What an AI agent actually is

An automation does what you specified. An agent decides what to do. What that difference means in practice, the six decisions that make up most of the work, what it looks like in a small business, and where the time really goes.

The word agent is being used for everything at the moment, from a chatbot on a website to a system that files your invoices. For somebody running a business, the useful definition is narrower than the marketing one, and the difference it describes is worth understanding before you buy anything.

The distinction that actually matters

An automation follows a path you worked out in advance. When an invoice arrives, take these five fields, check them against the purchase order, and put the result in your accounting system. Every run is the same shape. If something turns up that the path did not anticipate, it stops and waits for a person.

An agent is given a goal and a set of tools, and works out the steps itself. Resolve this supplier query. To do that it might read the email thread, look up the order, check whether the delivery was signed for, and then either draft a reply or decide the matter needs a human. Nobody wrote that sequence down. It chose it.

Which means the honest advice is unfashionable: if your task is the same every time, you want an automation. It is cheaper, more predictable, and easier to check. Agents earn their place where the work varies so much that writing the path down was never going to work.

What they are genuinely good at

Work that follows the same intent but never quite the same shape. Questions that need three systems checked before they can be answered. First drafts of things a person will review anyway. Triage, where the job is to decide what each incoming thing is and where it should go.

What they are not good at is anything where being confidently wrong is expensive. An agent that cannot find the answer will often produce a plausible one rather than stopping, and it does not sound any less certain when it does.

Setting one up: the six decisions

Most of the work in an agent project is not building anything. It is deciding these, in this order, and writing them down.

The boundary is the one people skip, and it is the one that decides whether the thing survives contact with your business. Write it as a list of what it may do without asking. If that list is uncomfortable to read, it is too long.

What it looks like in a small business

Five that come up repeatedly, none of which require you to be a technology company.

Inbox triage with drafted replies: every incoming enquiry read, sorted, filed against the right customer, and a reply drafted for somebody to glance at and send. Quoting: an enquiry read against your price list and your rules, producing a draft quote with the assumptions it made written underneath. Chasing: the agent notices what is missing from a job or a file, works out who has it, and asks them, repeatedly and politely, until it arrives.

Then the two that tend to surprise people. Preparing for a decision: before a weekly meeting, pulling the relevant numbers and recent history out of three systems and writing the half page somebody would otherwise assemble by hand. And first-pass review: reading a long document against a checklist and flagging what deserves human attention, which is the opposite of letting it decide anything.

Where the time actually goes

It is worth being precise about this, because the usual pitch is not quite right. An agent rarely removes a job. What it removes is the first twenty minutes of a great many jobs: the looking up, the gathering, the staring at a blank page before the actual thinking starts.

The second saving is less obvious and often larger. Work that waits for somebody to get to it stops waiting. The enquiry that would have sat until Tuesday has a draft reply against it on Friday afternoon, and the person who picks it up begins most of the way through rather than at the beginning.

That is a real saving and it is not the same as replacing anybody. Businesses that buy an agent expecting the second thing tend to be disappointed, and businesses that buy it expecting the first are usually pleased.

The things worth being careful about

Anything that reads input from outside your business can be influenced by that input: a message can contain instructions as easily as information, and the agent has no reliable way to tell the difference. This is why what it is allowed to reach matters more than how clever it is.

Agents also cost money per step, and one that loops can spend a surprising amount achieving nothing. And they change underneath you: the model gets updated by its provider, and behaviour that was settled in March can shift. All three are manageable. None of them manage themselves.

The short version

An agent is software that decides its own steps toward a goal you set. Use one where the work varies too much to write down, and an ordinary automation everywhere else. Most of the project is deciding the job, the tools, the boundary, the knowledge, the hard cases and the record. Expect it to remove the first twenty minutes of many jobs rather than a job, and keep a person wherever being confidently wrong would cost you.