Goldratt’s Theory of Constraints says a system’s performance is determined by its constraint. Most attention in AI-assisted development still goes to model speed, even when the model is already the fastest part of the process.
What happened between 17:00 and 09:00
One task started at 17:00 and reached production at 09:00 the next morning. The human wrote and clarified the task, AI implemented it, and the human reviewed the code. Control moved from AI back to the human four times. Active work still added up to only six hours.
Observed path
- AI work
- 3h
- Human work
- 3h
- Waiting
- 10h
16h lead time / 37.5% flow efficiency
With waiting reduced
- AI work
- 3h
- Human work
- 3h
- Waiting
- 2h
8h lead time / 75% flow efficiency
Eight of the ten waiting hours came from the human being unavailable overnight. Two more came from ordinary handoff delays: stepping away, noticing the AI had finished late, or working on something else.
A team working across shifts would not need to wait eight hours for one person to return. With the same three hours of AI work and three hours of human work, the task could reach production in roughly eight hours.
This task alone says nothing about team throughput. Parallel work, capacity, rework, and handoff cost still matter. It shows a narrower result: lead time can be cut in half without making the model faster.
Teams could work across time zones, use explicit handoff artifacts, share review ownership, or route the next action to whoever is available. The right change depends on the measured constraint.