Chatbot or agent? The difference is action.

A chatbot answers. An AI agent acts. You send a chatbot a message, it writes a reply, and then it waits while you do the work. You give an agent a goal, and it does the steps itself. It searches, opens files, runs code or clicks through a website, checks the result, and keeps going until the job is done or it needs you.

The model underneath is often the same. What changes is what the AI is allowed to touch, and how many steps it takes before a person looks. That is the real AI agent vs chatbot difference, and it decides what each one is good for, what it costs and what can go wrong.

01What is the difference between an AI agent and a chatbot?

A chatbot is like calling a knowledgeable friend. They tell you what to do, then you hang up and do it. An agent is like handing that friend your laptop and a to-do list. They do the steps and call you back when the work is done, or when they hit a question only you can answer.

Three ingredients turn a chatbot into an agent:

• Tools. The ability to do things outside the chat: search the web, read and write files, run code, use an app, move a mouse.

• A loop. Instead of one reply, the model takes a step, checks the result and chooses the next step, many times over.

• A goal with a finish line. “Make the failing test pass” has an end state the agent can check. “Tell me about testing” does not.

Take away any one of them and you are back to a chatbot, however smart the model is. How much it remembers from one session to the next is a separate question, covered in our guide to AI agent memory.

02AI agent vs chatbot, side by side

The worst-case row is the heart of it. A chatbot’s mistake sits on your screen until you act on it. An agent’s mistake has already happened by the time you see it.

03Copilot, agent, autopilot: a dial, not two boxes

Real products rarely sit at one extreme. Most let you choose how much the AI does before it checks in with you. It helps to picture four settings on one dial.

Each step up the dial saves you more time and hands over more risk. The right setting depends on the task, not the product. The same agent can run on autopilot for sorting files and on tight approvals for anything that sends email or spends money.

04Three requests, handled both ways

The clearest way to see the difference is to give both the same request.

“What does this clause in my lease mean?”

Chatbot: You paste the clause. It explains it in plain words, flags the part about leaving early, and suggests two questions for your landlord. Done in seconds.

Agent: It would do much the same, perhaps after searching your local tenancy rules. That is more time and cost for little gain.

Verdict: chatbot. When the answer is the product, you do not need an agent.

“The signup form stopped sending confirmation emails. Fix it.”

Chatbot: You paste the code and the error. It suggests a cause and a patch. You apply it, hit a second problem and paste again. A few rounds later, it works.

Agent: It opens the project, reproduces the bug and reads the logs. It finds an outdated email setting and a missing check, fixes both, runs the tests, and shows you the change to approve.

Verdict: agent. It can run the code, so it tests its own guesses instead of handing them to you to test. Our first-agent tutorial walks through a similar tool-use loop in code.

“Put last month’s 40 receipts into a spreadsheet for my accountant.”

Chatbot: It can read receipts you upload and list the amounts. You still gather the files, check each line and build the sheet yourself.

Agent: It reads the receipts folder, pulls the date, vendor, amount and tax from each one, and builds the spreadsheet. Then it lists the three receipts it could not read, so you can check them by hand.

Verdict: agent, with a check at the end. It does the grind, and you review the flagged items before anything reaches your accountant. Our guide to automating repetitive tasks has more jobs like this, with prompts you can copy.

05When to use a chatbot and when to use an agent

A simple test: look at what you would do with the answer. If you would read it and be done, use a chatbot. If you would copy it somewhere and then do five more steps, those steps are the agent’s job.

Pick a chatbot when:

• You want to understand something, brainstorm or get a first draft.

• The task is a single step, or every step needs your judgment.

• You want to be the one who acts, as with a delicate message to a client.

Pick an agent when:

• The job has many steps you know how to do but would rather not.

• Success is easy to check: the tests pass, every file is renamed, every row is filled.

• The work lives where the agent can reach it: your files, your code, a website, your apps.

For ten ready-made examples like this, from inbox triage to a bug-fix loop, see our list of agentic workflows you can run today.

A good agent brief reads differently from a chat message. It states the goal, the finish line, the limits and when to stop and ask.

Goal: Put last month's receipts into a spreadsheet for my accountant.
Where: The files are in Documents/Receipts/August.
Done when: There is one row per receipt with date, vendor, amount, tax and category, plus a total.
Limits: Read the receipts, but do not move, rename or delete them. Do not email anyone.
Ask me: If a receipt is unreadable or looks like a duplicate, list it instead of guessing.

06What agents cost, and what can go wrong

Agents do more work per request, so they use more tokens: the small chunks of text a model reads and writes, and the unit AI usage is billed in. Anthropic shared its own numbers when it explained how it built Claude’s Research feature, in which several agents work together:

That is a good trade when the task is worth it, because an hour of your time usually costs more than the extra compute. It is waste when a one-line answer would do. If cost matters, a smaller model for simple jobs is the easiest saving, and our guide to choosing an AI model explains the trade-offs.

Risk grows the same way. A chatbot that gets something wrong gives you a bad paragraph. An agent that gets something wrong can delete the wrong folder, email the wrong person or buy the wrong ticket. Mistakes also compound, because an early wrong guess feeds every step after it.

There is one risk chatbots barely have. An agent reads web pages, emails and documents while it works, and some of that text may be written to trick it. This attack is called prompt injection, and it is one reason well-designed agents pause before sensitive steps.

We cover the full risk picture, and the settings that reduce it, in is it safe to let an AI agent control your computer?

07FAQ

Read next: what an AI agent is and how it works, or how computer use agents click and type for you.