TL;DR: When you move from one task to another, a slice of your attention stays behind on the first one. Psychologist Sophie Leroy named this attention residue in 2009, and her research found it is worst when the task you just left was unfinished or done under time pressure – which is the exact state a half-finished prompt or a “still generating” window leaves you in. The AI era makes this acute for a simple structural reason: AI returns work in seconds, so a single hour now contains far more handoffs between instructing a machine and collaborating with people than a pre-AI hour ever did, and every handoff leaves residue. The verified costs are large: it takes about 23 minutes to fully return to an interrupted task (Gloria Mark, 2008), switching can burn up to 40% of productive time (APA, 2001), and knowledge workers already toggle between apps roughly 1,200 times a day (HBR, 2022) – all before you add AI, which 75% of knowledge workers now use at work (Microsoft and LinkedIn, 2024). This article turns those findings into an original Human-AI Switch Ledger and a scoreable self-audit. The move: budget your switches the way a CEO allocates scarce capital, and stay enough of a student to learn where your own residue leaks fastest.
You sit down to write something hard. You hand a chunk of it to an AI – “draft the intro” – and while it generates, you flick to Slack, answer a colleague, glance at email, then come back to judge the AI’s output. The whole loop took ninety seconds and felt efficient. It wasn’t. You left at least three tasks half-open, switched cognitive modes four times, and came back to the writing carrying the residue of all of them. The draft on screen is fine; your ability to evaluate it is not. This is the quiet tax of working alongside machines, and almost nobody budgets for it.
The science behind that tax has a name and a clear origin. In 2009, organizational psychologist Sophie Leroy published a paper with a title that sounds like a sigh: “Why is it so hard to do my work?” Her answer was attention residue – the finding that when you switch from Task A to Task B, part of your attention stays cognitively stuck on Task A, and that lingering residue measurably degrades your performance on Task B. The effect, crucially, is strongest when Task A was unfinished or done under time pressure. The AI era did not invent attention residue. It industrialized the conditions that produce it.
What attention residue actually is (and what the research really found)
Leroy’s core claim is narrow and well-supported: you cannot fully bring your attention to a new task while a slice of it is still processing the old one. In her experiments, people who were interrupted mid-task – especially when they expected to have to finish later under time pressure – carried residue into the next task and performed worse on it than people who reached a natural stopping point first. The residue is not a character flaw or a willpower failure. It is a normal property of how attention transitions between goals.
A later study sharpened the practical takeaway. In 2018, Leroy and Theresa Glomb (Organization Science) showed that a brief “ready-to-resume” plan – taking a moment to note where you are and what you will do next before you switch away – significantly reduced attention residue and protected performance on the interrupting task. That single finding is the most actionable lever in this entire article, and we will come back to it.
Attention residue is one specific cost inside a larger, well-documented family of switching costs. Before we build any framework, here is the verified ground truth – each figure traces to the named source.
What the research robustly supports (verified)
| Finding | What the research shows | Source (year) |
|---|---|---|
| Attention residue is real | Switching tasks leaves part of your attention “stuck” on the prior task and lowers performance on the next one; worst when the prior task was unfinished or time-pressured | Sophie Leroy, Organizational Behavior and Human Decision Processes (2009) |
| You can shrink the residue | A brief “ready-to-resume” plan (note where you are and what is next, before switching) significantly reduces residue and protects performance | Leroy & Glomb, Organization Science (2018) |
| Recovery is slow | After an interruption, work was resumed on average in about 23 minutes and 15 seconds, often with two intervening tasks before returning | Gloria Mark et al., The Cost of Interrupted Work, CHI (2008) |
| Attention fragments fast | Average attention span on a screen fell to roughly 47 seconds (median 40s) in recent years, from about 2.5 minutes in 2004 | Gloria Mark, Attention Span (2023) |
| Switching is the tax | Brief mental blocks from switching between tasks can cost as much as 40% of productive time, rising with task complexity | Rubinstein, Meyer & Evans – APA / J. Exp. Psychology (2001) |
| Toggling is already constant | Knowledge workers toggle between apps and windows about 1,200 times a day, spending close to four hours a week (~9% of work time) reorienting | Harvard Business Review study of 137 users across 20 teams (2022) |
| AI is now in the loop | 75% of knowledge workers use generative AI at work, and 78% bring their own tools (“BYOAI”) | Microsoft & LinkedIn, Work Trend Index (2024) |
Read the table as one sentence: human attention was already fragmented and expensive to reload before AI arrived – and AI does not remove a switch, it adds a new partner to switch to.
