The difference is whether anything decides
Automation follows a path somebody drew. An agent chooses a path. Everything else, the pricing and the risk and how hard it is to test, follows from that one difference.
If every branch is knowable in advance, you want automation. It is cheaper, it is predictable, it can be tested exhaustively, and when it breaks it breaks in a way somebody can read. Adding judgement to a process that does not need judgement buys you cost and uncertainty.
If the input varies in ways nobody can enumerate, whether free-text enquiries, documents that never arrive in the same shape twice, cases where the right answer depends on context a rule cannot capture, then a rule-based system will fail on the cases that matter and an agent is worth the extra.
Told apart in one question
Draw the process as a flowchart. If you can complete it without ever writing a box that says and then it works out what to do, it is automation. If you cannot, it is not.
That test is worth applying to anything you are quoted for, because a great deal that is described as AI is a scheduled job with a language model doing one step in the middle. That is frequently the correct build. It should not be priced as though it were the other thing.
What each one costs you, beyond the build
- Automation costs mostly at the connection points. Once it runs, it runs, and it fails loudly enough to notice.
- An agent costs at the connection points too, and then keeps costing: it needs review, it needs a defined behaviour for when it is unsure, and somebody has to look at what it did. That is not overhead to be trimmed. It is the thing that makes it safe to run.
- The expensive mistake is buying the second and operating it like the first. It works for a while, which is exactly what makes it expensive.
Most real systems are both
In practice the useful build is usually mostly automation with judgement at the two or three points that genuinely need it. The predictable parts stay predictable and cheap; the model is spent where variation actually lives.
That shape is also easier to explain to the people who have to use it, and easier to unpick when something goes wrong, which are not small considerations, because they are what decides whether it is still running in a year.