# Chatbot, assistant, agent or copilot? A plain-English guide

> The four words are used loosely. Here is what each one means in practice, how much it is allowed to do on its own, and how to choose.

Source: https://conversationalaidesign.com/guides/chatbot-assistant-agent-copilot.html. Author: Nelson Abacherli Kawakami. Reviewed November 3, 2026.


## The short answer

Ask two questions about any product: **can it only talk, or can it also act in your systems?** And **who is accountable for each action?** The answers place it on the table below.

| | Chatbot | AI assistant | AI agent | Copilot |
|---|---|---|---|---|
| **What it does** | Follows scripts or intents and answers known questions | Understands open language and answers from knowledge, often with retrieval | Works toward a goal over several steps and uses tools to take actions | Works beside a human, drafting, summarizing and suggesting |
| **Who acts** | The customer (the bot tells them what to do) | The customer | The system, within limits you set | The human, who approves and sends |
| **Typical example** | A website FAQ bot with buttons | A help-center assistant that answers from your articles | An assistant that checks an order, verifies eligibility and issues a refund up to a limit | A suggested reply and ticket summary shown to a support agent |
| **Main risk** | Dead ends when the question is new | Confident wrong answers | Wrong or unauthorized actions, and attacks that steer the agent | Over-trusting the draft; quality drifting unnoticed |
| **Learn more** | module 2, module 11 | module 3, module 6 | module 8, module 9, module 10, module 18 | module 3, module 14 |

The words overlap in marketing. A vendor may call a scripted flow an "agent". Judge by what the system can do and who is accountable, not by the label.

## The autonomy ladder

Autonomy is a dial, not a switch. Most useful support systems sit on the middle rungs.

1. **Informs.** Answers questions. Changes nothing.
2. **Suggests.** Proposes an action or a reply; a person decides.
3. **Acts with approval.** Prepares the action and waits for a yes (a refund draft that a person confirms).
4. **Acts within limits.** Completes low-risk actions alone, inside rules the system enforces in code (refunds under a set amount, once per order).
5. **Acts and reports.** Works on its own across many steps and gives you a log to audit.

Move up one rung at a time, only after the rung below has passed evaluation. See module 10 for limits that hold even when the model is wrong.

> **Why 'agent' changes the risk**
> A bot that only talks can embarrass you. An agent that acts can move money, change records and leak data. That is why [the safety module](../modules/18-safety.html) treats agents separately, and why the lethal trifecta cheat sheet exists.

## How to choose

- **Customers mostly need answers.** Build an assistant grounded in your content (module 6). Keep actions out until you need them.
- **Customers need something changed** (cancel, refund, rebook, update). Build an agent with narrow tools, explicit limits and a human approval step for the risky ones (module 8, module 9).
- **People own the conversation** (a team handling complex or sensitive cases). Start with a copilot. It is the lowest-risk way to learn what the model gets right in your domain.
- **Not sure.** Start at rung 1 or 2 and measure. Evaluation tells you when it is safe to move up (module 14).

## How this course uses the words

This course says **assistant** for the product the user talks to, and **agent** for a model running in a loop that can decide to call tools. A single assistant can contain an agent, a deterministic flow and a retrieval step at once. That mix is the normal case, and module 9 explains how to combine them.


## Sources
- Anthropic: Building effective agents (workflows versus agents): https://www.anthropic.com/engineering/building-effective-agents
- OWASP Top 10 for LLM applications (excessive agency): https://genai.owasp.org/llm-top-10/