Three things together: high repetition, rules clear enough to write down, and a time cost somebody can measure. Miss any one of them and the build is not worth starting.

The three-part test

Repetition tells you the payback compounds. Clear rules tell you the agent can be trusted without a human checking every output. A measurable time cost tells you whether the saving is real or just a feeling that the task is annoying.

What fails the test

Work that looks repetitive but hides a judgement call in the middle. Processes that change every quarter, where the agent would be rebuilt as fast as it ships. And anything where nobody can say how many hours it currently takes, because that is the number the whole business case rests on.