Three different tools
A chatbot follows a script or retrieves a passage and stops. A fixed automation follows a path you drew: if the status is this, send that, write this field. An agent is given a goal, a small set of tools, and a stop rule. It can choose a next step. It can also choose a wrong one.
The choice is not a maturity ladder. Many businesses should stay on the fixed path. An agent is justified when the input does not fit a form and the next step depends on reading it.
| Tool | Input it expects | What it should not do |
|---|---|---|
| Chatbot | A question you anticipated | Change a record or make an offer |
| Fixed automation | A known field and a known status | Invent a step that is not on the path |
| Agent | Unstructured text or a file, plus tools you named | Act outward without the review you specified |
A workflow, not a personality
Describe the agent as a clerk with a tray. Something arrives. The clerk classifies it, looks up the allowed source, and prepares a draft or a routed task. The clerk does not own the customer relationship.
Write the stop rule in the same paragraph as the goal. “Draft a reply and stop” is a different system from “resolve the ticket.” If you cannot name the stop, you are not ready to build.
What enters, what is prepared, who approves
Take an inbound email about an order. The email enters. The system classifies the request, retrieves the order from the system of record, and drafts a reply using only that record and an approved explanation. A person on the team sends the reply or throws the draft away.
If the order cannot be found, the system stops and queues the email. Guessing an order number is a failure, not a helpful recovery.
When the simpler tool is the right one
If a form would collect the same facts, build the form. If the branch is a status you already store, use a rule. If the cost of a wrong outward message is high and you have no one to review the queue, do not put an agent in front of customers.
Related service: Agentic AI and automation