What is an AI agent?
An AI agent does a task that involves reading, assessing or writing, such as assessing an enquiry or drafting a reply. It puts it in front of you, or does it itself if you want it to.
The term is used for all sorts of things, so first what we mean by it: an AI agent does one task, without anyone having to start it each time.
The difference from a chatbot
A chatbot waits. You type a question, it gives an answer, and that is that. Everything that happens starts with you.
An agent does not wait. It starts by itself, because something comes in or because it is a particular time. It does its steps and leaves something ready. You only come into the picture when there is something to judge.
The difference from an ordinary automation
An ordinary automation follows rules: if this comes in, put that in that folder. That works fine as long as everything looks the same.
An agent can do something a rule cannot: understand the content. Work out from a mail which customer it is and what it is about, even when one sender writes "urgent" and the other "could you have a look". That is the part people were needed for until now.
The three things that hold an agent together
A trigger. A mail, a form, a time, a status that changes.
An assignment with sources. What it has to do, in which order, and where it may look. No more than that.
A moment where a person decides. For anything that goes out or touches money, the result is put ready first.
Why that last one always belongs
An agent that sends mails to customers by itself is an agent that one day sends the wrong mail. That is why approval is not an extra option but the starting point: you decide per task whether it may finish the job or has to put it forward first.
That also matches what the AI Act expects of companies: where decisions affect people, a person keeps oversight.
What an agent is not
Not an employee. Not a jack of all trades. Not a system you switch on and then forget. It is a piece of work you have written down and handed over, on the understanding that you can read every run back.
Frequently asked questions
Is an AI agent the same as automation?
Almost. An automation follows fixed rules. An agent may, within those rules, write the text itself, make the summary or propose the choice — and it puts that choice forward.
Does an agent think for itself?
No. It carries out an assignment you have set out, with the sources you have pointed it at. What it makes of it can differ each time; what it is allowed to do cannot.
Can an agent learn from mistakes?
Not by itself. You adjust the assignment when something goes wrong. That is an advantage: nothing changes without someone having decided it.
How many tasks can one agent do?
One. That is deliberate. An agent that does three things is three times as hard to check when something goes wrong.