What is an AI agent, and what would one do in a small business?

An AI agent is software that is given a job to do. It reads what is in front of it, decides the next step, and acts through the tools it has been given, within limits a person sets. When it is unsure, it stops and hands the case to someone.

How an agent differs from a chatbot

A chat assistant waits for you to type. You ask, it answers, and nothing happens until you ask again. An agent works without being prompted: give it the job of entering supplier invoices and it opens each one as it arrives, reads it, and enters it.

The other difference is tools. A chat assistant produces text. An agent can also act, for example by looking up a purchase order or writing a row into the accounting system. Because it can change things, it needs limits.

What an agent does in a small business

The useful agents are narrow. Each does one job that someone on the team does by hand today.

• Enters invoices

  Reads each supplier invoice, matches it to its purchase order, and enters the ones that match.

• Drafts follow-ups

  Finds quotes with no reply and drafts a follow-up for a person to approve.

• Sorts the inbox

  Turns after-hours email into tasks assigned to the right person.

• Builds the weekly report

  Pulls the numbers from each system and lays them out the same way every week.

• Checks the numbers

  Compares this week with last week and flags anything that looks wrong.

What a good agent has around it

• One job, written down, with a clear finish.
• Limits on what it may do without approval.
• A rule for when to stop and ask a person.
• A log of everything it did.
• A named owner who is told when something looks off.

With these in place, a person can see what the agent did and stop it when something goes wrong.

When an agent is the wrong tool

If the steps never change, you do not need an agent. A plain workflow that follows fixed rules is cheaper and easier to trust. An agent is worth the extra care when the inputs vary, such as documents that never look quite the same or messages written by people.

An agent is also the wrong tool when a mistake cannot be caught. If nobody can review the result before it matters, keep a person doing the work.

Where to start

Pick one job that repeats every week and that someone can describe step by step. You will need past examples to test against. Build the agent for that job only. Start it in draft mode, with a person approving its work, and widen its limits as the results hold up.