The test
Ask one question about the task: if the input arrived slightly wrong, would a sensible person still know what to do?
If the answer is no, because the task is a rule and a wrong input should simply stop it, you want an automation. An invoice passes its due date, a reminder goes out. A form is filled, a call is booked. The steps never vary and nobody should be improvising inside them.
If the answer is yes, because a person would read the wrong-looking input, work out what happened and carry on, you want an agent. A call transcript arrives with three decisions buried in forty minutes of talk. A folder of files lands with no names that mean anything. A contract has a clause nobody has seen before. The steps are not fixed; the judgement between them is the work.
What an automation is
A chain of tools with a trigger at one end and an output at the other. It runs on the tools you already own: a form tool, a CRM, an accounting system, a calendar, a chat app. Every step is a rule someone wrote down, and the chain is only as good as the week someone spent inside the work before writing the rules.
Where it goes wrong: at the edges. The customer with two records. The invoice with no due date. The form filled in by a bot. A well-built automation stops at the edge and tells a human; a badly built one carries on and does something confidently wrong.
What an agent is
Software with four parts you can open and read: its instructions, the tools it is allowed to touch, what it remembers, and the rules that stop it. It starts when something happens, reads what it needs, decides, acts, and leaves a receipt of what it read and what it did.
Where it goes wrong: at the boundaries nobody set. An agent with no guardrails will do the task it was given in a way nobody wanted, because it can. The guardrails are the product. Ours run as code, not as a sentence in a prompt, and a human still says yes to anything that leaves the building.
Why you usually want both
The work in most businesses is a line with a few places where a person has to think. The line is the automation. The places are seats for an agent. A call ends, which is a trigger; the transcript is read and the decisions pulled out, which is judgement; the CRM is updated and the follow-up drafted, which is a rule again.
That is the shape of the post-call brief, and it is why we sell it as an agent pack rather than an automation: the loop in the middle is the part that earns its keep.
What to do first
Pick the task that someone re-does every week and ask the test question. If it is a line, buy or build the automation and let it run for a month. If there is judgement in it, the agent is worth the extra care, and the first one to build is the one where the judgement is smallest and the receipt is easiest to check.