This frames it right — property management was a workflow problem before AI ever showed up, so bolting an agent onto a broken process just fails faster. The tests that stuck with me are the practical ones: can it draft a real reply to a real maintenance message, and does it know when to stop and escalate instead of guessing? The landlord example lands too — five units across three neighborhoods is exactly where the juggling starts to eat the day. And the point about every missed call being a vacancy day is the honest math most automation pitches skip.