Chatbot vs AI agent: the difference that matters

The difference in chatbot vs AI agent is what happens after the model produces output. A chatbot returns text and stops — the next action is yours. An agent treats its output as a plan, executes it against real systems, reads what happened, and keeps going until the goal is met or it needs you. Everything else is a consequence of that.

The same interface, a different contract

They often look identical. A text box, a conversation, a response. That similarity is why the terms get used interchangeably, and why buyers end up disappointed.

The contract is what differs. When you tell a chatbot “unsubscribe me from these newsletters”, it tells you how to unsubscribe. When you tell an agent, the newsletters are unsubscribed.

That gap sounds small in a sentence and is enormous in engineering. Producing correct instructions requires the model to be right. Executing them requires the system to be right about authentication, permissions, error handling, partial failure, and what to do when the third of twenty-three operations returns a 429.

Where a chatbot is the better product

Reaching for an agent when a chatbot would do is a common and expensive mistake.

When the human should decide. Drafting, brainstorming, explaining, summarising a document you already have. The value is in the thinking, and inserting execution adds risk without adding benefit.

When there’s nothing to act on. A support bot answering “what are your hours” has no state to change. Give it retrieval, not tools.

When the cost of being wrong is asymmetric. A wrong paragraph is edited. A wrong action is a support ticket, a refund, or an apology.

When you need speed. No tool round-trips, no planning loop. A chatbot answers in a second; an agent working a real task takes tens of seconds because it’s actually doing something.

A well-built chatbot that knows its limits is a better product than a badly-built agent that doesn’t.

The three things an agent needs that a chatbot doesn’t

Tools with real credentials. Not a demo key — scoped, expiring access to your actual inbox, calendar, or database. This is where most of the security engineering lives, and it’s invisible in a demo.

A loop that reads results. The model proposes, the system executes, the result comes back into context, and the plan is revised. Without the feedback edge you have a script that a model wrote once, which fails the moment reality differs from the plan.

A gate on irreversible steps. A chatbot needs no permission model because it can’t do anything. The moment a system can send, delete or pay, you need an explicit answer to “what can it do without asking?” — and both “everything” and “nothing” are wrong answers.

The comparison, honestly

Criterion	Chatbot	AI agent
Output	Text	Changed state
Latency	~1s	Seconds to minutes
Cost per interaction	One inference	Several, plus tool calls
Failure mode	Bad answer, you ignore it	Wrong action, already taken
Needs credentials	No	Yes, scoped
Needs approvals	No	Yes, on destructive steps
Auditability	Chat log	Full run trace

The failure row is the one to sit with. Chatbot errors are absorbed by the reader. Agent errors are absorbed by your systems, and the blast radius equals the permissions you granted.

How to tell what you’re being sold

Marketing has thoroughly blurred this, so test it:

• Ask it to do something with a side effect. If the answer is instructions, it’s a chatbot.
• Ask what happens when a tool call fails. Real agents describe retry, escalation or abort. Chatbots-with-plugins have no answer because they never hit the case.
• Ask what it can do unsupervised. A real product has a policy — reads run freely, irreversible writes stop for a human.
• Ask to see a past run. An agent produces an ordered trace of steps and tool calls. A chatbot produces a transcript.

The useful middle

The strongest products aren’t purely one or the other. They start conversational — you describe an outcome in plain language — then become agentic once the plan is clear, and hand control back at the moments that matter.

That’s the shape we built: talk to it like a chatbot, watch it plan like an agent, and approve anything irreversible before it happens.

For the fuller definition, see what an AI agent is — or drive one yourself.