Agent, chatbot, automation: three things, one name

The owner of several pharmacies in Constantine receives three proposals. The first talks about a chatbot, the second about an agent, the third about an automation. All three use the phrase "artificial intelligence", all three give a convincing demonstration, and none of them does the same work.

The confusion is not a vocabulary fault: it has consequences. These three systems fail differently, are monitored differently, and cost in different proportions. A buyer who cannot tell them apart cannot read a quote.

This article gives the distinction that decides — who starts the work and who decides what happens next — then what each costs to keep. It compares no models and names no brands: what is at stake is the nature of the system, not its supplier.

An automation talks to nobody. It waits for something to happen — a form filled in, a file dropped, a time of day, a row added to a spreadsheet — and executes a sequence of steps decided in advance.

The important point is that it needs no artificial intelligence to exist, and that most of this work is not artificial intelligence. The model, where there is one, handles one precise step: reading a document, classifying a message, drafting a summary. The rest is plumbing.

It fails silently, which makes it more dangerous than a chatbot. A chatbot that stops answering is visible within the minute; a stopped automation produces nothing, and the absence is only noticed when somebody looks for a result that never arrived.

Its cost is mixed: a fixed part that runs even with no activity, and a variable part with volume. It is also the only one of the three that can do damage at scale, because it repeats the same mistake across a thousand rows without tiring.

Its fragile part is not the model, it is the trigger. A renamed tab, a column inserted in the middle, a field added to the form by somebody who did not know a machine was reading it: each of those gestures is enough to stop the chain, and none of them looks like a technical change to the person making it. That is why, beside each automation, you write down which file and which form serve as its input.

In a real business the three coexist and hand work to each other. A chatbot receives a request for a quote; an automation creates the file and alerts the salesperson; an agent prepares the draft by fetching prices from the catalogue.

That combination is the right architecture for a simple reason: each part does what it can do predictably. It also has a practical consequence — one link’s failure does not look like another’s, and monitoring has to be differentiated.

That is why a project is better described by its handovers than by the name of the system. Where the chatbot ends, where the automation begins, at what moment a human takes over: those three boundaries describe the project better than any diagram.

A supplier who cannot draw those boundaries on a sheet of paper in five minutes has not yet designed the system, whatever the demonstrations show.

The most frequent answer is: the most boring one. The most repetitive, least ambiguous task, the one nobody asks for and everybody does — and it almost always belongs to automation rather than to an agent.

There is a reason that is not technical. All three ask the same thing of you: that the work be described. A business that cannot write down its own process will get nothing from any of them, and automation is what makes that gap visible fastest and cheapest.

The second criterion is volume. Below a few dozen occurrences a week, none of the three pays for itself: the go-live and monitoring work exceeds what it saves, and the honest answer is to buy nothing.

The third criterion is the forgotten one: is there somebody to watch. A system with no watcher degrades, and the bad choice is not the wrong tool, it is a tool nobody looks at.

Those three criteria are asked in that order and no other, and the order does the work. Starting from volume leads to picking the most frequent task, which is often also the most ambiguous. Starting from the watcher leads to picking the most available person’s task, which is not necessarily the most expensive one. The description first, then the volume, then the watcher: it is the only sequence that does not turn against you at the third question.

We make the section 11 list with you, place each line on the two axes, and say which of the three families applies — including when the answer is "none, your volume does not justify it". We then draw the handover boundaries before a line of code is written.

We refuse to deliver an agent that executes an irreversible action without human confirmation. It is not a start-up stage to be lifted later: it is the condition that makes the rest defensible.

We also refuse to sell the word "autonomous" without completing it. A quote promising autonomy without naming the last decision left to a human describes a property with no definition.

What you can do without us is the sheet itself. Many businesses discover there that their need is a three-step automation, and the exception queue it will produce will be more useful reading than any brochure.