Agent or chatbot. The difference isn't how intelligent it is, it is what it is allowed to do.
It is the most expensive confusion of the moment. Two things that look like close relatives are sold under the same name, have prices that differ by an order of magnitude, and fail for opposite reasons. The line separating them is sharp and can be recognised with a single question.
A chatbot produces text. An agent produces a change. At the end of a conversation with a chatbot there is an answer. At the end of an agent's work there is a line written in management software, an email in draft, a document generated.
Four capabilities make an agent: retrieving from authorised sources, deciding within rules written beforehand, carrying out an action, and stopping when evidence is missing.
The fourth is the one nobody sells, and it is the most important. A system that can't abstain isn't autonomous, it is just quick to get things wrong.
You almost always need less than what you are offered. If the step is always the same, a script costs less and breaks less.
The definitions one by one are in the glossary of AI and automation. Here what matters is the choice, that is, which of the two to ask for and how to notice you are being sold one passed off as the other.
The question that separates the two
After the system has answered, does something have to happen?
If the answer is no, you need a chatbot, and a good chatbot costs little. A customer asks for opening hours, return conditions, whether a product contains an allergen. The system draws on documents that already exist and answers. End of task. All the value lies in the quality of the sources and the clarity of the answer.
If the answer is yes, you are in different territory. A customer writes to move a booking, and something has to change in the calendar. A supplier invoice arrives, and something has to end up in a spreadsheet. A foreign buyer asks for a technical sheet, and someone has to generate it in the right language. Here text is the means, the result is elsewhere.
The four capabilities that make an agent
An agentic system is recognised by four things it can do. If one is missing, what you have in front of you is a chatbot with buttons.
| Capacity | What it means | How it is checked |
|---|---|---|
| 01 · Retrievefrom authorised sources | It searches inside a closed list of documents, folders and systems it has been given, and says where each answer comes from. |
Ask a difficult question and see whether it cites the file. If it doesn't cite, it didn't search: it remembered. |
| 02 · Decidewithin written limits | It chooses between options someone listed before it was built, not between every imaginable option. |
Ask for the list of permitted actions. It must exist on paper, before the code. |
| 03 · Actand leave a record | It carries out the action inside a real system and writes what it did, when, on which record, with what outcome. |
Ask to see the execution log. If the log doesn't exist, the execution can't be verified. |
| 04 · Stopwhen evidence is missing | When the source isn't there, the case falls outside the scope or a criterion fails, it abstains and alerts a person. |
Ask it a question it can't answer. If it answers anyway, that system will lie to one of your customers too. |
The fourth is what distinguishes a serious supplier from a salesperson. A system that stops looks less powerful in a demo and costs much less in production.
Where a chatbot goes wrong
A chatbot almost always fails for one reason: it was given sources nobody kept up to date. It answers confidently using last year's price list, or a procedure the company changed in March. The model isn't wrong, the maintenance of the sources is.
The second way is the undeclared scope. An assistant that should answer on products and opening hours ends up giving advice on a dispute, because nobody told it where to stop.
Where an agent goes wrong
An agent fails in more expensive ways, and all of them predictable.
It does something it shouldn't have. The permitted actions weren't a closed list, so it found one more. It is prevented by writing the list beforehand, not by correcting afterwards.
It does the right thing at the wrong moment. It sends the payment reminder to a customer who paid yesterday, because it was reading old data. It is prevented by deciding which source is right when two contradict each other.
It carries on when it should have stopped. The case was outside the scope and nobody had written an abstention condition. It is the mistake that reaches the customer.
The three things that make an agent safe are written before the first line of code, or they are never written.
When you need neither
It is worth saying because nobody selling AI says it. If a step has stable input, stable transformation and stable output, the right solution is a deterministic script: it costs less, breaks less, and when it breaks you understand straight away why.
A report that every Monday takes the same fields and puts them in the same spreadsheet doesn't need to interpret anything. Putting a language model in the middle adds a per-run cost, uncertainty and one more point of failure, in exchange for nothing. Autonomy comes at a price, so it goes only where the process really has to interpret a changing context.
The four questions to ask whoever proposes it to you
They work with any supplier, and don't require knowing how to program.
Which sources does it read, and who authorised them? A good answer is a list. A vague answer means it will also read things it shouldn't.
Which actions can it take, one by one? If the list doesn't exist in writing, the limit doesn't exist either.
What happens when it can't find the answer, or a service doesn't respond? The right answer describes a precise behaviour: it stops, alerts, and whoever receives the alert has a name.
Where is what it did recorded, and who can read it? Without a log there is no control, and without control there is no compliance. The Italian AI law asks for exactly this: traceability and a person able to correct.
In one line
A chatbot saves you the answers. An agent saves you the steps. The first is judged by how up to date its sources are, the second by how tight its limits are. And almost every company that asks for the second discovers, at the first question, that the problem is solved by the first, or by neither.
How we design a system of this kind, and what we measure before building it, is on the method page. The pieces already built, one by one, are in the services.
Which of the two you need depends on the work, not the technology.
The difference between the two matters when you look at a precise task: how many enquiries arrive each week, how much time the replies written by hand cost, what must stay in a person's hands. The Diagnostic measures five dimensions in five minutes. No spam: the report is yours, and we write to you only if you ask us to.