An AI agent is a piece of software that can read a task, work out what needs to happen next, and take that action itself, rather than just answering the one question you type into it. A chatbot waits for your next message; an agent keeps going, pulling in the information it needs, drafting the reply or the report, and moving on to the next step in a written job. The language model gives it the reasoning to understand a messy, real-world request; the agent part is what lets it act on that understanding instead of just describing what it would do. In a small business, that might look like an agent that reads a new supplier email, checks what's already been agreed, and drafts a reply, all without you prompting each step in turn.
You've probably typed a question into an AI tool already: something that answers a query, drafts an email if you ask it to, or summarises a document you paste in. Every one of those answers depends on you doing the typing, in real time, one message at a time. If a customer chases an overdue order, you still open the order system yourself, look up what happened, and write the reply. The AI answered your question; it didn't do your job.
An agent is the word for what happens when that changes: software that can do the whole job, not just supply the one reply, without you sitting over it prompting every step. That's the part that's hard to picture if you've only used AI as a search box with better manners, and it's also why "agent" now gets used both for something genuinely useful and for products that are, underneath, still waiting for you to type the next line.
An agent runs on a written job, not a single question you type
Ask a chatbot something and it answers from what you typed and whatever it already knows; ask again tomorrow and it has no memory of yesterday's conversation unless you paste it back in. An agent works differently because it starts from a written job rather than a one-off question: a description of what it's responsible for, what tools it's allowed to use, and when it runs. In practice that written job is a real file, not a vague setting. It names the task the agent owns, the systems it can touch, and the limits on what it's allowed to decide without checking with a person first. That file is what lets an agent pick up the same kind of task on Tuesday that it handled on Monday, in the same way, without you re-explaining it each time.
The language model does the reasoning; the agent does the deciding and doing
The "AI" in AI agent is a language model: the part that reads an email, a spreadsheet row, or a customer message and works out what it means, in plain language rather than rigid rules. On its own, a language model only ever produces text. It's the agent wrapped around it that turns that understanding into a decision and an action: read the invoice, check it against what was ordered, flag the mismatch, draft the query to the supplier. The model supplies the judgement about what the message means; the agent is the part that holds the job, remembers where it got to, and carries the answer through to something done, or drafted, rather than just said.
One agent rarely covers a whole job by itself
A single agent tends to do one thing well: chase invoices, or answer a common customer question, or keep a spreadsheet current. Most real back-office and customer-facing work is made up of several of those things handed off to each other, the way a person handing a task to a colleague would. A small business adopting agents usually ends up with a handful of them, each with its own written job, working as a team rather than one agent trying to be everything at once. We've looked separately at which jobs suit a team of agents in the first place in where AI agents work best.
An agent that acts still needs a person checking its work
The fact that an agent can act, rather than just answer, is exactly why the check on that action matters more, not less. An agent reading a customer message can misread an ambiguous request and act on the wrong interpretation with total confidence, because sounding certain and being right are two different things for a language model. That's a real, documented failure mode, not a hypothetical one, and it's the reason a well-built agent's job description also spells out where it stops and hands a decision to a person rather than acting alone.
An agent explained well can still sound more capable than it is, and that's worth admitting plainly rather than glossing over. The job description, the tool limits, and the human check are what make an agent trustworthy day to day; strip any one of those out and you have software that acts on its own without anyone able to say why. The idea of an AI agent is genuinely useful. What makes one safe to run in a real business is never the AI alone.
None of that means an agent needs constant supervision to be worth having. It means the work you hand over is real work, done properly, with the one step that carries risk, sending, spending, or committing to something, still landing in front of a person before it happens.