Chatbot or agent: what is the difference, and when do you need which?

Most companies have a chatbot somewhere by now. Fewer have an agent, and the two words are used as if they meant the same thing. They do not, and the difference decides what you can reasonably expect.

The distinction in one line

A chatbot answers. An agent acts.

Ask a chatbot about a customer's outstanding invoices and, at best, it tells you what it has been told. Ask an agent and it looks them up in your finance system, checks them against the payment history, and drafts the reminder. The language model behind them is the same. What differs is whether it has been given tools and permission to use them.

What changes when it can use tools

It takes several steps. An agent decides what to do next based on what the previous step returned. That is what makes it useful on messy tasks, and also what makes it harder to predict.

It touches real systems. Reading is one thing. Writing is another. The moment an agent can create, update or send, the question stops being "is the answer good" and becomes "what happens when it is wrong".

Errors compound. A chatbot that gets one answer wrong wastes a minute. An agent that gets step two wrong builds steps three to seven on top of it.

None of this is an argument against agents. It is an argument for scoping them properly, which is ordinary engineering rather than anything exotic.

Where the line should sit

Useful agents in a company setting almost always have four things:

• Narrow permissions. Access to the specific systems the task needs, at the level the task needs. Read-only wherever reading is enough.
• A stopping rule. Explicit conditions under which it hands back to a person rather than proceeding.
• A record. What it did, in what order, on what basis. Not for compliance theatre, for the Tuesday morning when someone asks why a customer got that email.
• A person on the outcome. Not approving every step, which defeats the point, but accountable for the result.

Which do you actually need?

For questions over your own material, where a person reads the answer and decides what to do, a well-built retrieval chatbot is often the right and much cheaper answer.

For work that is repetitive, rule-bound, spans several systems and currently eats a person's afternoon every week, an agent earns its keep. The test is not how impressive the demo is. It is whether the task has a definition of done that someone can check.

Next step

The honest starting question is not "should we build an agent". It is which recurring task in your organisation has clear rules, clear inputs and a checkable result. If nothing on the list qualifies, the answer is a chatbot, or nothing yet.