By Mehdi HassanMost writing on AI interaction models argues for one side: chat is dead, agents are the future, copilots are the sensible middle. In practice, the right choice depends on the task, and many products need all three. This comparison puts them side by side and gives a verdict for each common use case. It builds on our guide to AI product UX.
What is the difference between a copilot, an agent and a chatbot?
A chatbot answers questions in a dedicated conversation. A copilot works beside the user inside their current task, aware of what they're looking at, and suggests or drafts. An agent receives a goal and performs a sequence of actions to reach it, often across tools, while the user supervises.
- Chatbot: the user asks, the AI answers. Context is mostly what the user types or attaches.
- Copilot: the user works, the AI helps. Context is the document, record or screen the user is in.
- Agent: the user delegates, the AI does. Context is the goal plus whatever tools and data it can reach.
The term "copilot" is also a product name for some companies. In this article, it means the general pattern: an assistant embedded alongside the user's work.
How do copilot vs agent vs chat compare on UX?
Copilot, agent and chat differ most in who controls the next step, how much context the AI has, how long the work takes, and what happens when it's wrong. This table summarizes the trade-offs.
Factor | Chat | Copilot | Agent |
|---|---|---|---|
Who drives | User, message by message | User, with suggestions | AI, user supervises |
Where it lives | Separate chat screen or panel | Inside the current work | Task view, background runs |
Context | What the user provides | Current document or record | Goal, tools and connected data |
Typical duration | Seconds | Seconds | Minutes to hours |
Cost of a mistake | A wrong answer to ignore | A bad suggestion to reject | A wrong action to undo |
Key UI patterns | Composer, thread, citations | Inline suggestions, side panel | Plans, approvals, logs, rollback |
User effort | High: must know what to ask | Low: help appears in context | Low per task, higher to set up |
Trust needed | Low | Medium | High |
The "trust needed" row explains why many teams start with chat and move toward agents slowly. Every step toward autonomy needs more supervision design. Our guide to agentic UX covers that supervision, and AI chat interface design covers the chat side.
When should you use a copilot?
Use a copilot when the user owns the outcome and works in a complex tool, such as a document editor, a spreadsheet, a design tool or a CRM, and AI help is most useful in the context of what they're already doing. The user stays in charge, and the AI shortens the work.
Pros
- Uses the current context, so users don't have to explain what they're looking at.
- Low risk: the user accepts or rejects each suggestion.
- Fits existing workflows without a new screen to learn.
Cons
- Users may not see what context the copilot is using.
- Suggestions that appear too often feel like interruptions.
- Doesn't save time on repetitive multi-step work the way an agent can.
The smallest form of a copilot is an inline action, like "Rewrite" or "Summarize" on selected text. Our guide to inline AI patterns covers those, and AI loading states covers what users see while a suggestion generates.
When should you use an AI agent?
Use an agent for tasks that are multi-step, repeatable and tedious, where the result is easy to check or reverse: triaging tickets, cleaning data, researching a list of companies, preparing reports. Avoid agents when a single wrong action is costly and hard to undo, unless every risky step needs approval.
Agents shine when the user would otherwise spend a long time clicking through several tools. They fail when the task is ambiguous, the data is unreliable or the user can't easily tell whether the result is right. Before committing, prototype with simulated agent behavior to see whether users actually read plans and approvals. Our guide to prototyping AI features covers Wizard of Oz tests and fake doors for exactly this question.
A quick test: copilot or agent?
Ask three questions about the task.
Does the user need to make judgment calls at most steps? If yes, use a copilot.
Can the result be checked in under a minute, or undone easily? If no, use a copilot or an agent with approval on every risky step.
Does the same task repeat often with similar inputs? If yes, an agent will save the most time.
Two or three answers pointing to an agent is a good sign. Otherwise, start with a copilot and promote the most repetitive tasks to agents later.
When is a chatbot still the right choice?
A chatbot is still the right choice for open-ended questions, research and exploration, customer support, and products where users don't know in advance what they need. It is a poor choice as the only interface for structured tasks that a button or a copilot suggestion would handle in one click.
Chat works well when it shows its sources, so users can verify answers. Our guide to citations in AI answers covers that pattern, and AI error handling covers what happens when answers go wrong.
Which AI interaction model fits each use case?
The right interaction model depends on the use case. This table gives our verdict for the cases we see most often in SaaS and AI products.
Use case | Best model | Why |
|---|---|---|
Writing and editing documents | Copilot | User owns the text, AI helps in place |
Customer support on a website | Chat | Questions are unpredictable |
Research across many sources | Chat with citations, or an agent | Chat for questions, agent for full reports |
Bulk data cleanup | Agent with batch approval | Repetitive, checkable, undoable |
Sending outreach emails | Agent drafts, user sends | Sending is irreversible |
Analytics questions | Copilot or chat | Answers in context of the dashboard |
Coding assistance | Copilot, with agents for scoped tasks | Developers want control and speed |
Onboarding a new user | Copilot hints | Help in context beats a chat window |
As you add more than one model, keep them consistent with shared components. Our guide to design systems for AI products covers this.
How do you explain the choice on your website?
Explain your interaction model on your website by showing it, not naming it. Buyers care less whether you call it a copilot or an agent and more about what it does, where they stay in control, and how it fits their workflow.
Show a real task from start to finish, with the plan and approvals if it's an agent. Our guides to designing an AI agent website and explaining AI products on your website cover how. Once the site is live, tracking AI referral traffic in GA4 shows whether answer engines are sending buyers your way.
