How Long Does It Take to Refocus After an Interruption?

Editorial Note: This article brings together research on interruption, attention residue, and task resumption to examine why the same interruption can carry very different costs across different kinds of work. The two-axis framework presented here is an editorial synthesis of that research, not an established scientific taxonomy.

Why the Same Interruption Costs Some People Far More

The same interruption, a dramatically different cost

At three in the afternoon, two very different things are happening in the same building.

On the third floor, an account director has just hung up her seventh call of the day. That one concerned integration scheduling. The previous one concerned payment milestones in a renewal. The one before that involved two departments inside the customer's own organization disagreeing about priority. Three unrelated topics. Three seconds after hanging up, she walked into an in-person product review and picked the thread up cleanly — tone, pacing, and context all intact.

On the fifth floor, a research scientist received a message at roughly the same moment. He had been working through a chain of reasoning: one hypothesis, three constraints, two paths already ruled out. Answering the message took forty seconds. When he returned, he sat at his keyboard for twenty minutes without typing. He needed to confirm that his earlier reasoning still held.

Describing this as "some people are good at multitasking, some aren't" misses the variable that actually matters. The difference lies in how interruptible the underlying work is. The same switch can impose dramatically different recovery costs depending on the kind of task.

Understanding why requires looking at what a switch actually consumes.

What an interruption really costs: rebuilding state, not moving attention

The physical act of shifting attention from task A to task B can be quick. The larger cost may come from reconstructing the implicit state that existed only in someone's head.

While writing code or working through a research problem, a person holds a coupled network in working memory: how the call chain resolves, how this assumption constrains that variable, how far the derivation has progressed, why two particular paths were already eliminated. Most of that information was never written down. It exists as a live reasoning chain. Cut the chain and what dissipates includes not only progress but the validity of the premises underneath it. Recovery can therefore involve two operations: locating where you were, and re-verifying whether the earlier reasoning still stands. In highly state-dependent work, the second operation may be the more expensive one.

Microsoft Research researchers Chris Parnin and Christopher Rugaber quantified this in Programmer, Interrupted. They analyzed 10,000 programming sessions from 86 developers, supplemented by a survey of 414 people. Resuming editing after an interruption took ten to fifteen minutes on average. When the interruption landed mid-method, only 10% of cases resumed within a minute. The same study recorded the improvised measures developers use: navigating back through jump lists to rebuild context, deliberately leaving a compile error in place as a tripwire, and treating the diff view as a last resort. Every one of those tricks serves the same need — get internal state onto an external medium as fast as possible.

Gloria Mark of UC Irvine has studied interruption and attention for years. Her most-cited figure — an average of 23 minutes and 15 seconds to return to an interrupted task — has drawn sustained methodological criticism, including questions about sample and measurement. Her longitudinal research provides a useful measure of how rapidly people switch attention on screens: the average fell from roughly 2.5 minutes in 2004 to 75 seconds in 2012, and later studies from roughly 2016 through 2020 found an average of about 47 seconds. The trend is more informative than treating any single figure as a universal measure of attention.

Sophie Leroy's 2009 work on attention residue supplies the complementary mechanism. Even after a person has moved to a new task, some attention can remain focused on the previous one. Leroy's experiments found that unfinished tasks can create this attention residue and impair performance on the subsequent task. This helps explain why an interruption can be especially disruptive when the previous task has not reached a clear stopping point.

If interruption cost depends partly on how much state must be reconstructed, then people who appear to multitask effortlessly may be working in tasks that require less reconstruction. Their apparent switching ability does not necessarily mean that the same switching cost would apply to more state-dependent work.

Case A: Why eight unrelated calls in a day barely cost anything

Consider a typical day for the account director.

At 10 a.m., an integration team at Customer A disputes a mismatch between interface documentation and actual behavior. At 11:30, procurement at Customer B asks to move payment milestones from quarterly to monthly. At 1 p.m., two departments at Customer C are deadlocked over the priority of the same feature, and she has to mediate. At 3 p.m., a customer with a two-week renewal window suddenly asks for a price reduction. Eight calls, eight unrelated topics.

Her work has three structural properties.

