The short answer
A chatbot answers messages. A copilot helps a human do the work faster. An agent does the work: multi-step action through real tools, finished tasks, judgment along the way. Vendors use the three words interchangeably because agents price highest. The test that cuts through: what exists in your systems after the conversation ends?
The words are used interchangeably in every pitch you will hear this year, and the confusion is not accidental: agent is the premium word, so everything gets called one. The machines are genuinely different, and buying the wrong one wastes a budget precisely.
The three machines
The chatbot is a conversation interface over knowledge: it answers questions about your pricing, your policies, your docs. Its unit of output is an answer, and when the conversation ends, the work it discussed still exists, untouched, for a person to do. The copilot sits inside a human's workflow and accelerates them: drafts the reply, summarizes the thread, suggests the formula. The human remains the actor; the copilot is leverage on their hours. The agent is the actor: given a goal, it reads context, decides, acts through tools, checks results, and continues until the task is done or a boundary says stop. The full anatomy is here; the one-line version is that agents complete work, not messages.
The comparison
| Chatbot | Copilot | Agent | |
|---|---|---|---|
| Unit of output | An answer | A better draft | A completed task |
| Who does the work | A human, later | A human, faster | The system, with humans on exceptions |
| Acts through tools | No | Inside one tool | Across your systems |
| Economic effect | Deflects questions | Multiplies a person | Subtracts a queue |
| Failure mode | Wrong answers | Bad drafts a human catches | Wrong actions: needs boundaries, evals, monitoring |
| Engineering burden | Low | Low to medium | High: production discipline required |
The tells
Two questions expose any offering in under a minute. First: what exists after a conversation with it ends? Nothing but an answer is a chatbot. A better version of something you were already writing is a copilot. A booked meeting, an updated record, a processed document, a completed task in your systems: that is an agent. Second: what happens when a step fails? Real agents act through real systems, so their builders have answers about boundaries, escalation, and monitoring; costume agents have adjectives. The industry data says costumes dominate: Gartner estimated only about 130 of the thousands of self-described agentic vendors are real.
The test is one question: what exists in your systems after the conversation ends?
Which one you actually need
All three are legitimate purchases for different problems. Repetitive questions about known facts: a chatbot, cheaply. Skilled people spending hours on drafting and summarizing inside their tools: copilots, usually bought not built. A queue with volume, structure, and judgment at the edges, where the work itself should stop consuming a team: an agent, and now the decision becomes build versus buy and the engineering bar becomes production discipline. We run an agent on our own funnel, which is why this taxonomy is drawn from operating one rather than from a pitch: the differences above are the ones that show up in the invoice and the incident log.