The Twenty-Three Minute Myth

The published paper is worth reading precisely because of how it differs from its own legend. In the 2005 study, Mark's team shadowed information workers minute by minute and found that when interrupted work was resumed the same day, which 77 per cent of it was, the average time before returning to it was 25 minutes and 26 seconds, with a standard deviation of nearly 55 minutes. During that gap, people were not staring at a wall recovering. They completed, on average, 2.26 other tasks. In the Gallup interview a year later, a slightly different cut of the same data became 23 minutes and 15 seconds, and two decades of retelling did the rest: returned to the task became recovered focus, the intervening tasks vanished, and a variance twice the size of the mean disappeared into one confident number. I would call this stat laundering: a figure passed between secondary sources until the primary is unfindable and the claim has quietly changed shape. The productivity industry runs on such numbers, and more of its famous figures dissolve the same way when you ask for the paper.

What makes the myth genuinely damaging is that it points the diagnosis in the wrong direction. Mark's own follow-up experiment, published in 2008, found something the recovery-clock story cannot accommodate: interrupted work gets finished faster. People completing an email task under interruption took just over 20 minutes against nearly 23 minutes uninterrupted, with no measurable drop in quality. People compensate, writing less and working faster. The bill arrived somewhere else. The interrupted groups reported significantly higher stress, frustration, time pressure, and effort. The cost of interruption, in other words, is not a 23-minute hole in the calendar. It is paid in the currency I have argued is the founder's true constraint, energy (the argument in Return on Energy), and in the tolerance for error that goes with it. Output holds. The human absorbs the difference. A fragmented organisation does not look slow. It looks busy, tired, and increasingly unable to catch its own mistakes.

Meanwhile the real numbers have moved somewhere the myth never imagined. Mark's screen-logging studies tracked the average time a person spends on one screen before switching: about two and a half minutes in 2004, 75 seconds by around 2012, and roughly 47 seconds in her most recent work, a figure others have replicated within a few seconds. The careful framing matters: this is time on one screen, not attention span. And the environment those screens sit inside has become an interruption machine. Microsoft's own telemetry, published in June 2025 and worth the vendor-research caveat, found the average employee interrupted every two minutes during core working hours, 275 times a day, with 117 emails daily and 153 Teams messages each weekday. Set that against the myth and it collapses in the strangest way: it is too optimistic. A 23-minute recovery clock implies you have 23 uninterrupted minutes in which to recover. At an interruption every two minutes, there is no recovery window at all. There is only the compensation mode, all day.

Follow the corrected diagnosis and the standard fix falls apart. If interruption cost a recovery clock, the answer would be to defend blocks of time, which is what the entire focus industry sells: deep-work calendars, no-meeting days, notification hygiene. But blocking time only reschedules interrupt-shaped work. The 275 arrivals still arrive, compressed into the unblocked hours, and the energy bill is unchanged. If the real cost is energy and error tolerance, the only fix that touches it is removing whole classes of interruption from the human entirely, so they never reach a person at all. Look honestly at what those arrivals contain and most are machine-shaped: a status someone could not see, a document someone could not find, a first-line question with a documented answer. That is agent work, and absorbing it is precisely what agents are for. The hard part is not the absorbing. It is knowing which classes can safely leave the human, because some interruptions are the job: the escalation, the judgment call, the client who needs a person. Draw the line wrong in one direction and you have automated away your own responsiveness. Draw it wrong in the other and your people still answer 275 pings a day.