Topics are self-contained. Each call has a clear boundary, and three sentences at the top restore the context, because most of that context lives in the other party's head rather than in hers.

State is externalized. Call outcomes land in the CRM, in email, or in a ticket, and they take the form of a deliverable: a decision, an action item, a confirmation.

Units are divisible. Every call ends at a natural commit point. There is no intermediate state where stopping destroys information.

In database terms, she handles short transactions: request arrives, work happens, commit, release. Once committed, a transaction consumes almost no downstream capacity, so the next request finds her state clean.

The more important point is that switching functions as her output mechanism rather than as overhead. Her role creates value through connection and dispatch — aligning information scattered across other people's heads and pushing decisions to the next stage. In Paul Graham's framing of maker's schedule versus manager's schedule, this is manager's time: scheduled in hourly blocks, where a fragmented meeting damages nothing structural.

So the statement "she switches twenty times a day and stays effective" can be true without implying unusually strong multitasking ability. In this case, the structure of the work makes frequent switching relatively inexpensive.

Case B: Why a forty-second message can wreck an afternoon

Now the scientist on the fifth floor.

At 10 a.m. he is testing a hypothesis: if a particular constraint holds, the current approach can meet the performance target without added complexity. He has eliminated two alternative paths, each for a distinct reason involving unacceptable side effects. Every link in that chain depends on the links before it, so a change anywhere upstream invalidates conclusions downstream.

At 10:20 a message arrives. A colleague asks something entirely unrelated to his current work. He spends forty seconds answering.

What awaits him is not "continue where you left off" but a state audit. Where exactly was I? Do the reasons for eliminating those two paths still apply? What were the boundary conditions on the assumption I was testing? None of these questions have external answers; they may require substantial re-verification or re-derivation. Parnin and Rugaber's 10-to-15-minute figure comes from interrupted programming work, so it should be treated here as evidence of the scale of resumption cost in a specific domain, not as a floor for this scenario.

His schedule carries a second problem. There is a meeting from 10:20 to 11:00, two ad hoc consultations split 11:00 to noon, and a review from 2:00 to 3:00. The first uninterrupted block longer than two hours arrives at 9 p.m. Many people in this position describe it as "I can only do real work at night and on weekends," and attribute that to discipline or diligence. The more accurate description is that no usable continuous block exists during the day. That is a property of the schedule, not of willpower.

A self-check framework: is your context externalized, and is your work unit divisible?

Placing the two cases side by side yields a two-dimensional test.

Axis one: is state externalized? Has the critical information been committed to a document, a codebase, a system, a chart, a CRM record? Externalized state makes recovery roughly a read operation. Un-externalized state makes recovery roughly a recomputation.

Axis two: is the unit divisible? Can work pause at a natural boundary without information loss? Divisible units have commit points. Indivisible units contain intermediate states where stopping is destructive.

The four quadrants suggest different interruption tolerances. Externalized and divisible — support tickets, approvals, routine Q&A — should generally be relatively inexpensive to resume. Externalized but indivisible — assembly following a fixed procedure, operations running against a protocol — may still carry a moderate recovery cost. Implicit but divisible — well-modularized development, writing in discrete sections — can carry a higher cost because some reasoning state remains internal. Implicit and indivisible — complex debugging, proof derivation, architecture design, long-form composition — should be among the most interruption-sensitive forms of work.

A useful question beneath both axes is whether the task is generating new structure or invoking an existing template. Generation produces something new at every step, with no ready-made schema to lean on, so state must be maintained by reasoning. Template invocation compresses judgment into pattern matching, which can make switching relatively inexpensive.

The practical value of this framework is that it explains variation within a single role. The same surgeon shows very different interruption tolerance during routine suturing versus a complex reconstruction. The same lawyer is far more fragile while drafting a clause that propagates through an entire agreement than while running case-law retrieval. The more actionable questions are therefore three specific ones:

Does my core output generate new structure or apply an existing template?
Does my critical state have an external carrier?
Are my interruptions ones I chose, or ones other people pushed onto me?

