"AI agent" is used for everything from a chat window to a fully automated back office. That makes it hard to buy one, build one or tell whether a vendor is selling one. Here is a definition you can use.
The short definition
A business AI agent is software that is given a goal, reads the systems your business already uses, takes steps with tools to reach the goal, and checks its own work, with a person approving the steps that matter.
Four parts make that work:
- A goal. "Prepare a quote for this request," not "be helpful."
- Access to your data. It reads from your ERP, CRM, inventory or spreadsheets.
- Tools. It can do things: look up a price, draft an email, create a purchase order for approval.
- A loop with checks. It tries, looks at the result, and corrects itself or asks a person when it is unsure.
Chatbot, copilot or agent?
| Chatbot | Copilot | Agent | |
|---|---|---|---|
| What it does | Answers questions | Helps a person do a task | Does the task, then reports back |
| Who drives | The person asks each time | The person, step by step | The agent, within limits |
| Needs your data | Sometimes | Often | Always |
| Typical example | Answering "what are your hours?" | Suggesting wording in an email | Preparing a quote from a request, for approval |
The lines blur, and many useful products sit between the columns. The practical question is not what to call it. It is who does the work and who checks it.
What a business agent actually does
The jobs that suit agents are repetitive, rule-shaped and tied to data you already have:
- Answer customers using your real stock, order status and policies.
- Prepare quotes from a request, a price list and a customer's history.
- Reorder stock when forecast demand and current inventory say it is time.
- Follow up on open quotes, late invoices or accounts that have gone quiet.
In each case a person stays in charge of the decision that carries risk, such as sending a quote or spending money, and the agent does the preparation.
What an agent needs to work
Trustworthy data. An agent that prepares a quote reads prices, customer history and stock. If those disagree between your systems, the agent will pick one with complete confidence. This is why agents sit at the top of the four levels of data value: they work best on top of a platform where each number is defined once.
Narrow permissions. Start with read-only access. Give write access one step at a time, and only where a person can approve the result.
A measure of success. Decide before you build what number should move: quote turnaround, stockouts, response time. An agent without a measure is a demo.
An owner. Someone on your team who knows how it works and can change it. A system that only the builder understands is a risk.
How to scope a first agent
- Pick one process. The one that is repetitive, painful and well understood by the people who do it today.
- Name the number. For example, "the time from request to quote goes from three days to a few hours, with a person approving each quote."
- Check the data. Can the agent read what it needs, and do the sources agree? Fix that first.
- Begin read-only, with approval. Let the agent prepare, and let a person send. Loosen limits only as results justify it.
- Measure from day one. Compare against how the process runs today.
Questions to ask anyone who sells you one
- Which systems does it read, and with what permissions?
- What happens when it is wrong, and who sees that?
- Which steps need human approval, and can I change that?
- How will we know in weeks, not months, whether it is working?
- Who on my team will be able to run it without you?
See one at work
In The Yagrumo Market, one of the shops has a furniture workshop whose quotes take three days. You can watch an AI agent draft one in seconds, with a person approving it. It is a good way to see the pattern before you build your own.