The most dangerous thing an AI can do is act without permission.
AI systems that act autonomously create unpredictable risks. An email sent to the wrong person, a payment processed incorrectly, a database record updated with bad data — these aren't theoretical risks, they're daily realities of autonomous systems.
Stewart separates thinking from doing. Analysis, drafting, planning, and organizing happen freely — these don't affect anything outside the system. External actions — sending, publishing, purchasing, updating — require your explicit approval every time.
Over time, as you see consistent, reliable behavior, you may choose to pre-approve certain routine actions. But the default is always: ask first, act second. Trust is earned through demonstrated reliability, not assumed from the start.