An AI agent is software that can take actions towards a goal, not just produce text. It reads data, decides what to do, and calls real functions to do it: raising an invoice, sending a message, creating a task. The difference from a chatbot is that a chatbot tells you what to do and an agent does it.
The word agent is doing a lot of work in a lot of marketing, so it is worth pinning down what it actually means.
The distinction is not how clever the model is. It is whether the software can reach out and change something.
A language model produces text. That is all it does. Given your question it produces an answer, and the answer might be excellent, and nothing in the world has changed.
An agent is that model wired to a set of functions it is allowed to call. Instead of writing raise an invoice for £400, it calls the function that raises the invoice. The model is still only producing text, but some of that text is now an instruction the surrounding system carries out.
That wiring is the whole of it. There is no separate agent technology. There is a model, a list of things it may do, and code that does them.
Something starts it: a schedule, a message, a button. The agent reads whatever it is allowed to see. It decides what to do. It calls a function. It tells someone what happened.
Everything interesting about a particular agent is in the middle three steps: what it can see, what it may decide, and what it is permitted to call.
Before letting one near a business, these are the ones that separate a useful agent from a liability.
An agent is only as good as the records it reads, and only as safe as the records it reads. An agent with access to everything is a bigger problem than an agent with access to your invoices.
Ask for the list of functions. It exists, it is finite, and it is the real answer to what this thing does. Anything not on that list is something the agent cannot do no matter how it is asked.
The most important question, and the one most likely to be answered vaguely. Software that drafts and waits is a fundamentally different product from software that acts and reports.
An agent that acts leaves consequences. If there is no record of which action it took and when, then the first time you find out is when a customer complains.
It will be. The question is whether being wrong produces a draft you delete or an email your customer has already read.
A good agent uses a model for the parts that need judgement and ordinary code for the parts that do not.
Deciding which invoices are overdue is a date comparison. Putting that through a model adds nothing, costs money, and makes the answer change between runs. Writing the message that chases one is genuinely a language problem. Splitting the two is what makes an agent explainable, because the half you might disagree with is a rule rather than a mood.
The concept is simple. What it is worth depends entirely on what it is wired to.