What an AI agent is, and when a business needs one

A chatbot answers. An agent does the work.

A chatbot takes a question and returns text. What happens next is up to the person who asked.

An agent takes a goal. It plans the steps, uses the systems it has been given and carries the work through to a result. If the goal is “answer this customer”, a chatbot can suggest a reply. An agent reads the message, looks up the account, checks the order, drafts the reply and puts it in front of a person to approve.

The model inside can be the same. The difference is what sits around it: tools, rules and a clear definition of what “done” means.

How an agent differs from classic automation

Classic automation follows fixed rules. If a field has this value, do that. It is fast and cheap, and it is the right choice when the input always looks the same.

It stops working when the input is language. An email can ask the same thing in a hundred ways. A contract can put the same clause on page two or on page nine. That work needs reading and judgment, and that is where an agent fits.

A good system uses both. Rules handle what rules can handle. The agent handles the part that needs reading.

What an agent needs to work in production

A demo needs a model and a prompt. A production agent needs more:

• One clear task. “Handle the operations inbox” is a task. “Help the operations team” is not.
• Tools. Access to the systems where the work happens, with only the permissions the task needs.
• Approval rules. You decide which steps the agent takes alone and which wait for a person. Anything that commits the company usually waits.
• An evaluation set. Real cases from your own work, with the right outcome for each. This is how you know how well the agent performs before it goes live, and whether a change made it better or worse.
• Logs. A record of every run: what the agent read, which tools it used and what it decided.

Without the last two you cannot say whether the agent works. You can only hope.

When an agent is worth building

An agent is a good fit when most of these are true:

• The work is repeated every day and follows roughly the same steps.
• It needs reading or judgment, so rules alone never covered it.
• It touches several systems, and people spend time moving information between them.
• A mistake can be caught, because a person reviews the result or the step can be undone.
• You can tell a good result from a bad one.

When it is not

Do not build an agent when a simple rule does the job. Do not build one when a wrong result cannot be caught and the damage is large. And wait if nobody can say what a good result looks like, because then nobody can test it.

Where to start

Pick one task with a result you can measure. Build the agent against real cases. Run it next to your team until the results are trusted, then give it more to do.

Our use case for an operations inbox shows what that looks like. The AI Agents page describes how we build one with you.