There's a number that shows up repeatedly in the research on attention and focus.
It's 23 minutes and 15 seconds.
That's how long it takes, on average, to fully recover your focus after a single interruption — according to Gloria Mark's lab at UC Irvine. Not to get back to where you were. To fully recover.
One interruption. Twenty-three minutes.
Gloria Mark, UC Irvine — observed knowledge workers in real workplace settings, not lab conditions.
4 interruptions per morning = ~90 minutes of lost focus capacity, even while sitting at your desk the whole time.
Most engineers don't know this number. The ones who do often file it under "interesting" and move on. But do the math:
If you get interrupted 4 times in a morning — each time by a Slack ping, a code review request, an AI suggestion that snaps you out of deep work — you've lost roughly 90 minutes of recovered focus time. Not 90 minutes of elapsed time. Ninety minutes of focus capacity, even though you were sitting at your desk the whole time.
The interruption itself takes seconds. The recovery takes twenty-three minutes.
This is the transaction cost nobody talks about.
Gloria Mark has been studying attention in workplace settings since the early 2000s. Her team has observed thousands of hours of workers in real offices — not labs, real offices — tracking every switch in attention, every interruption, every return to a task.
Her core finding: the average knowledge worker is interrupted or switches tasks every 3 minutes and 5 seconds. And after each one, it takes 23 minutes to return to the same level of focus.
The math is brutal.
But there's a second finding that matters more for this conversation: the quality of the recovery depends on what you do during those 23 minutes.
If you spend those 23 minutes checking email, reading Slack, or scrolling — you're not recovering your focus. You're replacing one shallow task with another. You come back to the original task less focused than when you left it.
The recovery isn't automatic. It's a choice about what fills those 23 minutes.
Here's what's specific about AI's effect on this dynamic.
Traditional interruptions — a colleague stopping by, a meeting, a phone call — have a natural boundary. When they're over, you come back. There's a beginning and an end.
AI doesn't have that boundary. AI runs alongside your work. The suggestion bubble stays open. The next prompt is always one click away.
This creates a new kind of interruption that doesn't announce itself: the ambient availability loop. Your AI tool is always there, ready to answer, ready to generate, ready to finish your thought. And each time you engage with it, you're switching out of your own problem-solving and into a receiving mode.
The 23-minute clock starts. But instead of recovering in silence — instead of sitting with the not-quite-knowing — you fill those 23 minutes by asking the AI.
You don't recover focus. You borrow it.
And the cost compounds because you often come back to your own work having made less progress on it than if you'd stayed in the loop the whole time.
So here's what this means in practice.
After you use AI to solve something — don't immediately return to the main task.
Wait.
Not because you've done something wrong. Because you've interrupted yourself — even if the interruption was voluntary — and you need the full 23 minutes to recover.
After any significant AI-assisted task: set a 5-minute timer before you return to deep work. During those 5 minutes, don't use screens. Walk. Look out the window. Make coffee. Let the AI output settle, and let your own focus recover.
It's not a break. It's a recovery buffer. And the research suggests that what fills those minutes matters as much as the minutes themselves.
Beyond the 23-minute buffer, Mark's research and related work point to a few things that reliably improve focus recovery:
The 23-minute rule isn't about being more productive. It's about what's actually happening when you work.
You can't manufacture deep focus. You can only create the conditions for it — and one of those conditions is time between interruptions. Not time in the day. Time between switches.
AI makes switching effortless. That ease is the feature. It's also the cost.
The engineers who use AI without eroding their focus are the ones who are deliberate about when they use it, and intentional about what they do in the recovery window after.
Twenty-three minutes isn't a lot. It's also everything.