Chatbot or agent? The difference isn't the interface

The input box looks the same either way, which is why most companies get the chatbot or AI agent question wrong. The difference is not the interface, it is the responsibility. One of them answers you. The other one goes and does something inside your systems. Mix those up and you either get too little, or you let something loose that nobody is checking.

What is the difference between a chatbot and an agent?

A chatbot talks. An agent acts. A chatbot answers questions from material you handed it in advance, and when the conversation ends, nothing in your systems has changed. An agent carries out steps: it looks things up, assembles, writes, sends. So with a chatbot the stake is the content, with an agent it is the consequence.

The interface blurs all of this. Input box, typing, answer: at a glance the two are identical. The difference sits behind them, where the system touches your world. At worst, a chatbot says something wrong. At worst, an agent does something wrong, and a correction no longer covers it.

Our own site is a concrete example. It has a chatbot, and today it stands in an informational state. When it does go live, the server side rule is that it may state facts only from an approved corpus, and it never gives a price, a deadline or a client count. That is where a chatbot's responsibility ends, and no further. What it says has to be true, and where there is no source, it should say nothing.

For an agent that bar is not enough. There the question is not whether the sentence is true, but whether the step can be undone, whether you can see afterwards what happened, and where the system stops when it is unsure. That is what building agentic systems is about: not smarter prose, but a logged workflow you can halt.

If you want to decide it in one sentence, ask yourselves this: after the conversation ends, is there a trace of it anywhere in your systems? If there is none, you are talking about a chatbot. If there is, you are talking about an agent, even when it shows up to the visitor in the same friendly speech bubble.

When is a chatbot enough?

When most of the questions are the same ones, and the answers already exist in writing somewhere. Opening hours, delivery, warranty, spec sheets: nothing has to be carried out here, the answer only has to be found and said clearly. If your support time goes on repeat questions, a chatbot takes exactly that layer off.

The strength of a chatbot is not in the model, it is in what you hand it. We call that the corpus: an approved body of text the bot may work from, and outside of which it may not speak. Without it, the bot talks about your company out of its own general knowledge, and from then on you do not know what your customer is hearing.

Three signs that a chatbot is the right answer for you right now:

• The question repeats. The same handful of topics comes back week after week, and a person types every one of them out again.
• The answer exists. It is written down in a document, a spec sheet, a contract template, it is just hard to find.
• Nothing has to happen at the end. The conversation succeeds because the person understood, not because something started running in a system.

In that case AI customer service is not an IT project, it is editorial work. The hard part is getting your company knowledge base in order, and the bot is only the surface on top of it. And if the corpus is missing, better to say so up front: you need that step even if you end up building an agent.

A well written corpus has a side effect: it is not only good for the bot. The same text goes to a new colleague in their first week, and the same answer goes out by email. The bot does not manufacture new content here, it just makes available what has been living in three people's heads.

When do you need an agent?

When something still has to be done after the answer, and today a person retypes it from one system into another. Putting a quote together, entering data, filing, updating a status, sending it around. An agent does that work, logs every step, and stops where a human decision is needed.

An agent pays off when there is a process you can write down step by step, and that someone runs by hand several times a week. The question is not how clever the model is, but whether the process can be stated. What is not stated, an agent cannot carry out either. It will only act as if it had.

Automation here does not mean the machine takes over everything. It means the machine does the mechanical part and puts the decision points in front of a person. With a properly built agent you can say at any time what it did, from which data, and how to undo it. If those three are missing, it is not an agent, just a button with nobody behind it.

This is why we suggest starting with a single agent, on the dullest and most repetitive task you have. One working agent teaches you more about your own company than any survey, and the second one builds on that. We wrote about it separately: how to pick your first AI agent.

Writing the process down is the harder half of the work, and that is good news, because it is the half you know best. One page is enough: what starts it, what data it works from, what comes out at the end, and the few cases where it has to go to a person.

What does choosing wrong cost you?

The first cost is not money, it is work handed back to people. Put a chatbot where an operation was needed, and your team still does everything by hand, only now they clean up after the bot too. Put an agent where an answer would have been enough, and unchecked steps run where nobody has time to look.

The chatbot in the wrong place is the common case. It goes live, it answers, and then at the hard question it hands the conversation to a person who no longer knows what was said before. The work did not disappear. It just got a waiting room.

The other direction is rarer and costs more. If an agent can reach your systems but has no log, no permissions and no stopping point behind it, you notice the fault not where it happened but where the consequence showed up. At that stage it can rarely be patched. It has to be rebuilt, because nobody can tell where it went off.

There is a third cost people rarely name: the trust you lose inside the company. If the first attempt at AI looked like a failure to your colleagues, the next one has to be sold to them, not to a supplier. That is why it is worth starting with the small, checkable step, even when it promises less.

The hardest item to win back is the time that checking eats. Somebody always ends up standing next to an unreliable system to verify its output, and one job quietly becomes two. That cost never shows up on an invoice, but it sits in the calendar every week.

Can you have both?

You can, and in most places that is the right order. The chatbot comes first, because it forces your answers to be written down in one place. Agents then sit on top of that ordered knowledge. The other way around rarely works: over messy knowledge, an agent makes bad calls fast.

In practice the two are layers of the same system. Written knowledge underneath, and above it whatever reaches for it: one only reads and repeats, the other works with it. The same corpus serves both, so you never have to build it twice.

The line between them has to be drawn sharply, and it is better said on day one:

• The talking part touches nothing. If an answer has a side effect, it is no longer an answer, it is an operation, and it gets its own permission.
• The acting part invents no facts. Whatever an agent writes into a system has a source behind it, not an inference.
• A person sees both. In one place, after the fact too: what was asked, what was said, what was done.

That is the order we follow as well. On the site the chatbot is the talking layer, built on a single approved corpus, and the acting layer is meant to sit on that same written knowledge. So the choice is rarely chatbot or AI agent forever. It is much more about which one you start with, and what the other can later be built on top of.

The order has one very practical benefit too. While the chatbot is working, you find out what visitors and colleagues actually ask, and that list is more accurate than any planning session. Most agent ideas come from there: from the places where the bot keeps saying that this one needs a person.

We do not write down anything that has no source.