An agent takes actions. That is the whole distinction.
A chatbot answers. An agent does. That is the line, and almost everything else said about agents is elaboration on it.
Ask a chatbot about an invoice and it tells you about invoices. Give an agent the same question and it can open your system, find the invoice, notice it does not match the order, and either fix it or tell somebody it could not. The language part is the same. What differs is whether it is allowed to reach into anything.
Which means the interesting question about an agent is never how clever it is. It is what it is permitted to touch, and what happens when it is wrong.
What one looks like in a business
Concretely, in the kind of business we work with, an agent is usually one of these:
- Something that watches. A deal that has gone quiet, a job drifting past a date, a supplier invoice that does not match its order. Attention is expensive and machines have an unlimited supply of it.
- Something that answers from your own material. What did we quote them last time, and why? A question your systems already contain the answer to and nobody can find in under ten minutes.
- Something that prepares. Drafts the quote, assembles the pack, fills the form to the point where a person checks it and sends it.
Notice what is not on that list: nothing that has to be right every single time with nobody checking. That is not a limitation of today's models so much as a design rule.
The three questions that actually matter
When somebody proposes an agent, these are the questions that determine whether it is a good idea, and none of them is about the model.
- What can it reach? Which systems, which records, and whose data. This decides your exposure, and it is a structural decision rather than a setting.
- What happens when it is not sure? A well-built agent stops and asks. A badly-built one guesses fluently, which is the expensive failure because it does not look like one.
- Who sees what it did? If nobody can review the actions afterwards, nobody can trust it, and the first surprise ends the project.
Where the word gets stretched
A good deal of what is sold as an agent is a scheduled task with a language model attached to one step. That is often exactly the right thing to build, because it is cheaper, more predictable and easier to reason about. But know which one you are buying: the price and the risk are different.
The practical test: if you can draw it as a flowchart with no branch that says and then it decides, it is automation. That is not a criticism. It is a different product.