For most people, once AI saves them some time, those twenty minutes quickly go into replying to email, a meeting, or the next thing they are being chased about, and the next day everything is as before.
When that happens, it is not necessarily that the AI is being used badly; it is that the flywheel never started turning. What really separates outcomes is not only how much time is saved, but where the saved time ends up.
1. Compounding: the only arrow that really matters#
Save X minutes a day, put those X minutes back into optimising your own workflow, and two or three months later X has compounded into a surprising number.
Written down this looks like a platitude, but it is the only hard step in the whole diagram, because putting time back into optimisation is something nobody sees at the time: no ticket gets closed, no manager knows, and all of the return is in the future.
So the real bottleneck of the flywheel is not technical; it is whether you can tolerate a stretch of time with no visible output.
2. Delegation: what you spend is no longer labour, but waiting#
Some tasks, even if the AI needs to run for a long time, four or eight hours, can from then on be offloaded entirely, as long as it can run alone without human intervention.
The task still takes time, but what I spend is no longer labour, but waiting. Labour has a limit, tires, and occupies my attention; waiting does not.
Once a task can be reshaped into something that can be left to run, its cost shifts from my stamina and attention to machine time.
3. A smaller window of flow#
As for work that still needs a human, say a task where I have to spend half an hour reading code or a paper and half an hour producing the output, with the whole hour in uninterrupted flow, which means I need a full, unbroken one-hour window before it can be scheduled at all.
It is not that office workers have no spare hour, but that they lack an hour guaranteed to stay continuous.
Even if AI cannot take over the whole task, it can compress the part that needs flow substantially, so the same task fits into a smaller window. For anyone who gets interrupted by calls and messages, this is worth far more than saving time: it turns tasks that could not be scheduled into tasks that can.
4. Don’t draw a line#
Many people use AI by first drawing a line in their heads: this side it does, that side “it can’t”, and then spending their time resenting the part it can’t do.
My approach is the opposite: first do everything the AI can do right now, and optimise it.
Precisely because you can run several sessions at once delegating different things, everything it can do should be taken out, done, and optimised. By the time you have spent N weeks optimising the current process, the next version has improved again and you will find more it can do, and from then on you keep optimising along with each new version.
That line moves every few months, so whoever drew it first is effectively judging today’s model by an impression from a few months ago.
5. It evolves with you#
Every time you use an AI agent it keeps recording what happened and what you said, and you in turn come to understand more and more where this agent’s limits and personality currently lie.
AI is the kind of prosthesis that fits better the more you use it.
This is also why the previous point says to do everything the AI can do first: the breaking-in is itself an investment. The extra twenty minutes you spend today are not wasted on a one-off task; they go into an interface that accumulates.
When the wheel doesn’t turn#
Back to that ★ in the diagram. If the time you save is immediately filled with new work, all five of these fail, not because the AI is not strong enough but because the flywheel is missing the segment that sends the momentum back to the start; you end up just speeding things up for someone else while staying where you are.
That is not a technical problem but an environmental one, and environmental problems are usually not solved by getting better at prompting.
Further reading: if you have not started yet, what is stuck is usually not the idea but the first two steps. Seven Opening Moves is the concrete how-to →
If your process is stuck at knowing it can be optimised but never finding that stretch of unproductive time, that stretch can be outsourced.
This kind of build can be commissioned from me: scope, pricing, how to start →