Chatbot, assistant, agent, automation: four words sold to you as the same thing (wrongly)

An owner-manager tells me over the phone that he's been offered "an AI agent" to send his follow-ups. I ask him two questions, and it turns out he doesn't need an agent at all, just a plain automation, ten times cheaper and risk-free. The salesperson had used the word that impresses. I see this kind of mix-up every week, and it costs money. So let's lay the four words out flat, once and for all.

The assistant: you ask, it answers

This is the best known, because it's ChatGPT or Claude. You open a window, you make a request, it answers, you carry on. You're in the loop at every step, at the controls. Nothing happens if you type nothing. It's a colleague you call on whenever you want, to write, summarise, find ideas. Useful, immediate, and with no autonomy at all: it waits to be spoken to.

The chatbot: an assistant stuck in a window

The chatbot is often that little bubble in the bottom right corner of a website that answers visitors' questions. Technically, it's an assistant, but placed in one specific spot, with a narrower scope and a single purpose: inform, direct, capture a request. Good when it knows its limits and hands over quickly, tiresome when it goes round in circles pretending to understand. The word sounds dated, but the thing itself is still very common.

Automation: a rail, not a conversation

Here, we change register entirely. An automation doesn't talk to anyone. It's a fixed sequence of actions, triggered by an event: "when this happens, do that, always in the same order." One example I installed for an online shop: every delivered order triggers, a few days later, a review request sent on its own. More than 850 customers tracked one by one, one send a day, without anyone touching anything. It doesn't think, it doesn't choose anything. It follows a rail. And that's exactly right: for a task that always repeats the same way, you really don't want a machine to "decide".

Automation follows a rail. An agent chooses its own path. That's why an agent needs watching.

The agent: give it a goal, it chooses the steps

The agent is the level above, and the most misused commercially. You don't give it a sequence of actions, you give it a goal, and it chooses its own path to get there, chaining several steps together. My own system is one: every night, it reads the day's figures, cross-checks them, and prepares me a shortlist of around thirty decisions, ranked from most profitable to most marginal. Nobody dictated the steps to it one by one. It decides its own route. And that's exactly why an agent demands a level of oversight an automation doesn't: when a machine chooses, it can choose wrong, and you need a safety net.

Why the confusion costs money

The problem isn't vocabulary for its own sake. It's that people are often sold the most expensive word for the simplest task. You're offered an "agent" where a fixed rail would have done: you pay for complexity and inherit a risk you didn't need. Or the other way round, more rarely: you're sold a rigid automation for a need that genuinely called for judgement, and it breaks on the first unexpected case. Knowing which box your need falls into is already half the decision, and it saves you from paying for a word.

One last word to be fair: in real life, these boundaries blur a little, and even I use "assistant" loosely. What matters isn't reciting the definitions, it's spotting, behind the salesy word, what the thing actually does and what it demands of you. An agent isn't "better" than an automation; for a rail, it's even a poor choice. If you want to place these building blocks against each other, I set them out in the three levels of AI in a business. And when it comes to laying the right rail on your tools, without overselling, that's the heart of my automation work.