What an AI agent actually is

An agent is judged on what it does, not on what it says. A chatbot's job ends at a reply. An agent takes the next real step itself: books the slot, updates the record, sends the follow-up.

The distinction only matters if the agent is allowed to touch something real. A calendar it can actually book into, a record it can actually update, a message it can actually send, not a sandbox copy that still needs someone to make the change by hand afterward. An agent connected to nothing real is a chatbot with a different name.

Working from the business's own data is the second requirement, not general knowledge with the company's name attached. A generic model asked about a clinic's opening hours will guess something plausible and confident; an agent built on that clinic's actual calendar and actual policies either knows the answer or says so.

The part that keeps an agent from causing damage is the same guardrail list every system here is built with: a written line between what it may decide alone and what has to reach a person, agreed before anything goes live. An agent built here starts at 2,000 euros for the first action, the same starting point as an AI chatbot, because underneath the different name they are frequently the same build.