The word 'agent' has been applied to everything from a decision-tree chatbot to genuinely autonomous systems. The distinction that matters practically is simple: can it take actions in your systems, or can it only produce text?
Chatbots produce text
A chatbot receives a question and returns an answer. A good one retrieves from your actual documentation so the answer is grounded rather than invented, but the output is still just words.
This is genuinely valuable — deflecting repetitive support questions saves real money — and it's comparatively safe, because the worst case is a wrong answer that a human can correct.
Agents take actions
An agent has tools. It can query your CRM, call an API, create a record, send a message, issue a refund. It works in a loop: decide what to do, do it, observe the result, decide again.
That's far more useful and far more consequential. A chatbot that misunderstands gives a bad answer. An agent that misunderstands might issue a refund to the wrong customer.
What this means for how you build
- Give agents the narrowest set of tools that lets them complete the job
- Put a human approval gate on anything irreversible — refunds, deletions, outbound customer messages
- Set hard limits on spend, volume and iteration count
- Log every action with its reasoning so you can audit what happened
- Define explicit escalation triggers rather than hoping the model asks for help
Where each fits
Start with a chatbot if your problem is answering the same questions repeatedly. It's cheaper, faster to build, and the failure mode is mild.
Move to an agent when the valuable part is the work after the answer — looking something up, updating a record, moving a process forward. Just budget for the guardrail engineering, because that's where most of the real work is.
A rough guide: if a mistake would cost you an apology, a chatbot is fine. If it would cost you money or trust, you need approval gates.