Why working with AI is uniquely residue-heavy
It is tempting to assume AI reduces switching, because it does some of your work. In practice it changes the structure of your day in three ways that each map directly onto Leroy’s residue triggers.
1. AI multiplies the number of switches per hour. The whole appeal of AI is speed: it returns a draft, a summary, or an answer in seconds. But fast returns mean more opportunities to switch. Where a pre-AI task might have run for twenty uninterrupted minutes, an AI-augmented version becomes prompt, wait, evaluate, reprompt, wait, evaluate – a dozen micro-handoffs in the same window. More handoffs is simply more residue, because residue is paid per switch, not per hour.
2. AI leaves tasks in the exact “unfinished” state residue loves. Leroy found residue is worst when you leave a task incomplete. A prompt that is still generating is the platonic ideal of an unfinished task: you have started it, you cannot finish it yet, and you switch away to fill the wait. When you return, you are resuming an open loop – the highest-residue condition there is.
3. AI forces a cognitive mode switch, not just a task switch. This is the part most people miss. Working with AI is not one activity; it is two opposite ones. Generating with AI is fast, fluent, and uncritical – you are in a permissive, productive register. Evaluating AI output requires the opposite stance: skeptical, slow, looking for the error. Flipping between “create” mode and “critic” mode many times an hour is a deeper switch than moving between two similar tasks, and it carries heavier residue. And layered on top is the social register switch the title points to: instructing a machine is transactional, while collaborating with a colleague is relational, and bouncing between the two leaves residue of its own.
None of this means AI is bad for focus. Used well, it is leverage. It means AI changes where the cost lives: the work gets faster, but the switching surface gets larger, and that surface is where attention residue accumulates. A CEO would say the unit economics changed – and you do not manage a changed cost structure with the old habits.
The Human-AI Switch Ledger
Here is the original framework at the center of this article. A ledger lists where a resource is spent so you can see the leaks. The Human-AI Switch Ledger names the five switch types that dominate an AI-augmented day, maps each to why it generates residue (using Leroy’s mechanism), and gives the CEO+Student lever to reclaim it.
This is an analytical framework, not a measured study – it does not claim a precise “this switch costs X minutes” figure, because no such per-switch dataset for human-AI work exists. It is a decision aid built on top of the verified switching-cost research above. Use it the way you would read a household ledger: to find the line that is bleeding.
The Human-AI Switch Ledger (CEOtudent framework, 2026)
| Switch type | What actually happens | Why it leaves residue (Leroy mechanism) | The CEO+Student lever |
|---|---|---|---|
| 1. Delegate-and-drift | You hand a task to AI and fill the wait by jumping to email, chat, or a feed | You leave the original task unfinished and open a second loop – the peak residue condition | Wait inside the task. Stay on the same problem while it generates (outline the next section), or batch prompts so waits overlap, not multiply |
| 2. Generate-to-critic flip | You switch from fluent “create with AI” mode to skeptical “evaluate the output” mode | A deep register switch; create-mode residue contaminates critic-mode judgment, so you under-scrutinize | Separate the passes. Generate several outputs first, then switch once into a dedicated evaluation block with a checklist – not flip per output |
| 3. Reprompt churn | Iterating with AI in many small, partial exchanges that never reach “done” | Each partial exchange is an unfinished task; the loop never closes, so residue compounds | Define “done” before you prompt. Write the acceptance criteria first; stop when met instead of endlessly nudging |
| 4. Tool-hopping | Bouncing between several AI tools and browser tabs within one piece of work | Classic toggling tax (the ~1,200/day problem) plus reorientation cost on every hop | Standardize your stack. Pick one tool per job; close every tab but the active one during deep work |
| 5. Human-to-machine register shift | Alternating between collaborating with people and instructing AI | A social-to-transactional mode switch; relational residue lingers into machine work and vice versa | Cluster by register. Group human collaboration into windows and AI work into others, so you change register a few times a day, not constantly |
Three things jump out once the ledger is on the table. First, Delegate-and-drift (line 1) is usually the biggest single leak, because the “free” waiting time feels like a bonus and is actually the moment you open your most expensive loop. Second, the Generate-to-critic flip (line 2) is the most dangerous, not the most frequent – it is where residue quietly damages your judgment of AI output, which is precisely the human contribution AI cannot replace. Third, reprompt churn (line 3) is the easiest to cut, because most of it disappears the moment you decide in advance what “good enough” looks like.
