AI agent vs chatbot: what is the difference for a business?

"AI agent" is the most overused phrase in software this year, and a lot of what is sold under it is a chatbot with a new label. The difference matters, because the two solve different problems, cost different amounts to build safely, and fail in different ways.

The short answer

A chatbot answers questions. You ask, it replies, and a person decides what to do with the reply.

An AI agent answers and then acts. It can look up a customer in your CRM, update a deal, draft and send an email, book a meeting or raise a support ticket — using tools you have given it permission to use.

If your problem ends with someone reading an answer, you need a chatbot. If it ends with someone clicking through three systems to do something with that answer, you may need an agent.

What a chatbot is good at

Modern chatbots are built on large language models and, done well, are grounded in your own content — help articles, policies, product documentation — so they answer from your facts rather than the model's general knowledge. This is usually called retrieval-augmented generation, or RAG.

Good fits for a chatbot:

• Answering product and policy questions on your website or WhatsApp
• Helping staff find the right internal document or procedure
• First-line support that hands over to a person with the conversation attached

The risk is limited: the worst a chatbot does on its own is give a wrong answer. That is why a chatbot is the right first step for most businesses.

What makes something an agent

Three things separate an agent from a chatbot:

• Tools. The agent can call functions — read a CRM record, create a ticket, check stock, send a message.
• Planning. Given a goal ("follow up with everyone who asked for a quote last week"), it works out the steps rather than following a fixed script.
• Acting on the result. It uses what one step returns to decide the next one.

That is powerful, and it changes the risk completely. A wrong answer is embarrassing. A wrong action — an email sent to the wrong customer, a discount applied to the wrong deal — costs money.

Where agents earn their place

Agents pay off on work that is repetitive, multi-step and spread across systems:

• Sales follow-up: reading new enquiries, enriching them, creating the lead in the CRM and drafting the first reply for a person to approve
• Support triage: classifying incoming requests, pulling the customer's history and routing the case with a suggested answer
• Operations: chasing missing documents, updating order status, reconciling records between two systems

The common thread is that a person already does these steps by hand, the same way, many times a day.

How to keep an agent safe

The difference between a demo and something you can trust in production is mostly guardrails:

• Least permission. Give each agent only the tools its job needs. A support agent that can read orders does not also need to issue refunds.
• Approval steps. Anything risky or irreversible waits for a person to approve it. Loosen this only where the logs show it is safe.
• A full audit log. Every step the agent took, and why, recorded where your team can review it.
• Evaluation before launch. A set of real past requests the agent must handle correctly before it touches live data — re-run on every change.

Which one should you build first?

For most teams the honest answer is: start with a grounded chatbot, measure what people ask, then turn the most repetitive follow-up actions into an agent with approvals. You learn what your users actually need before you hand software the keys.

If you already know the workflow you want to automate, we can map it with you and tell you whether it needs an agent at all. See how we approach AI agent development, or our wider AI and machine learning work.