AI agent vs chatbot: the difference, and which one you need

A chatbot answers. An AI agent answers and acts. A chatbot holds a conversation and gives information. An AI agent works toward a goal: it decides which steps are needed, uses tools to look things up and do things in your systems, and hands over to a person when it should. A chatbot can tell a customer how to return an item. An agent can start the return.

The terms are used loosely, often by people selling one as the other. This guide sets out the practical differences, what each costs, and how to work out which you need.

#Three things that get called a chatbot

Rule-based chatbot. A decision tree with buttons and keywords. It follows a script someone wrote. It is predictable and cheap, and it fails the moment a customer asks something in words the script did not expect. Most chatbots before 2023 were this kind, which is why chatbots have a poor reputation.

AI chatbot. A language model that answers freely in natural language, usually grounded in your own documents so it answers from your help center, policies and product information and can show its source. It understands almost any phrasing. It still only talks.

AI agent. The same kind of model, given tools and a goal. It can look up an order, change a booking, update a record, send a message or create a ticket, and it works out the steps for itself. It also knows when to stop and ask a person.

#Side by side

#The same request, three ways

A customer writes: "I ordered the wrong size. Can I swap it for a medium, and will it still arrive before the 14th?"

• The rule-based chatbot shows a menu: Orders, Returns, Shipping. The customer picks Returns and gets a link to the policy.
• The AI chatbot understands the question and explains, correctly, how exchanges work and how long shipping usually takes. The customer still has to do it all.
• The AI agent finds the order, checks that an exchange is within policy and that a medium is in stock, works out the delivery date from the shipping option, tells the customer it will arrive on the 12th, and arranges the exchange once they confirm.

Only the third one finishes the job. It is also the only one that needs access to your order system, a rule about what it may change, and testing to prove it gets this right.

#What an agent adds, and what that costs you

It adds:

• Tools. Defined actions such as "look up order", "start exchange", "book slot", each with its own rules.
• Planning. Working out which tools to use, in what order, and what to do when one fails.
• Memory of the task. Keeping track across several steps and several messages.
• Judgment about handover. Recognizing what is outside its authority.

It costs you:

• Integration work. Connecting to your systems is most of the effort and most of the price.
• A new kind of risk. A chatbot's worst case is a wrong answer. An agent's worst case is a wrong action. That is why agents need permissions, limits, approval steps and logs.
• Testing. An agent should be measured against a set of real examples before launch and before every change.
• Running costs. An agent takes several steps per request, so it uses more of the AI model per conversation than a chatbot does.

The costs are covered in detail in how much does an AI agent cost.

#Five questions to work out which you need

1. Do people mainly want information, or do they want something done? Information points to an AI chatbot. Action points to an agent.
2. Does answering need live data? If the honest answer to most questions starts with "let me check", you need at least read access to your systems, which is the first step toward an agent.
3. Are the steps the same every time? If yes, you may need neither. Plain automation is cheaper and more predictable.
4. What does a mistake cost? Where a mistake is cheap, an agent can act alone. Where it is expensive, it should prepare the action and a person should approve it.
5. What volume are you dealing with? A few requests a week rarely justify an agent. Hundreds or thousands usually do.

A rough guide from those answers:

#A sensible path from one to the other

You do not have to choose once and for all. A path that works well:

1. Start with an AI chatbot grounded in your content. It is quick to set up, and the conversation logs show you what people actually want.
2. Read the logs. Look for the requests where the chatbot explained how to do something that it could have done itself.
3. Add the most common action, with a narrow rule and an approval step.
4. Remove the approval step for that action once it has been right for a few weeks.
5. Repeat for the next action.

Each step is small, measurable and reversible. The knowledge base and the evaluation examples carry forward, so nothing is thrown away.

#How to spot a chatbot sold as an agent

"Agent" is the fashionable word, and plenty of products have been renamed without being rebuilt. Ask these questions:

• What actions can it take in my systems, and can you show me one? If the answer is about conversation quality, it is a chatbot.
• How are permissions and limits set for each action?
• What happens when an action fails halfway?
• Where do I see a log of everything it did?
• How is accuracy measured before a change goes live?

Clear answers to all five mean a real agent. Vague ones mean a chatbot with a new label, which may still be what you need, at a chatbot price.

#Where to start

If you are unsure which you need, write down the ten requests you get most often and mark each one "tell" or "do". Mostly "tell" means an AI chatbot, and many help desk tools include a good one. Mostly "do" means an agent. Either way, a free 20-minute call is enough for us to tell you which, and whether a product you already own can do it. When a custom build is the answer, that is our AI agent development service, from $9,800 as one fixed price.

For the wider picture, see AI agents for business.

Common questions.

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

  A chatbot holds a conversation and gives answers. An AI agent works toward a goal: it decides which steps are needed, uses tools to look things up and take actions in your systems, checks the result, and hands over to a person when it should. Put simply, a chatbot can tell a customer how to return an item, and an agent can start the return.

• Is ChatGPT a chatbot or an AI agent?

  In its basic form it is an AI chatbot: it answers from what it learned in training and what you type. When it is given tools, such as web browsing, code execution or connections to your apps, and works through a task in several steps, it is behaving as an agent. The same underlying model can power either. What makes it an agent is the tools and the autonomy around it.

• Are AI agents more expensive than chatbots?

  Usually, yes. A rule-based chatbot can cost very little. An AI chatbot that answers from your documents typically costs a few thousand dollars to set up properly, or a modest subscription. An AI agent costs more because it has to be connected to your systems, given permissions and tested, typically $8,000 to $40,000 for a custom build. It also does far more.

• Can a chatbot be upgraded into an agent later?

  Often, and it is a sensible path. Start with an AI chatbot that answers from your content. Watch what people ask it to do that it cannot. Then add the one or two actions that come up most, each with its own permissions and limits. The knowledge base and the conversation logs from the chatbot stage carry straight over.

• Do I need an AI agent, or is automation enough?

  If the steps are the same every time, automation is enough, and it is cheaper and more predictable. A new form submission that creates a record and sends an email needs no agent. You need an agent when the input is messy language, the steps differ from case to case, and something has to be decided along the way.