AI agent, automation and chatbot — three different things, one name

Three tools sold today under the same "AI for business" heading do three different things. One pushes data along a path somebody drew in advance. One answers questions. One gets a goal and works out how to reach it.

In the companies we talk to, the difference surfaces after the fact — when the tool that was bought turns out to do something other than what the process needed. What that costs is usually a lost quarter rather than money.

Below: three working definitions and the criterion that settles the choice.

Automation: the path written down in advance#

Automation is a written rule: when X happens, do Y, then Z. Everything that happens was anticipated and described by a person beforehand.

The same input gives the same output. On the first run and on the thousandth. That is exactly why automation runs invoicing, reminders and the movement of data between systems — the places where "nearly right" is not right.

Failure is loud. A rule either runs or stops with an error at a named step. There is no state in which it works, but differently from yesterday.

The price is predictable too. A rule costs the same at a hundred runs and at ten thousand, because there is no model inside it to ask each time.

One condition: somebody has to know the path in advance.

Chatbot: an answer that changes nothing#

A chatbot takes a question and returns text. Whether it answers from general knowledge or from your own documents, its output ends on the screen.

That is an advantage where the job is knowledge: new starters asking about procedures, finding a clause in a contract, checking where something stands, drafting a letter. The risk is bounded by definition — a mistake stops at the screen and leaves no trace in any system.

It is a limitation wherever the answer is meant to produce a change. A chatbot that tells you flawlessly how many jobs are waiting to be quoted quotes none of them. Somebody still types the result into the system: the work has not gone away, it has moved one step.

Agent: a goal instead of a path#

An agent gets a goal and a set of tools it may use. It picks the steps itself.

"Itself" is the whole difference. Nobody wrote the order down in advance, so two runs of the same task can look different and both be correct.

Where the path cannot be described, it is the only option. The typical case: input arrives in a dozen formats at once, exceptions are the rule, and deciding requires context that will never fit in a form field.

The same property is the entire cost. If the agent chooses its own steps, you do not know how many there will be, what they will cost, or whether tomorrow's run will resemble today's.

How to work out which one you need#

Three questions, in this order.

Can you write the steps on a sheet of paper? If you can, and they are the same every time, this is automation and nothing more is needed. An agent here means paying for a decision already made.

Does something have to change in a system? If an answer is enough, a chatbot will do. If a record has to be created, a document sent or a status changed, a chatbot is too small, however well it answers.

How many exceptions are there? Five branches, and a rule handles it. Forty, with a new one every month, and maintaining the rules costs more than an agent reading the situation afresh on every run.

Why an agent where a rule would do is an expensive mistake#

Four costs, easiest to spot first.

The bill grows with the number of runs. Every step the agent takes is its own query to the model, so one run's cost multiplies by the number of steps, then again by the runs in a month. A rule doing the same job never queries the model, so its bill does not grow.

You lose repeatability where it was the point. A process meant to look the same every time starts looking similar. In accounting, contracts and anything somebody will eventually audit, "similar" is worse than slower.

Diagnosis gets more expensive. A broken rule shows the step it stopped at. With an agent you read what it concluded and why before you can locate where the error began.

The team checks anyway. The cost nobody plans for. When people cannot tell whether the process took the same route as yesterday, they look at the output — and instead of a saving you have a new task.

The mistake runs the other way too. Rules piled onto a process with too many exceptions become, within a year, a tangle nobody can change without risk. The answer there is an agent — at one specific step, not along the whole process.

What works in practice: the mixed setup#

Solutions that work are rarely one of the three things.

The typical arrangement: a rule catches the event and prepares the data, an agent handles the single step nobody could describe in advance, and the team asks a chat where things stand.

The practical principle: an agent only on the step you cannot write down. The rest as rules. The bill stays small, diagnosis stays simple, and repeatability survives where it matters.

The split pays off organisationally too: a rollout where the agent owns one clearly marked part of the process is easier to defend inside the company than "AI that will take care of the process".

Summary#

Automation runs a path somebody wrote down. A chatbot answers and stops there. An agent picks the steps itself, because there was no path.

The choice is not about which is more modern, but about whether you can describe the route to the goal. If you can, an agent is a more expensive version of something you already have. If you cannot, no number of rules will substitute for one.

The wider context — what agents do in a company, where they fail and how to stay in control — is in the guide on AI agents for business.