AI agents and chatbots are not the same thing
Confusing chatbots with agents costs money: companies buy a chatbot expecting work to get done, or build an agent where answering questions would have been enough. They are two different categories of software.
- Published
- Updated
What exactly is the difference?
A chatbot is a conversational interface that produces text answers; an AI agent is a system that takes a goal, plans steps, executes actions through tools connected to real systems, and verifies the outcome.
The distinction is not the language model — both can use the same one. The distinction is permission to act. A chatbot is done when it has produced text. An agent is done when the state of the world has changed: a ticket closed, an invoice booked, an order dispatched.
| Dimension | Chatbot | AI agent |
|---|---|---|
| Output | A message | A completed task |
| Tools | None, or text search | APIs, databases, internal systems |
| State | Conversation context | Task state across sessions |
| Error handling | Repeats or asks you to rephrase | Retries, changes strategy, or escalates to a person |
| Auditability | Message history | A log of every action, with input and output |
| Typical risk | Saying something wrong | Doing something wrong |
When is a chatbot enough?
A chatbot is enough when the value is in the answer rather than the execution: questions about documentation, internal policy, catalogues or FAQs.
If 80% of requests are resolved by saying something correct and the remaining 20% go to a person, a chatbot wired to your document base is cheaper to build, easier to audit and faster to ship than an agent.
When do you actually need an agent?
You need an agent when the work consists of changing data or triggering processes in other systems, and the bottleneck is the time a person spends moving information between screens.
- Classify an incoming email, extract the fields and write them into the CRM.
- Read an invoice, check it against the purchase order and flag the difference.
- Pull information from three systems and assemble a case file before a person picks it up.
- Close tickets whose resolution is a concrete, repeatable action.
A chatbot does not help here, because the user does not want an explanation of how to enter the data: they want the data entered.
What risks does an agent add?
An agent adds the risk of action: an error stops being a wrong answer and becomes a modified record, a sent email or an initiated payment.
- Explicit action limits: what it can do, on which records, up to what amount.
- Mandatory human confirmation for irreversible operations.
- A full log of every tool call, with input, output and the decision attached.
- Evaluations that run on every change to prompt, model or tools.
- A rollback plan for actions that can be undone.
How do cost and timeline change?
An agent costs more than a chatbot because the work is not in the conversation but in the integrations, the permissions and the testing.
As a reference from our own projects: a documentation assistant reaches production in weeks, and most of the effort goes into organising the content. An agent operating across two to five systems needs 4 to 8 weeks to reach limited production, and half of that time goes to integrations and the evaluation suite, not the prompt.
How do you decide in five minutes?
One question settles it: if the system answered perfectly every time, would the work be done, or would someone still have to go and do it?
If the work is done, you need a chatbot. If someone still has to execute it, you need an agent. And if it is not clear what the work is, you need neither yet: you need to describe the process.