The CEO move: budget your switches before the day spends them for you
A CEO facing a scarce resource does not try to manufacture more of it; they decide in advance where it gets allocated and protect the highest-value uses from being nibbled away. Attention is that resource, and switches are how it gets spent without anyone deciding to spend it.
The architecture that follows from the ledger is concrete:
- Protect one or two deep blocks where you do not switch registers at all. The highest-judgment work – the evaluation that AI cannot do for you – belongs in a window before the day’s toggling has fragmented your attention. Pay your most important attention bill first.
- Batch your AI work into defined sessions instead of sprinkling it through the day. Sprinkled AI use maximizes the switching surface; clustered AI use lets you stay in one register and pay the mode switch once.
- Treat “generating…” as a stay-put signal, not a switch-away invitation. The single habit that kills the most residue is refusing to drift during the wait. If the wait is genuinely long, batch several prompts so the waits overlap rather than each one triggering a fresh switch.
- Standardize the stack. Every extra AI tool is another window to toggle into. Decide which tool does which job and stop comparison-hopping mid-task.
Notice what is absent from this list: trying harder. Residue is not defeated by discipline any more than it is caused by laziness. It is defeated by architecture – by designing a day in which the expensive switches simply do not happen, so that your scarce attention is, by default, already pointed at the work that compounds.
The Student move: run the ready-to-resume experiment on yourself
The CEO allocates; the student observes and adjusts. The most evidence-backed personal intervention in this whole field is also the smallest: Leroy and Glomb’s ready-to-resume plan. Before you switch away from any task – especially before you drift during an AI wait – spend fifteen seconds writing down exactly where you are and what you will do next. The research found this reliably reduces the residue you carry into the next task. It works because it closes the loop in your head even though the task is objectively unfinished.
Then study your own ledger for one week. Notice three things: which switch type from the table you commit most (that is your biggest leak), when in the day your evaluation of AI output gets sloppy (that is the Generate-to-critic flip charging interest), and which reprompt loops you could have ended earlier with a clear “done.” You are not chasing someone else’s number; you are learning the shape of your own attention so you can allocate it better next week than last. Run it like a CEO, learn it like a student.
The deepest reframe is this: in an era where AI makes producing output nearly free, the scarce and valuable thing is the uncontaminated attention you bring to deciding whether that output is any good. Attention residue is the tax on that exact capacity. The people who will get the most out of AI are not the ones who switch to it the fastest – they are the ones who protect the focused human judgment that switching quietly erodes.
The Attention Residue Self-Audit
Score yourself 0-2 on each line for a typical workday (0 = rarely, 1 = sometimes, 2 = often). This is a self-reflection tool, not a clinical instrument.
- I leave a task open and drift to email/chat/feeds while AI generates. (Delegate-and-drift)
- I accept or reject AI output in the same fast mode I used to create it, without a separate critical pass. (Generate-to-critic flip)
- I keep nudging an AI in small reprompts without a clear definition of “done.” (Reprompt churn)
- I work a single task across many AI tools and browser tabs at once. (Tool-hopping)
- I bounce between talking to people and instructing AI many times an hour. (Register shift)
- I switch away from hard tasks without noting where I was or what is next. (No ready-to-resume)
- By mid-afternoon my judgment of whether AI output is good gets noticeably worse. (Residue interest)
Reading your score. 0-4: your switching surface is well managed; protect it. 5-9: residue is a real drag on your judgment; install the ready-to-resume habit and batch your AI work. 10-14: switching, not workload, is likely your main bottleneck; start with the single highest-leverage fix below.
The single highest-leverage fix, for almost everyone, is line 1: stop drifting during AI waits. It is the most common switch, it opens your most expensive loop, and closing it costs nothing but the decision to stay put.
Frequently asked questions
Is attention residue the same thing as multitasking?
No, and the distinction matters. Multitasking usually means trying to do two things at literally the same time. Attention residue is what happens in the transition between tasks done one after another – the part of your mind that stays behind on Task A when you have already moved to Task B. You can avoid classic multitasking and still bleed residue all day through rapid sequential switching, which is exactly what AI-augmented work encourages.
Does AI actually make focus worse, or am I just disorganized?
