A chatbot answers. An AI agent acts. Both use a language model, and the difference is whether the software can call functions that change something: raise the invoice, send the message, create the task. A chatbot hands the work back to a person, which is why installing one often changed nothing.
Plenty of businesses added a chat widget, watched it answer questions competently, and found that the amount of work they had to do stayed exactly the same.
That is not a failure of the chatbot. It is what a chatbot is.
| Chatbot | AI agent | |
|---|---|---|
| Produces | An answer | An answer, and a change |
| Reaches | Its own knowledge and the conversation | The business records and systems it is wired to |
| After it responds | A person does the work | The work is done, or waiting for approval |
| Runs when | Somebody opens the chat | On a schedule, a trigger, or a conversation |
| Fails by | Saying something wrong | Doing something wrong |
| Needs | Content to answer from | Permissions, guardrails and a record of what it did |
The line has blurred: many chat products now call tools, which makes them agents by this definition for those specific actions. The question to ask a vendor is not which word they use, it is which functions the thing may call.
A chatbot that is wrong tells a customer something untrue. That is bad, and it is recoverable, and the customer usually knows to double check.
An agent that is wrong has already done something. The invoice is raised, the message is sent, the record is changed. Recovery means undoing, and some things do not undo.
This is why the useful question about any agent is not how capable it is. It is what it does without asking, because that is the exact set of mistakes you will find out about late.
It often is, and pretending otherwise sells the wrong thing.
A business whose problem is answering the same six questions forty times a week does not need an agent. It needs those six answers somewhere a customer can find them.
A business whose problem is that nobody chased last month invoices has a different problem, and no amount of answering will touch it.
Support drafts answers to customer questions, which is chatbot-shaped work, and the drafts wait for approval before any customer sees them.
Sales, Ops and Finance are agent-shaped: they read the business records, decide something, and either produce a task, produce a draft, or in the case of a stablecoin invoice, close the loop entirely by settling it against the chain.
Both shapes are in the product because both problems are real. What we do not do is call the first one an agent.
The work a service business loses money on is the work nobody got round to. That is what an agent is for.
Last reviewed .