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

A chatbot talks: it takes a question and returns an answer. When the conversation ends, nothing has been done. An AI agent executes: it reads data, decides, writes to third parties, waits for replies, chases the ones who go quiet and leaves a result inside a system. The test that separates them is to ask what is left behind when nobody is watching. If the answer is a transcript, it is a chatbot. If it is a quote put together, a CRM record updated or a supplier comparison ready to review, it is an agent.

When do I need an AI agent instead of a chatbot?

You need an agent when the work does not end with the answer. If after the conversation somebody still has to copy details into a spreadsheet, email suppliers, chase replies or assemble a document, that follow-on work is what costs you money, and it is exactly the part a chatbot never touches. A chatbot is enough when the real problem is that nobody answers quickly enough to questions that already have a known answer.

Does an AI agent need human supervision?

Yes, but at one specific point, not everywhere. The shape that works is for the agent to run the mechanical part on its own, collecting, cross-checking, writing, chasing and sorting, and then stop at the decision that commits money or reputation: awarding a supplier, approving a quote, confirming a booking. An agent with no control point at all is a black box, and a black box cannot be corrected when it gets something wrong.

How much does an AI agent cost for a small business?

It depends on the process, not on the size of the company. The honest way to frame it runs the other way round: first measure how many hours a month the manual process eats and what that hour costs you. A process that burns 80 hours a month justifies an investment that a four-hour one does not. If nobody has measured those hours, the first step is not to ask for a quote. It is to measure them.

AI agent or chatbot: which one does your business need?

Two node structures: on the left a closed loop that returns to its starting point, on the right a network branching out to several destinations

Most conversations about this start badly, because they start with the technology. Which model, which vendor, GPT or Claude. That is not the question.

The question is duller and far more useful: when this thing finishes talking to someone, what is done?

The lunch-break test

Imagine you install the tool and go to lunch. You come back and look at what happened.

If what you find is a transcript of conversations — people who asked things, answers that were given — you have a chatbot. It did its job: it replied. That is all.

If what you find is a quote put together, a record created in the CRM, six emails sent to suppliers and three replies already lined up in a table, you have an agent.

The distinction that matters

A chatbot moves information. An agent moves work. If your problem is that people cannot find answers, a chatbot solves it. If your problem is that someone on your team spends the afternoon copying data from one place to another, a chatbot does not touch it.

Why the two get confused

Because they show up the same way: a little chat box in the corner of a website. From the outside they are identical. The difference is what happens after the visitor types.

Given "I need a quote for 200 square metres of flooring", a chatbot says something reasonable and its involvement ends there. Someone on your team will read that conversation tomorrow, open the spreadsheet, look up the price and write the email.

An agent, from that same message, asks for what is missing — exact area, finish, lead time, where it is being installed — checks the price against the table you gave it, generates the quote and drops it into the CRM with the full context. When your salesperson looks at it in the morning, they do not have to start. They have to review and send.

It is the difference between arriving at a blank page and arriving at a draft.

Three questions to tell which one you need

1. After the conversation, does somebody have to do something by hand?

If the answer is no — people ask, they get their answer, done — you need a chatbot, and probably a simple one. Opening hours, delivery terms, how to return an item: that is information, and a chatbot sitting on your documentation handles it well.

If the answer is yes, that "something by hand" is your real problem. And it is the one the chatbot will not touch.

2. Does the work depend on third parties who are slow to reply?

This is the boundary you feel the most. Chasing is agent work, not chatbot work.

Asking twelve suppliers for a price is easy: it is twelve emails. What takes five days is not writing them. It is that four reply the same day, five on the third, and three need chasing twice. That loop of waiting, checking who is missing and writing again is where the time goes, and it is exactly what an agent can do on its own, every day, without anyone having to remember.

3. Is there a decision that commits money?

If there is — awarding a supplier, approving a discount, confirming a booking for twenty people — do not automate it. Automate everything that comes before it, and let the decision reach a person with the numbers already in place.

A well-built agent is not the one that decides on its own. It is the one that makes deciding take two minutes instead of two hours.

A case with numbers

The easiest way to see it is a real process. In KONBUD, our procurement agent for construction firms, the process was this: a bill of quantities comes in for a site, you have to price every material with several suppliers and build a comparison to decide.

One quoting round across twelve suppliers on the same package took five days. With the agent it takes four hours.

The interesting part is where the time actually was. Not in writing the emails: in waiting for them, chasing them and sorting them once they arrived in twelve different shapes — one with the price in the body of the email, one with a PDF attached, one with a counter-offer and unusual terms.

A chatbot would not have moved that number by a single minute, because no part of that work is conversation.

The pattern, without the build

Swap "suppliers" for "hotels", "clinics" or "candidates" and the problem is identical: requests that arrive incomplete, third parties who take their time, and one person in the middle acting as glue. That human glue is what an agent replaces.

When a chatbot is the right answer

You do not always need an agent, and saying otherwise would be overselling.

If your problem is a high volume of repeat questions that you are slow to answer, a chatbot built properly on your documentation solves that and costs considerably less. If you run a shop with a thousand orders a month and half the emails are "where is my parcel", that is a chatbot.

The mistake is not choosing a chatbot. The mistake is buying a chatbot expecting it to hand you back operational hours, then discovering three months later that the team is doing exactly what it did before, except now there are also conversations to review.

How to work out where to start without paying anyone

Before you ask anybody for a quote, do this sum. Take the process that hurts most and answer:

How many times it happens per month. Do not estimate it high: count it for a week.
How long each one takes, start to finish, waiting included.
Who does it and what their hour costs.

Multiply. If it comes out at forty times a month at two hours each, that is eighty hours a month and close to a thousand a year. With that number on the table, the conversation about which tool you need gets a lot shorter.

And if the sum comes out at four hours a month, you already have the best possible answer: do not automate anything yet.