Both framings miss the structural point. AI does not directly destroy focus; it enlarges the switching surface – the number of handoffs available per hour – because it returns work so fast. More available switches plus the same human attention equals more residue, unless you deliberately re-architect when and how you switch. It is not a discipline problem; it is a workflow-design problem, which is good news, because workflows are easier to change than willpower.
What is the single most effective habit to reduce attention residue?
The “ready-to-resume” plan from Leroy and Glomb’s 2018 research: before switching away from a task, take fifteen seconds to write down where you are and what you will do next. It reliably reduced residue in their experiments because it lets your mind treat an unfinished task as “parked” rather than “open.” In AI work, apply it the instant before you drift during a wait.
How long does it really take to refocus after an interruption?
Gloria Mark’s 2008 study found that interrupted work was resumed, on average, after about 23 minutes and 15 seconds, and people typically handled around two other tasks before returning to the original. That figure is about returning to the work, and full cognitive re-immersion can take longer. The practical implication: a “quick” two-minute switch is rarely two minutes once you count the reload.
Should I just stop using AI for deep work, then?
No – that throws away real leverage. The move is to cluster AI use rather than sprinkle it. Use AI inside defined sessions where you stay in one register, batch your prompts so waits overlap, and reserve a separate, unhurried block for evaluating its output with a checklist. Used that way, AI adds capability without shredding your attention; used as an always-open window that pings for a decision every thirty seconds, it becomes the single largest source of residue in your day.
Is “deep work” still possible if my job requires constant AI interaction?
Yes, but you have to redefine the block. Deep work in an AI-augmented role is not necessarily long stretches with no tools; it is stretches where you do not switch registers – where you are either creating, or evaluating, or collaborating, but not flipping between all three every minute. Protecting register-stability is the modern version of protecting uninterrupted time.
Sources
Sophie Leroy. Why is it so hard to do my work? The challenge of attention residue when switching between work tasks (Organizational Behavior and Human Decision Processes, 2009) – introduced the concept of attention residue and showed that switching tasks leaves part of one’s attention on the prior task, lowering performance on the next, with the effect strongest when the prior task was unfinished or time-pressured.
Sophie Leroy & Theresa M. Glomb. Tasks Interrupted: How Anticipating Time Pressure on Resumption of an Interrupted Task Causes Attention Residue and Low Performance on Interrupting Tasks and How a Ready-to-Resume Plan Mitigates the Effects (Organization Science, 2018) – demonstrated across four studies that a brief plan to resume an interrupted task significantly reduces attention residue and protects performance on the interrupting task.
Gloria Mark, Daniela Gudith & Ulrich Klocke. The Cost of Interrupted Work: More Speed and Stress (Proceedings of CHI, 2008), University of California, Irvine – found that interrupted work was resumed on average in about 23 minutes and 15 seconds, typically after handling two intervening tasks.
Gloria Mark. Attention Span: A Groundbreaking Way to Restore Balance, Happiness and Productivity (2023) and associated University of California, Irvine research – documenting that average attention span on a screen fell from roughly 2.5 minutes in 2004 to about 47 seconds (median 40 seconds) in recent years.
Joshua Rubinstein, David Meyer & Jeffrey Evans. Executive Control of Cognitive Processes in Task Switching (Journal of Experimental Psychology: Human Perception and Performance, 2001), as summarized by the American Psychological Association – brief mental blocks from switching between tasks can cost as much as 40% of productive time, with costs rising as tasks grow more complex.
Harvard Business Review. How Much Time and Energy Do We Waste Toggling Between Applications? (2022) – a study of 137 users across 20 teams at three large firms found workers toggled between apps and windows about 1,200 times a day, spending close to four hours a week (around 9% of work time) reorienting.
Microsoft & LinkedIn. Work Trend Index Annual Report: AI at Work Is Here. Now Comes the Hard Part (2024), based on a survey of 31,000 knowledge workers across 31 markets – reported that 75% of knowledge workers use generative AI at work and 78% bring their own AI tools to work.
Editorial note: This article is part of CEOtudent’s fully AI-assisted editorial process. The Human-AI Switch Ledger and the Attention Residue Self-Audit are original CEOtudent decision aids – analytical frameworks for managing task switching, not measured studies; in particular, no precise “human-AI switching costs X%” figure is claimed, because no such dataset exists. The supporting figures are drawn from the publicly available sources listed above and were verified as of June 2026. This article is general educational commentary on attention and productivity, not medical, psychological, or clinical advice.
This post is also available in:
