Thomas Davenport and Steven Miller studied 29 real AI deployments and found the same pattern again and again: augmentation, not automation, wins.