What is the difference between a chatbot, copilot and agent?

Ask most people what they are using when they type a question into ChatGPT, and they will say a chatbot. Ask someone who uses Microsoft Copilot in Word, and they will probably say the same. They are not wrong, exactly, but they are using a word that has stretched to cover three genuinely different things.

Chatbot, copilot and agent are now used almost interchangeably in marketing material and headlines, and the blurring matters because each type of AI tool works differently, is suited to different tasks, and carries different risks. Understanding the distinction is not about being technically precise for its own sake. It changes what you should expect from the tool, how much you should trust it, and what you should never ask it to do unsupervised.

A chatbot, in the most straightforward sense, is a conversational AI that responds to questions and instructions with text. You put something in, you get something back. The exchange stays within that window. A chatbot does not go and do things in other applications on your behalf. It does not send your emails, update your calendar, or book anything. It generates a response, and then you decide what to do with it. ChatGPT in its basic form, Claude, and the standard version of Google Gemini are all chatbots in this sense, even though the quality and capability of the underlying models has grown enormously. The defining feature is that the action stops at the text they produce.

That changes the nature of the risk significantly. If a chatbot gives you a wrong answer, you can spot the error before it causes any harm, provided you are paying attention. The AI does not do anything; it tells you something. What happens next is your decision.

The word copilot describes something more embedded. Rather than sitting in its own interface and answering general questions, a copilot is integrated into a specific application and helps you work within it. Microsoft Copilot in Word can suggest rewrites, summarise a document, and draft new paragraphs while you are working. GitHub Copilot suggests code as you type in a development environment. Google Workspace’s AI features can draft emails directly in Gmail, or pull relevant data into Sheets as you build a spreadsheet. The copilot is working alongside you inside the tool you are already using, not in a separate conversation window.

The difference in how you experience this is significant. A copilot’s suggestions land directly in your document, your inbox, or your codebase. You are still in control of whether to accept or reject them, but the outputs land somewhere real rather than in a chat window you can close. This is why many organisations that have been relaxed about employees using standalone chatbots are more cautious about copilot tools: what the AI produces goes somewhere, and the friction between AI suggestion and actual output is lower.

The third category, the AI agent, is where things get genuinely different. An agent does not just answer questions or assist within a single application. It takes action across multiple tools, systems and services in order to complete a task that has been given to it. Where a chatbot talks and a copilot assists, an agent does. You might tell an agent to research a topic, compile the findings into a document, send a summary to a list of contacts, and book a follow-up meeting. The agent would handle each step in sequence, switching between tools as required, without a human in the loop at every stage.

This is the version of AI that is generating the most excitement in 2026, and also the most legitimate concern. The ability to act, rather than simply respond, makes agents far more powerful and far more capable of causing harm if something goes wrong. A chatbot that hallucinates gives you a wrong answer. An agent that hallucinates might send an email to the wrong person, delete the wrong file, or submit a form with inaccurate information. The stakes are higher because the AI is the one pressing the button.

Current agents vary enormously in how capable and how reliable they are. The best of them handle focused, well-defined workflows reliably, particularly where the steps are predictable and the consequences of a mistake are recoverable. They are less reliable for long-horizon tasks that require judgement at multiple points, and they can struggle with anything that requires understanding ambiguity or knowing when to stop and ask a human. The marketing around AI agents in particular tends to outrun the reality, which is worth bearing in mind when assessing claims.

The practical question for anyone choosing between these types of tool is how much action they want the AI to take, and how much oversight they want to maintain. For someone using AI to research, draft and think through ideas, a good chatbot is often all they need. For someone who wants AI woven into the tools they use every day, a copilot makes more sense. For someone who wants the AI to handle a multi-step workflow with minimal intervention, an agent is the right category, but it requires more careful setup, clearer boundaries, and a genuine understanding of where human oversight needs to stay in the loop.

None of this means one type is better than the others. They are suited to different tasks at different levels of automation. The mistake is treating them as synonyms when they represent genuinely different levels of trust, capability and risk. The product names do not help: Microsoft calls its AI copilot even when it is edging into agent territory. OpenAI offers both chatbot and agent functionality under the same ChatGPT brand. Google is doing the same. The label on the box is less useful than understanding what the tool is actually doing on your behalf.

When you know whether you are talking to a chatbot, being assisted by a copilot, or giving instructions to an agent, you know how much to verify, how much to supervise, and where to keep your hands on the wheel.