What Is an AI Agent for Small Business?

Abstract graphic of a central node with orbiting points, representing AI agents, from Nantucket AI

Every software company is selling an AI agent this year. Some of what they are selling is a chat window with a new label on it. Some of it can send email from your address, move files in your folders, and change entries in your calendar.

Nobody explains the difference in the pitch, and that difference is the entire decision. So here is the plain answer to what an AI agent is for a small business, what agentic work actually looks like, and where the line sits between the two.

Chat writes. An agent acts.

A chat assistant produces text. You ask, it answers, you copy the answer somewhere and do something with it. You are the transport layer. Nothing happens in your business until you make it happen.

An agent does the moving itself. It reads the actual files, takes several steps in a row without asking between each one, and changes something real. The output is not a paragraph you might use. The output is a renamed folder, a drafted reply sitting in your outbox, a spreadsheet with two hundred rows filled in.

One question sorts almost any tool into the right bucket. What can this thing change without me touching it?

If the answer is nothing, it is an assistant, and your worst case is a bad draft you throw away. If the answer includes your email, your files, your calendar, your books, or anything a customer sees, it is an agent, and your worst case is considerably larger.

Notice that the question is about reach, not about intelligence. A modest tool with access to your accounting system is riskier than a very capable one that can only write into a text box. Reach is the number that matters.

What agentic work looks like on a Tuesday

The useful examples are boring, and that is a good sign.

An office manager has eleven years of board minutes in a folder. Finding when the parking policy last changed means opening files one at a time. An agent reads all of them and returns the three meetings where it came up, with the file name and date for each, so she can open those three and verify.

A property manager has two hundred supplier invoices for the season and a spreadsheet that has to be filled in from them. That is a full day of typing. Agentic work turns it into a run she watches and then checks against a sample of the originals.

A restaurant rebuilds the same labor cost report every month by hand from four exports. The pattern never changes, only the numbers do. That is exactly the shape of job worth handing over, because you already know what a correct answer looks like.

The common thread is not glamor. It is repetition, a clear definition of correct, and a result you can check.

Supervision is the actual job

The instinct is to ask how capable the tool is. The better question is how much you are still checking.

That is the axis behind the trust ladder, which sorts AI work into four rungs. Orient is learning what the tools are before you hand them anything. Assist is where you write the ask, the tool drafts, and you decide what ships. Supervise is where the tool runs a whole task while you watch. Delegate is where recurring work runs inside rules you wrote, and your job is reading the record afterward.

Agentic work starts at Supervise. That rung has three habits, and businesses that get burned usually skipped all three.

Scope the task small enough to check. If you cannot verify the result in a few minutes, you have not scoped it, you have gambled on it.

Make it reversible. Work on a copy. Know your undo. A task you cannot take back is a task you should not have started.

Read the permission prompts. When a tool asks whether it may do something, that prompt is the control. Clicking through them out of habit removes the only safeguard you had.

And when it finishes, check the result against the source material, not against the tool’s own summary of what it did. The summary is written by the same system that did the work. It is not an audit.

What a small business should try first

Pick one job with four properties. It repeats. You already know what correct looks like. Nothing customer-facing goes out at the end of it. And you can undo it.

Good first candidates are a recurring internal report, a folder of documents you need answers out of, a pile of receipts or invoices that has to become a spreadsheet, or a cleanup of file names and records nobody has had time for.

Bad first candidates are anything that sends, anything that publishes, and anything that touches money.

Run the job three times with your eyes on it before you form an opinion. The first run teaches you where it drifts. The second tells you whether the first was luck. By the third you know whether this belongs in your operation or not.

What to never hand off

Some work stays with a person, and not because the tools are weak.

Anything you sign. Anything that commits money or agrees to terms. Anything where a person is the subject, including hiring, firing, discipline, references, and reviews. Anything legal, medical, or financial that a professional is supposed to be accountable for. Anything a regulator or an insurer could ask you to defend later. And anything where the whole value is that a human being paid attention, like a condolence note or an apology to a customer you want to keep.

There is also a plain honesty rule that has nothing to do with risk tolerance. Never let a tool invent a fact about your business. Not a statistic, not a review, not an award, not a client name. If you did not supply the number, it does not go out.

Where to learn this properly

If you are still at the stage of sitting in meetings hearing the word “agent” and wanting a straight answer before anyone signs anything, start with What Is an AI Agent. It is free, takes about two and a half hours, names no products, and recommends no vendors. You finish able to ask the four or five questions that decide whether something is safe to try.

When you are ready to actually run a supervised task on your own files, Intro to Claude Code is the hands-on version. It is not a programming course. Every example is a business job, and you finish having completed one real task you did not do by hand.

Both are free. Start with the one that matches where you are, not the one that sounds more advanced.