Explained
What AI agents are and what they actually do in business
A chat writes text, an agent does the work. The difference looks small, but it decides whether this brings a benefit or just another piece of software nobody uses.
Over the past year the term "AI agent" has been attached to so many different things that it has almost stopped meaning anything. So let us start with the distinction that actually matters.
The difference between a chat and an agent
When you type into ChatGPT and get an answer, that is a chat. You ask, the model answers, and it ends there.
An agent differs in three ways:
- It starts working on its own, without you asking. When an email lands in the inbox, for example.
- It has tools, not just words. It can read the inbox, write a row into a spreadsheet, create a calendar event.
- It performs several steps in sequence and decides for itself which to do next.
In short: a chat writes text, an agent does the work.
What an agent does well in business
A precise answer beats a general one, so here it is concretely.
| Job | Why it fits |
|---|---|
| Sorting incoming email | Telling a buyer from a supplier and from an advert is recognition, not decision-making |
| Pulling data out of free text | Name, date, phone and rough amount from a message into a structure |
| Drafting a reply from known facts | When the answer is already in the price list, it only needs finding and phrasing |
| Watching for what has stalled | An enquiry with no reply, a proposal with no reaction, an unpaid deposit |
What an agent does badly
This part matters more, because almost nobody writes about it.
Facts nobody gave it
A model never says "I do not know" unless you teach it to. Asked about a price that is not in its knowledge base, it will offer a plausible number. To the client that reads as your promise. So an agent has to be tied to your price list rather than to its own memory, and it needs a clear instruction: what is not in the facts does not get said.
Decisions that carry responsibility
A discount, an exception, a quality dispute, a cancellation with losses. Those are questions of risk, not of language, and they have to stay with a person.
Work with no clear end
If the task sounds like "take care of the clients", the agent has no way of knowing when it is finished. Only work with a start and an end works.
A practical test. Before handing work to an agent, ask yourself: could I hand the same work to a new employee on their first day, giving them only the price list and ten rules? If yes, the agent will manage too. If not, neither will it.
Will the client be able to tell software wrote it
They will, if the text is generic. They will not, if the reply comes from your price list, in your tone and with your terms.
There is also a legal side. Since August 2026 the EU AI Act requires that a person knows they are dealing with artificial intelligence. That means where a reply is automatic, it has to be said. In practice one line at the end of the email is enough.
Experience shows that saying it openly does no harm. What harms is not that software replied, but that the reply was empty.
Where to start
Not with an agent. With the one topic that repeats most often. Usually that is availability, prices, or what is included. Write out the answers to those questions the way you would answer them yourself, and only then let the software use them.
For the first month nothing should go out without your approval. Automatic sending is worth switching on only for topics whose answers you have seen come out correct twenty times in a row.
If you want to see this with your own numbers
Give it thirty minutes and we will walk through one real enquiry of yours from start to finish. After the call you get a map of where data is retyped by hand at your company. If there is nothing worth automating, I will say so.
Let us look at your workflow