.

Narrow permissions. Give each tool the least access that does its job. A lookup tool reads one customer's orders, found by the logged-in customer's ID, not any order by number. A booking tool books within opening hours and cannot delete other people's slots. AI agent security goes through permissions in detail.

Confirmation before anything that changes a record. The agent states what it is about to do ("I'll move your appointment to Tuesday 10 November at 10:30, is that right?") and acts only on a yes.

Handover rules. When the agent is unsure, when the customer is upset, or when a request is outside its limits, it passes the conversation to a person with the context attached. Human handover sets out the triggers.

More testing. For a chatbot you check that answers are correct. For an agent you also check that it calls the right tool, with the right inputs, at the right moment, and that it refuses when it should. How to test an AI agent before launch describes the method.

Does an agent cost more to run?

Yes, roughly twice as much per exchange, but from a small base. An agent sends tool definitions with every request, and each tool call is an extra round trip to the model: one call to decide, a second to reply after the result comes back.

Take a support exchange of about 4,000 input tokens and 250 output tokens for a chatbot. The agent version that looks up an order uses about 4,500 input and 100 output tokens to decide, then about 4,800 input and 250 output tokens to reply. At list prices as of October 2026:

At a few thousand conversations a month, both are small next to the build cost and the time someone spends reviewing conversations. Put your own token counts and volumes into the AI API cost calculator to see the monthly figure on each model.

Should you start with a chatbot and add actions later?

For almost every business, yes. Build the bot that answers from your content first. Launch it, read the conversations for a month, and fix the content gaps they reveal. The logs will also tell you which action customers ask for most, which is rarely the one you would have guessed. Add that action next, with its own permissions, confirmation step and tests.

Starting with a full agent means debugging answers and actions at the same time, in front of customers. Starting with a chatbot means the answers are already trusted by the time the bot is allowed to change anything. If you are ready to scope the first version, how to build a customer support AI agent walks through it step by step.