The queue time example is the part that changed how I think about this. Fifty minutes of active work stretched across almost two days because research waited eight hours for review and approval waited another day, and that is exactly the kind of waste an agent can remove without touching the actual decisions. I also liked the line that capability is not permission. An agent that can send email, update a CRM, and spend money should not be allowed to do all three by default. Measuring cycle time and rework rate instead of token usage feels like the right scoreboard too. The most useful agent being boring is a good reminder that the win is a workflow that quietly finishes, not a demo that impresses.