AI agent, chatbot or assistant? The difference in one table

A chatbot talks, an assistant helps you do a task while you make the calls, and an agent does the task itself, choosing its own steps. The model underneath can be identical. What differs is who decides what happens next. For how an agent works inside, see What is an AI agent?.

The table

The rows below follow how the companies building these systems describe them, mainly Google Cloud's comparison of agents, assistants and bots, and OpenAI's guide to building agents.

Google's version of this table sorts the three by autonomy. Its bot follows pre-defined rules and responds to triggers or commands. Its assistant is reactive, and the user makes decisions. Its agent is proactive and goal-oriented, and can make decisions independently.

OpenAI draws the line in a similar place. In its guide to building agents, an agent uses a language model to control how a workflow runs, picks tools based on where the task stands, and can stop and hand control back to the user if it fails. An app that uses a language model without letting it steer the work is not an agent, in OpenAI's terms.

Where the lines blur

The table is tidy. Products are not.

• One product, two modes. OpenAI's ChatGPT agent, launched on 17 July 2025, sits inside the same ChatGPT chat window. OpenAI says you can move from a plain conversation to asking for actions in the same chat. The tools available to it include a visual browser, a text browser, a terminal and connectors to apps such as email.
• Agents that ask permission. OpenAI says ChatGPT agent asks before actions with real-world consequences, such as a purchase, and some tasks, such as sending email, need you to watch. You can interrupt or take over at any point. So even a product named "agent" hands some decisions back to the person.
• Workflows that look like agents. Anthropic separates "workflows", where code written in advance sets the order of steps, from "agents", where the model directs its own process and tool use. Anthropic also notes that some customers use "agent" only for fully autonomous systems that run independently for long periods.
• Assistants called agents. Microsoft's documentation for Microsoft 365 Copilot says agents work as "specialized AI assistants" for particular domains. Its simpler "declarative" agents do not support proactive interactions and rely on the user to start them. Its "custom engine" agents can trigger actions without direct user input.

How the words are used in marketing

Three patterns show up in the companies' own pages:

1. "Agent" and "assistant" share a shelf. Google Cloud's agent page lists integrated AI assistants and pre-built AI agents side by side in one product portfolio. Microsoft calls Copilot add-ons "agents" even when the user starts every interaction.
2. "Assistant" can mean an API, not a product. OpenAI's Assistants API was a developer interface, not a consumer assistant. OpenAI's documentation says it was sunset on 26 August 2026, and points developers to the Responses API instead.
3. "Chatbot" serves as the baseline. OpenAI's guide uses it as the example of an LLM app that does not count as an agent. Anthropic describes customer support tools as pairing a familiar chatbot interface with tool integration, so a chat window can sit on top of an agent.

Our take: the word on the box tells you which trend the product launched into, not how much it will do on its own. Two products both called "agents" can differ more than an agent and an assistant do.

How to tell which one you are using

Ignore the name and check three things:

1. Can it act outside the chat? If it only writes text for you to copy, it is a chatbot or an assistant, whatever it is called.
2. Who presses the final button? If you approve every action, it behaves like an assistant. If it sends, buys or deletes on its own, it is an agent, and its mistakes are actions.
3. Can it run when you are not there? Microsoft's split between agents that need a user to start them and agents that can trigger themselves is a useful test for any product.

If a product passes all three, treat it like a new member of staff with access to your accounts: give it only the connections the task needs. OpenAI's own advice for ChatGPT agent is to turn off connectors you are not using.