The third question is frequently overlooked and carries substantial weight. Deciding on your own that "this module is done, I'll go answer a message" saves state at a good breakpoint. Being torn out of a reasoning chain by someone else breaks it at its most fragile point. "I switch twenty times a day and stay efficient" and "I get interrupted twenty times a day and it wrecks me" can both be true at once. The distinction is who initiated the switch.

Which jobs are fragile to interruption — and which ones are surprisingly resilient

The framework predicts a broad set of occupations, and the predictions do not always match intuition.

Potentially fragile categories. Laboratory research can be interruption-sensitive for two different reasons. Some work depends on irreversible time windows: cell culture, animal studies, and many synthesis steps carry strict reaction times and handling windows, where interruption can compromise a batch rather than merely slowing it. Clinical and surgical settings can also be highly interruption-sensitive, although tolerance varies substantially by task. Legal and audit work involves long-chain logical construction and is structurally similar to programming — a change to one contract clause propagates through the document, and a broken cross-reference chain requires re-validating assumptions. Therapy and counseling add a further layer beyond the cognitive case formulation: an empathic state and a trust relationship, both of which the other person can perceive. Interruption there can do damage that exceeds the efficiency loss. Creative writing, translation, and interpreting lose register and stylistic state; an author's remark about needing to "find the character's voice again" and an engineer's remark about needing to "get back into the mental model of the code" describe the same mechanism. Precision manual work — conservation, watchmaking, fine assembly — depends on motor memory and spatial judgment that require recalibration after a break.

Surprisingly resilient categories. Emergency triage, ticket handling, content moderation, rough-cut editing, and routine image review all look high-intensity but carry low switching cost, because they rely on mature judgment schemas and each unit is self-contained. This is the same mechanism as taking calls.

Emergency department research is especially persuasive here because it measures lost outcomes rather than subjective recovery time. A time-and-motion study published in Quality & Safety in Health Care in 2010 observed 40 physicians across 210 hours in Australia and found an average of 6.6 interruptions per hour, with 11% of tasks interrupted — and 18.5% of those interrupted tasks never resumed. Documentation work attracted the highest concentration of interruptions. The "never resumed" figure makes a different kind of cost visible: some interruptions do not merely delay recovery but leave the original task incomplete.

The hardest case: one person playing both router and processor

High switching cost alone is rarely the decisive factor. The damaging combination is high switching cost plus high interruption probability, which typically arises when one person occupies two roles simultaneously.

A technical founder writes core code while also handling fundraising, hiring, and customer escalations. An early-career academic needs long immersive stretches for research while teaching, administration, expense reporting, and student matters insert themselves at any moment. A clinician's outpatient and ward work is high-frequency switching, while research and manuscript writing demand continuity, and both compete for the same calendar. Staff engineers and engineering managers at large companies oscillate between reviews, answering questions, and deep implementation. Law firm partners face the same structure.

The common feature is that Graham's two time systems are loaded onto one person. Manager's schedule operates in hourly units, where a fragmented meeting costs little. Maker's schedule operates in half-day units, where a single interruption can void an entire block. When both coexist, the conflict is structural rather than a matter of time-management technique.

One organizational implication deserves attention: efficiency in the router role is sometimes purchased with cost in the processor role. A quick "can you just check something" is nearly free for the sender and may represent twenty minutes of re-derivation for the receiver. Costs of this kind are usually invisible inside an organization, which means they rarely appear in anyone's performance accounting. Multiple organizational studies also find that roles absorbing interruptions — coordinating, answering questions, onboarding newcomers — are distributed unevenly across the workforce, and that this work is high-value, hard to measure, and continuously erodes the deep-time budget of the people doing it.

Structural problems call for structural responses.

How to reduce context switching cost: five practices and why they work

All five practices below rest on one principle: convert implicit state into explicit state as early as possible, and convert passive interruption into active switching. Each includes the mechanism and a minimal experiment that can run this week.

1. Design breakpoints so pauses land in a committed state. Split large tasks into units with a defined deliverable, such that each unit ends with complete, externalizable state. The mechanism: recovery cost depends on the completeness of state at the pause point, not on the total length of the task. Minimal experiment: define three commit points for the task you are currently on, and at each one stop for two minutes and write the state down.

