This is a useful framing. Treating AI as an async collaborator instead of an instant oracle feels more realistic, especially as reasoning modes get slower. Good workflows should assume latency and help humans stay productive while the model is thinking, not just show a spinner.