Chatbot, automation, agent
Almost every confusing conversation about AI in business is two people using one of these words to mean another. They differ in one specific way: who decides the order of the steps.
| Who decides the steps | Good at | Fails by | |
|---|---|---|---|
| Chatbot | Nobody — it responds and stops | Answering questions from material it has | Answering confidently and wrongly |
| Automation | You, at build time | Doing the same reliable thing every time | Meeting a case you did not anticipate |
| Agent | The model, at run time | Tasks where the steps depend on what it finds | Doing something plausible and wrong, repeatedly and quickly |
Source: Cybrial’s own framing. It is the distinction that decides which one a business should buy, and it is deliberately about control rather than technology.
Note the last column. An agent’s failure mode is the expensive one, because it acts. A chatbot that is wrong has said something wrong; an agent that is wrong has done something wrong, and possibly several things, before anyone looked.
Is ChatGPT an AI agent?
By default, no — in its plain form it is a conversational model: you ask, it answers, nothing happens in the world. With tools attached, it becomes one, because it can then decide to search, run code, read a file or call an API and use what comes back to decide the next step.
That is the honest shape of the distinction, and it is why the question keeps recurring. “Agent” is not a different kind of technology from the chat interface most people have used. It is the same model, given tools and a goal instead of a conversation.
The “five types of agent”, and why that list is not much use
The classification that turns up everywhere — simple reflex, model-based reflex, goal-based, utility-based, learning — comes from academic artificial intelligence and long predates language models. It is a genuine taxonomy and it is close to useless for deciding what to build.
The distinction that matters commercially is much blunter: how much can it do without a human agreeing first?
- It drafts, a person approves. Cheap to build, cheap to be wrong, and where the great majority of business value currently is.
- It acts within a fenced area — reading records, writing to one system, with limits on what it may change.
- It acts across systems unsupervised. Expensive to build, expensive to be wrong, and rarely the right first project.
Almost every business asking about agents wants the first one and has been quoted for the third.
What drives the cost of building one
No build price is published here, and none is quoted for other people either, because there is no primary source for what an agent costs to build and an invented range would be worse than silence. What can be said honestly is where the hours go, which is the useful half anyway.
- How many systems it must reach, and whether each has a documented API
- How wrong it is allowed to be — a draft somebody approves is a fraction of the cost of an action taken alone
- What it does when it does not know, which is a design problem rather than a coding one
- Whether real cases exist to test it against, or whether they have to be invented first
- How much it will be trusted with over time, because that decides how much logging and review it needs on day one
The running cost is usually the pleasant surprise: model usage for a small business task is normally tens of pounds a month. The expense is the build, and most of the build is integration rather than anything to do with AI.
When the simpler thing wins
If you can write the steps down, write the steps down. An automation that follows them will be cheaper, faster, more predictable and far easier to debug at four in the afternoon when something has gone wrong. Handing a fixed sequence to a model that might choose differently each time is paying extra for uncertainty.
The genuine case for an agent is a task where the steps depend on what it finds — reading an inbound enquiry and deciding which of six things it is, then doing the right one. That is not scriptable, and that is exactly where the technology earns its keep.
A useful rule of thumb: if a competent new starter could be given the task as a written procedure, build the procedure. If they would need judgement, an agent is worth considering — and so is asking whether the task should be automated at all.
