A Stripe engineering manager, Amol Sharma, posted a number recently: proposal review and cross-team alignment went from about 40 per cent of his week to 70 to 80 per cent, over three years. Not because he took on more direct reports. Because execution got cheap and initiative got abundant, and nobody adjusted the review load to match.
That is not a story about tools. It is a story about where the bottleneck moves once one part of the system speeds up and the rest does not.
Anyone can generate twenty pages now. Someone still has to find the two that matter. If the norm around documents does not change when the cost of producing them collapses, the org does not save any work. It just relocates the work onto whoever has to read it. The fix Sharma found was not reviewing faster. It was asking authors to own the compression themselves: a one-pager, the key points, the open questions, before anything reaches a reviewer's desk.
The second bottleneck was Sharma himself. Routing every decision through one person is fine when execution is slow enough that the queue never builds. Once the team around him accelerated and he did not, the queue built, and it took honest teammates naming it before he saw it. The fix was pushing small and medium decisions down to tech leads rather than holding onto all of them.
That move has a name older than any of this. L. David Marquet ran the USS Santa Fe by deliberately moving decision authority to wherever the information already was, rather than making every decision travel up to him first. Turn the Ship Around!, the book he wrote about it, is not about software. It describes the same structural problem: a single point of decision-making becomes the constraint the moment the people around it can act faster than it can decide.
None of this happens automatically once people start delegating more. Amy Edmondson's research on psychological safety, admitting an error without it costing you, is well established. The AI-specific evidence is newer and thinner: one recent study found it predicted whether employees started using AI at all, but not how much or how long they kept using it. Getting people to start is one problem. Getting the new way of working to hold is separate, and it is the one Sharma's account is actually about.
If you manage anyone, or anyone reports through you before something happens, both mechanisms are worth checking against your own week. Has your review load quietly doubled without your role changing? Are decisions queuing at your desk that a fortnight ago you would have made in the room? Neither is a sign you are doing something wrong. It is a sign the constraint moved, and moved to you, and the fix is not working longer hours. It is deciding, deliberately, what you no longer need to be the one deciding.