2. Leave a recovery anchor: thirty seconds, three lines. Before any interruption, write three sentences — what I am doing, what the next step is, where I am currently stuck. The mechanism: this downgrades recovery from recomputation to reading, and it narrows the scope that needs re-verification from the whole chain to the neighborhood of the anchor. The developer tactics recorded by Parnin and Rugaber — leaving a compile error, using the diff view — are low-fidelity versions of the same idea. Minimal experiment: keep paper beside the keyboard, write three lines before every interruption for one week, and log recovery times.

3. Batch communications and default to asynchronous. Handle messages, email, and ad hoc questions in fixed windows; prefer asynchronous channels whenever synchronous contact is not required. Two mechanisms operate here: interruptions move from a random distribution to a controllable one, and each switch happens at a breakpoint you selected. Minimal experiment: set two fixed communication windows per day, silence instant notifications outside them, and compare the number of deep tasks completed after two weeks.

4. Make time boundaries visible to others. Personal knowledge is insufficient; colleagues need to see the boundary. The mechanism: anticipated interruption imposes cost on its own — even when no interruption actually occurs, people underperform because they avoid entering a deep state. Publishing the boundary reduces both actual and anticipated interruption. Minimal experiment: block two three-hour deep-work periods per week on a calendar visible to your team, then track whether interruption counts fall.

5. Establish an interruption protocol and make the cost measurable. Agree as a team what justifies immediate interruption (safety, production incidents, hard deadlines) and what must go asynchronous, and keep a simple log of interruption counts and recovery times. The mechanism: invisible cost cannot be reallocated; measurable cost creates room for negotiation. Minimal experiment: have the team log interruption source and recovery time for two weeks, then bring the data to a conversation about changing collaboration rules.

Of the five, the second is the easiest to test immediately because it requires little change to the surrounding workflow. It leaves task structure untouched and simply downgrades reconstruction from recomputation to retrieval.

What changes when externalizing context becomes nearly free

Every practice above points at the same act: externalizing context. People have done this for a long time through documents, comments, sticky notes, TODO lists, and meeting minutes. What changed is not that the idea became novel. What changed is its price. Writing context detailed enough for your future self to resume used to be expensive, so people wrote sparingly and recorded only conclusions. Generating some forms of context has become much cheaper, and in some workflows a system can help produce them.

One direct implication follows: the "implicit state" axis is becoming movable. A task that previously held its state only in someone's head gains interruption tolerance once that context is continuously externalized.

A counter-risk deserves stating plainly. Over-externalization can erode the reconstruction capability itself. When reasoning is consistently offloaded to an external system, the opportunity to practice rebuilding state from scratch shrinks — and that practice is part of what deep-work capacity consists of. A second failure mode is subtler: the context sits complete in the system while the person no longer understands it. A checkpoint's value depends on whether its author still comprehends the reasoning inside it, not on whether it was stored.

A defensible position is to treat externalization as an aid to reasoning rather than a substitute for it. What deserves recording is why a conclusion was reached, not only where the work stands. The first preserves reconstructability; the second preserves only progress.

Back to the two people

The account director and the scientist faced interruptions of similar duration but paid very different costs. The difference is better explained by task structure — how much state must be reconstructed, whether units have natural stopping points, and whether the switch was planned or imposed — than by a simple difference in individual ability.

The more useful question is therefore not "can I multitask." It is: which quadrant am I in right now, and can I move the work toward the cheaper one?

The first two determinants respond to redesigning task structure. The third requires changing collaboration rules. All three are addressable, and none of them depend on willpower.

References

Chris Parnin, 2014 Programmer, Interrupted: Data, Brains, and Tools https://www.microsoft.com/en-us/research/video/programmer-interrupted-data-brains-and-tools/
Sophie Leroy, 2009 Why Is It So Hard to Do My Work? The Challenge of Attention Residue When Switching Between Work Tasks https://doi.org/10.1016/j.obhdp.2009.04.002
American Psychological Association, 2023 Why our attention spans are shrinking, with Gloria Mark, PhD https://www.apa.org/news/podcasts/speaking-of-psychology/attention-spans
Westbrook, J. I., et al., 2010 The impact of interruptions on clinical task completion https://doi.org/10.1136/qshc.2009.039255