Copilot, chatbot or agent: which one do you actually need?
Three very different products get sold under the same word. How to tell them apart before you commission the wrong one.
Half the AI projects that disappoint were specified as one of these and built as another. The distinction is not marketing — it changes the architecture, the cost and the risk profile.
A chatbot answers
It takes a question and returns information. It may be grounded in your documents, but it does not change anything. The failure mode is a wrong answer, which is recoverable, and the build is bounded.
A copilot assists someone doing the work
It sits inside an existing workflow and drafts, suggests or pre-fills. Crucially, a human stays in the loop and presses the button. Copilots are the highest-return AI product for most businesses because they lift throughput without requiring the system to be right unsupervised.
An agent does the work
It plans, calls tools, takes actions and completes tasks without a human at each step. The upside is real leverage; the cost is that every action needs permissions, audit and reversal, and the failure mode is no longer a wrong sentence but a wrong transaction.
Three products, three risk profiles
Chatbot
- What it does
- Answers questions from information
- Worst case
- Someone is misinformed
Copilot
- What it does
- Drafts inside a human's workflow
- Worst case
- A human rejects a bad draft
Agent
- What it does
- Plans and executes without a human per step
- Worst case
- A wrong transaction, unattended
| What it does | Worst case | |
|---|---|---|
| Chatbot | Answers questions from information | Someone is misinformed |
| Copilot | Drafts inside a human's workflow | A human rejects a bad draft |
| Agent | Plans and executes without a human per step | A wrong transaction, unattended |
How to choose
Ask what happens if the system is confidently wrong at 3am with nobody watching. If the answer is 'somebody is mildly misinformed', a chatbot is fine. If it is 'a customer is refunded twice', you are building an agent whether you called it one or not, and it needs agent-grade controls.
- Information only, no side effects → chatbot
- Drafting inside a human workflow → copilot, and start here if unsure
- Unattended multi-step execution → agent, with authorisation and audit budgeted
The upgrade path
Ship the copilot first even when the agent is the goal. You get value early, and the review queue becomes the labelled dataset and the confidence evidence you need to justify removing the human later. Going straight to autonomy skips the step that earns trust.
The upgrade path
Ship the copilot first even when the agent is the goal.
- 1
Copilot in the workflow
Value early, with a human pressing the button every time.
- 2
The review queue becomes data
Every override is a labelled example you already paid for.
- 3
Measure the approval rate
The evidence that would justify removing the human.
- 4
Automate the confident band
Narrowly, expanding only as the numbers support it.