The difference between an AI agent and a chatbot is what happens after the answer. A chatbot replies to a question, using a script or your documents. An AI agent can also look up an order, book a slot or send an email by calling the tools you give it. If the job ends with an answer, you need a chatbot. If it ends with something changed in a system, you need an agent, with limits and a person to approve risky steps.
A plant shop gets two messages. "Where is my order?" ends with an answer. "Can you move my delivery to Saturday?" ends with a change in the delivery system. That gap is most of the difference.
The three things people mean
1. A scripted button bot. The customer gets a menu: 1 Track order, 2 Returns, 3 Talk to a person. A human wrote every path. It is predictable and cheap to run, and breaks when a customer types something unexpected.
2. A chatbot that answers from your documents. A language model reads the message, finds the relevant part of your return policy or care guides, and replies in normal sentences. It does not touch your systems, so it cannot see this customer's parcel unless you paste the data in.
3. An agent. A language model plus tools. A tool is a function your developer writes, such as "look up order" or "change delivery date". The model picks a tool, gets the result and decides what to do next. Anthropic's documentation says the model "returns a structured call that your application runs" (tool use overview). OpenAI's function calling guide describes the same loop. The plant shop's agent can now say "your parcel left the depot this morning" and move the delivery date.
Your own code sits between the model's idea and the real action, so that is where limits and approval belong. Some products sold as "AI agents" are really type 2. Ask a vendor what the tools are and what each can change.
Side by side
| Button bot | Document chatbot | Agent | |
|---|---|---|---|
| Job it suits | Fixed menus | Questions with written answers | Tasks that change something |
| Ends with | A scripted reply | An answer | A change in a system |
| Needs from you | A designed flow | Up-to-date documents | Documents, system access, an approver |
| Goes wrong by | Dead ends | Wrong or invented answers | Wrong actions |
| Safe default | Always offer a person | Cite the source, offer a person | Limits in code, approval for risky steps |
What an agent needs and how it fails
A document chatbot needs correct documents: if the policy file says 7 days and the shop does 14, the bot says 7 with confidence. An agent needs that too, plus system access and a named person who owns approvals. Give each tool only the access it needs. OWASP's guidance on excessive agency says to limit permissions to "the minimum necessary".
A customer message can also try to steer an agent. OWASP lists this as prompt injection, LLM01 in its 2025 Top 10 for LLM Applications, and recommends human-in-the-loop controls for privileged operations. Sort every tool by risk, enforce limits in the tool's code and not in the prompt, and require approval for anything that moves money or cannot be undone. See how an AI customer service agent runs with human approval.
How to choose
Ask what the job ends with.
- An answer: a chatbot. Few, fixed questions need only a button bot.
- A change: an agent, starting with the safest tools.
- Both: start with the chatbot, add read-only tools such as "look up order", then add actions one at a time, each with a limit and an approver.
For the plant shop: documents first, order lookup second, delivery changes third, refunds last and only under a limit you set. If customers mostly ask what, when and how much, a document chatbot is enough.
On WhatsApp the choice is the same, with extra rules. See WhatsApp AI agent for business and how to create a WhatsApp chatbot.
Our concept project Pickup is an AI phone agent that books jobs and hands emergencies to a person. It is concept work, not a client project.
Not sure which you need? Tell us about your project.
Frequently asked questions
What is the difference between AI agents and chatbots?
A chatbot replies, using a script or your documents. An AI agent can also act, by calling tools you give it. It needs permission limits and human approval for risky steps.
Is ChatGPT an AI agent?
ChatGPT is mainly a chatbot, but it also has features that carry out tasks with tools, and OpenAI has offered agent-style modes whose names have changed over time. The test for any product: can it take actions in other systems for you, or only reply?
Is an AI agent like a bot?
An agent is a kind of bot, but not every bot is an agent. "Bot" covers scripted menu bots, document chatbots and tool-using agents. Only an agent's language model chooses its next step and calls tools.
Which should a small business start with?
A document chatbot, if most questions have written answers. It carries the least risk. Add an agent when something must change, one action at a time with limits. See our guide to AI customer service agents.