What an AI agent actually is, in plain English.
The word "agent" now turns up on almost every software sales page, attached to everything from a chat window to a scheduled email. For an owner or department head trying to decide what to buy or build, that makes the word close to useless. It has a fairly specific meaning, though, and once you know it you can sort most of what you're offered in a few minutes.
Three kinds of help that get called the same thing.
Most of what gets sold as AI falls into one of three groups. The first is chat: you type a request, the model writes something back, and you decide what to do with it. Drafting a reply to a difficult email, summarizing a long contract or rewording a job posting are all chat. A person reads every output before it goes anywhere.
The second is an automation. Someone writes down the steps in advance, and software follows them every time a trigger fires. When a form is submitted, copy the fields into a spreadsheet and send a confirmation. An automation can include an AI step in the middle (summarize this ticket, sort this message into one of five folders), but the path through the work is fixed. It does the same thing on Tuesday that it did on Monday.
The third group is agents, which is what people usually have in mind when the word comes up with some excitement in a meeting.
What makes something an agent.
An agent is given a goal and a set of tools, and it decides the steps for itself. It looks at the situation, picks an action, checks what happened, and picks the next action based on the result. It keeps going in that loop until it thinks the goal is met or it runs out of options.
Take a clinic front desk. An automation could send every new patient the same intake packet. An agent, given the goal "get this patient booked for the right appointment," might read the referral, see that a form is missing, look up the schedule, notice the requested provider is out next week, offer two alternatives and draft a note to the referring office. Nobody wrote those steps down in that order. The agent chose them.
That freedom is the whole point, and it is also where the risk comes from. Because the agent chooses its own path, two runs of the same request can go differently. It can also choose a step you would not have chosen, such as emailing the wrong contact or updating a record it should have left alone.
An automation follows the route you drew. An agent draws its own route each time, so you have to decide how far it is allowed to walk.
Why the label gets stretched.
Vendors have good reasons to call things agents. The word suggests that the software does the work instead of helping with it, and that is an easier thing to sell. So a chatbot with a search box becomes an "agent," and so does a two-step automation with an AI summary in it.
A quick test helps. Ask the vendor, or yourself, who decides what happens next. If a person decides after every reply, it is chat. If the steps were fixed in advance by whoever set it up, it is an automation, however clever the individual steps are. Only when the software chooses its next action based on what it just saw is it an agent.
None of this makes the simpler tools worse. In the work we do with teams, most of the time saved comes from chat used well and from plain automations that run reliably, and only a small share needs anything that plans its own steps.
When the extra freedom is worth it.
Agents earn their keep when the work is varied enough that you could not write the steps down in advance. Intake that arrives in a dozen formats is a good candidate, and so is a request that needs two or three systems checked before anyone can answer it. In each case a person currently spends their time deciding what to look at next.
They are a poor fit when the work is the same every time. Monthly invoice reminders and the weekly staff schedule do not need judgment, and handing them to something that improvises only adds ways for them to go wrong. A fixed automation is cheaper to run, easier to check and does exactly what it did yesterday.
There is also a middle case. Some work needs judgment at one point and nowhere else, such as deciding whether a message is a complaint or a routine question. That usually calls for an automation with a single AI step at the point where judgment is needed, which is far easier to test than a full agent.
A decision guide: chat, automation or agent.
Before anyone in your organization buys or builds something, walk the task through these questions in order. Stop at the first answer that fits.
- Will a person read and act on every output anyway? Use chat. Write a good set of instructions, save them where the team can find them, and stop there.
- Are the steps the same every time, even if the content changes? Build an automation. If one step needs a judgment call, add a single AI step there and leave the rest fixed.
- Do the steps change depending on what turns up along the way, so that writing them down in advance would mean dozens of branches? Consider an agent.
- If it is an agent, what can it read, what can it change, and what can it send in your organization's name? Write those three lists down before it runs once, and start with read-only access.
- Who will look at what the agent did, and how often? If the answer is nobody, go back to question 2 and find a version of the work that can be fixed in advance.
Most tasks stop at the first or second question, and that is normal. A school district office that answers parent emails with good saved prompts, and routes enrollment forms with a plain automation, has covered a lot of ground before it needs anything that plans for itself.
