From token costs to capability gaps, here’s what AI gets right — and where you’ll still need real people.