The short answer
Automation runs a fixed route you defined and never does anything else. An AI agent decides the order and the steps itself to reach a goal. Fixed steps and a fixed outcome call for automation. Varying input and a judgement per case call for an agent, and then always with an employee who approves.
The difference in one table
| Automation | AI agent | |
|---|---|---|
| Decides the steps | You, in advance | The model, per case |
| Same input, same outcome | Always | Usually, not guaranteed |
| Breaks on | Input that does not fit the rules | A goal described too vaguely |
| Cost per case | Predictable | Depends on how many steps the model takes |
| Auditable afterwards | Fully, the route is fixed | Only if you record every step |
| Suits | Retyping, routing, checking against fixed rules | Researching, summarising, weighing, drafting |
Why this matters
An agent that decides the order itself can also decide the wrong order. As long as it only prepares, that is annoying at worst. Let it send, book or commit, and it becomes a risk you do not see coming, because it fails not with an error message but with a plausible wrong answer.
So every agent needs a boundary: what may it read, what may it do, who approves, when does it stop and what does it record. Those five questions sit on the page about control and security.
What you build in practice
Usually both, inside the same workflow. The automation fetches the mail, prepares the attachment and writes the result away. The model does the step that asks for a judgement: what does this say, does it match the source, which reply fits here. The employee decides.
Whoever starts with an agent for everything builds something nobody can check. Whoever starts with automation for everything gets stuck on the first case that deviates. Drawing the line is the real work.
Which of the two do you need
Answer one question: can someone who knows the work write down in advance which step follows when, without exceptions? If yes, it is automation and you need no model. If no because it differs per case, that is where the agent sits. The free process check walks that question through with you.
Frequently asked questions
Is an AI agent the same as a chatbot?
No. A chatbot answers a question. An agent receives a goal and takes several steps by itself to reach it, for example looking up information, comparing it and drafting a result.
Can an agent enter my system on its own?
Technically yes, sensibly no. In the way of working I build, an agent may prepare and may never send, book or commit on its own. An employee approves and decides.
Which one is more expensive?
Automation has predictable costs, an agent does not: it charges per step the model takes, and that varies per case. At large volumes that difference is worth calculating.
How do I know whether my process needs an agent?
If someone who knows the work can write the steps down in advance without exceptions, you do not need an agent. If it differs per case and a judgement is required, you do.
