"AI agent" has become a catch-all term, which makes it a genuinely confusing thing to plan a budget around. The distinction that actually matters for product decisions isn't marketing language — it's what the system is allowed to do once it starts working.
A chatbot answers. An agent acts.
A chatbot, in the useful sense of the word, takes an input and produces an output — a conversational interface over a knowledge base, a support answer grounded in your documentation, a first draft of a reply. It's a single step. The interesting engineering problem is making sure the answer is grounded in real information, not making things up.
An agent takes a goal and plans a sequence of actions to get there, calling tools along the way — searching a database, filling in a form, triggering a workflow, checking its own output against a rule before proceeding. The interesting engineering problem shifts from 'is this answer accurate' to 'did this multi-step process stay inside the boundaries we defined, and can we tell what happened if it didn't.'
Why the distinction changes the build
A chatbot that gives a wrong answer is a bad experience. An agent that takes a wrong action can change data, send a message, or spend money. That's not a reason to avoid agents — it's a reason to design them differently from day one: narrow scopes, explicit tool permissions, and a human checkpoint sized to the risk of what the agent is doing.
- Start narrow: give an agent one well-defined job before generalising it.
- Log everything: every tool call and decision needs to be inspectable after the fact.
- Put the checkpoint where the cost is: low-risk actions can run unattended; anything irreversible gets a review step.
Most businesses need both, in the right order
In practice, the products we build usually start with a grounded chatbot or assistant — because it's faster to ship, easier to evaluate, and teaches everyone (including us) where the model is reliable and where it isn't. Once that trust exists, specific workflows graduate into agent territory: the assistant that used to draft a reply now files the ticket, updates the record and notifies the right person, with a person spot-checking the exceptions.
The mistake we see most often is reaching for an agent on day one because it sounds more advanced, before anyone has verified the model is even reliably right about the underlying facts. Get the answers trustworthy first. Let the actions follow.