An AI agent is not a chatbot

The two words are used as if they were a spectrum, with the chatbot at the timid end and the agent at the ambitious one. They are not the same kind of thing at all, and the difference is not how clever the model is. It is whether the software is allowed to touch anything.

The difference in one sentence

A chatbot returns text for a person to use. An AI agent takes actions in real systems on that person's behalf. Everything else follows from that: a chatbot's worst output is a bad paragraph you can delete, while an agent's worst output is an email in a customer's inbox, a refund on a card, or a row missing from a ledger.

This is why the question people ask about agents is never really about intelligence. Nobody asks whether the model can write the apology email. They ask what happens if it sends the wrong one to the wrong person at two in the morning.

Four things change the moment software can act

• Mistakes leave the building. A wrong answer in a chat window costs you a re-prompt. A wrong action costs you a relationship, a refund, or a correction you have to send under your own name.
• Cost becomes continuous. A chat turn is one bill you asked for. An agent runs many steps without being asked again, so the spend has to be visible while it happens and capped before it starts. We wrote about that in what a delegation costs.
• Permissions become the product. With a chatbot, access is a paste. With an agent, access is a standing grant to your mailbox, your CRM or your payment processor, and the interesting design work is deciding how narrow that grant can be.
• Review moves from output to action. You stop reviewing prose and start reviewing consequences: this recipient, this amount, this text, this time.

So briefing changes too

A prompt asks for text. A brief describes a finished piece of work, which is a different sentence to write. It carries what done looks like, the context to read first, the limits that must hold, and the moments where a person wants to be asked. That is the same thing you would tell a new member of staff on their first morning, and it is why we think the unit of work should be a brief rather than a chat thread.

The alternative that most tools reach for is a builder: draw every branch in advance, then watch it survive contact with reality. We have written about why we start from the brief instead in a brief beats a flowchart.

And supervision becomes a product surface

Once software can act, the supervision is not a setting buried in an admin page. It is the part you use every day. In Delagents it is the approvals inbox: an agent that is about to send, pay, post or delete stops and shows the exact action, what it costs, and who else could decide. One click approves it, and the decision is recorded under the name of the person who made it.

Under that, eight kinds of action always stop for a person, whatever an agent's instructions say. That list is deliberately short and deliberately fixed, and it is the subject of eight things an agent never does without a person.

When you actually want a chatbot

Often. For thinking out loud, for drafting in a document you are already editing, for a question you will act on yourself in the next thirty seconds, a chat window is the right tool and an agent is overhead. The honest rule we use: if you would not hand the task to a person and walk away, you do not need an agent for it either.

What you should not do is take a chat assistant, hand it a standing grant to your Gmail, and call the result an agent. That is a chatbot with a blast radius.

The test worth applying

When a product calls itself an agent, ask three questions. What can it change without asking? What does it show you before it acts? And what happens when you do not answer? A tool that cannot answer those three is not more capable than a chatbot. It is the same thing with the brakes removed.

In Delagents, every new agent's first run is draft-only, whatever its instructions say. Write access unlocks after a person has reviewed one complete delegation.