AI Agents vs Chatbots
Agents vs Chatbots — and Why Most People Asking Need Neither Yet
One says things, the other does things. That is the whole definition and it takes a sentence, which is why this page spends almost no time on it. The useful question is which your task needs — and the honest answer for a large share of the people searching this is that the task in front of them is not ready for either, because nobody has yet written down what the correct answer is.
Chatbot vs AI agent
It answers · It acts
- What it does
Replies to a question with information.Takes an action on your systems - The worst thing it can do
Say something wrong, which somebody may quote back at you.Do something wrong you must unwind - What has to be decided first
What it says when it does not know.What it may do without asking - Who signs it off
Whoever owns the answers — usually marketing or support.Somebody who can authorise spending - How you find its mistakes
Read the conversations. They are all text and all reviewable.A decision record, or archaeology - Realistic first build
Two to four weeks, narrow subject, live behind a handover.Longer, and it starts with zero authority - When it is the wrong choice
When the answer needs judgement or is not written down anywhere.When being wrong is expensive and rare
The row that decides most projects is the third. Both fail in the same way and for the same reason — somebody has to have written down what the correct behaviour is, and in most businesses that has never been done for the task in question. Until it is, you are choosing between two systems that will both confidently do the wrong thing, one in words and one in actions.
The real answer is usually neither, yet
A large share of the people comparing these two are looking at a task where nobody in the business has written down what a correct outcome looks like. That is not a failing — it has never needed writing down, because a person did it and used judgement. But it means neither option is buildable yet, and the useful next step is an afternoon spent listing what the right answer is for twenty real cases.
Where something IS written down, the choice becomes easy and it is not really about the technology. If the task ends with somebody knowing something, you want the thing that says it. If it ends with something happening in a system, you want the thing that does it. The two questions almost never both apply to the same task, which is why the comparison feels harder than it is.
What we would push back on is the assumption that an agent is the more advanced choice and therefore the better one. An agent is the more consequential choice. It is not more sophisticated to have software send emails on your behalf; it is riskier, and the sophistication is entirely in the permissions and the escalation, not in the model. A well-scoped chatbot that hands over cleanly beats a badly scoped agent every time and costs a fraction.
And plenty of real builds are neither, or both. A workflow with one decision point in it is not an agent, and calling it one changes nothing except the invoice. If somebody is selling you an agent for a task with no decisions in it, they are selling you a name.
An agent is not the more advanced choice. It is the more consequential one, and the sophistication is all in the permissions.
A three-question test that settles it
Answer these about the specific task, not about your business in general.
Is there one correct answer?
If the answer depends on judgement that lives in somebody’s head, neither is ready. Write down the right outcome for twenty real cases first — that document is the actual first deliverable.
Does the task end in knowing or doing?
Ends with somebody informed — a chatbot. Ends with a record changed, an email sent, a slot booked — an agent. The two rarely both apply.
What does being wrong cost?
Cheap and visible: either is fine. Expensive or hard to notice: keep a person in it, whichever you build, and put the ceiling in writing.
Do you have the answers written down?
A chatbot needs a small set of approved answers. An agent needs approved permissions. Neither needs a model you chose carefully — that is the least important decision in the project.
Is this actually a workflow?
A defined path with no decisions in it is a workflow, and calling it an agent changes the price rather than the thing. Most requests that arrive asking for an agent turn out to be this.
Could it be both?
Frequently. A chatbot that answers, and hands anything actionable to an agent step with narrow permissions, is a common shape and easier to trust than either alone.
What people ask about this choice
Is an AI agent just a better chatbot?
No, it is a different thing with a different risk. A chatbot answers, and the worst it can do is say something wrong. An agent acts on your systems, and the worst it can do is take an action you then have to unwind. That is not an upgrade — it is a decision about what software is allowed to do without asking anybody.
Which is cheaper to build?
A chatbot, usually by a wide margin, because most of its cost is assembling a small set of approved answers. An agent costs more because most of its cost is not the build at all — it is deciding and documenting what it may do, and the review process that lets you widen that safely.
We were told we need an agent. Do we?
Ask what decisions it makes. If the task follows a defined path with no judgement in it, that is a workflow, and calling it an agent changes the price rather than the thing. A large share of requests that arrive asking for an agent turn out to be a workflow with one decision in the middle.
Can we start with a chatbot and upgrade later?
Yes, and it is usually the right sequence. Running a chatbot for a few months tells you what people actually ask, which is exactly the input an agent needs and the thing nobody has when they start. It is also the cheapest way to find out whether the task has one correct answer.
What if we cannot decide?
Then the task probably is not ready. The test is whether somebody can write down the correct outcome for twenty real cases from your own records. If they cannot, that document is the first deliverable — not either system.
Does the model matter?
Less than almost anything else in the project, and considerably less than whoever is selling you one will suggest. What decides whether either works is what it draws on, what it may do, and where it hands over — none of which is a property of the model.