Most teams start with the wrong question. They ask what AI can do, when the useful question is which work is repetitive, rule-bound and currently done by someone who would rather be doing something else.
Work that automates well
Anything with a clear input, a known set of steps and a checkable output. Intake, routing, reconciliation, status updates and first drafts all fit. If you can write the steps on an index card, an agent can usually run them.
Work that doesn’t
Judgment calls with real stakes, relationships, and anything where the rules change every week. We keep people in charge of those and let the agent prepare the ground.
A quick test
Ask whoever does the task today to explain it in five minutes. If they can, it is a candidate. If they say it depends, it probably needs a human, with an agent doing the prep.
