The state of deep work.
We read every major study on deep work across eight knowledge professions. Here are our findings.
If you run a business, you've almost certainly been told that your team needs more deep work. You've probably read that it takes 23 minutes to refocus after an interruption.¹ You may have seen statistics claiming knowledge workers lose 2.1 hours a day to interruptions, or that they're interrupted 87 times daily. You've definitely encountered confident claims about how many hours of "real" focus a knowledge worker gets in a day.
Most of this is wrong, misattributed, or built on a foundation that doesn't exist.
We spent several weeks reading the actual source material behind the productivity industry's most-repeated claims, across eight professions:
Founders and CEOs
Developers
Designers
Product managers
Marketers
Operations and chiefs of staff
Salespeople
Consultants
What we found wasn't what the productivity literature suggests. There's no shared definition of deep work. There's barely any data for three of the eight roles. And three of the most-cited statistics in the entire field are misrepresentations of what the original researchers actually found, or untraceable to a real source at all.
This report is the result. It's written for founders and operators of small and medium-sized businesses who want to know what the research genuinely supports and a better definition of deep work.
What is deep work?
Common Misconceptions
The Gloria Mark "23 minutes" claim
This is probably the most-cited number in productivity writing. It's usually attributed to her 2008 paper, The Cost of Interrupted Work: More Speed and Stress.² There are two problems with this.
Problem one: the figure isn't from that paper. It comes from her 2005 paper, No Task Left Behind?, a field observation of 24 information workers.³ The actual finding was 23 minutes and 15 seconds to return to the original task, typically after working on two intervening tasks first. It isn't "time to refocus." It's "time to come back, after wandering through other work in between." Very different claim. (We've written separately about what the interruption recovery research actually shows for anyone who wants the properly-sourced version.)
Problem two: the 2008 paper found something the productivity industry has buried. Interrupted workers in that study completed tasks faster than the uninterrupted baseline. They reported significantly higher stress, frustration, and effort, but they didn't lose productivity. The "interruptions destroy your output" framing inverts the finding.²
The Basex Research numbers
You'll often see two statistics cited together: knowledge workers lose 2.1 hours a day to interruptions, and they're interrupted 87 times per day. These come from a 2005 research firm called Basex.
The problem is that the original methodology has never been publicly retrievable. The numbers circulate freely in productivity writing, business books, and consulting decks, but the sample size, study design, and how interruptions were defined aren't accessible.
The McKinsey Chief of Staff claim
You may have seen this one: "McKinsey found 56% of Chiefs of Staff work over 50 hours per week, and 25% work over 60."
We searched for the original source. It doesn't appear in any McKinsey publication. McKinsey's only Chief of Staff paper analyzed public records of around 250 Chiefs of Staff, and it contains no data on hours worked at all.⁴ The 56% figure is likely a corrupted reference to a job-market statistic or a compensation survey from elsewhere. Either way, it shouldn't be cited.
These are load-bearing facts for an entire genre of business writing. If three of the most-repeated focus statistics are wrong or untraceable, the broader literature deserves more scrutiny than it usually gets. So here's what we found…
Various professions compared
Here's the deeper problem. When the productivity industry talks about deep work, it implies a shared definition but there isn't one.
Different research bodies measure entirely different things:
Software engineering research measures time-to-resumption after interruption. Typically 10 to 15 minutes to begin editing again, and 15 to 30 minutes to reattain a flow state.⁵
Sales research measures the percentage of the working week spent on actual selling rather than admin.⁶
CEO research measures meeting share of total time. The most rigorous study found 72% on average across 27 large-company CEOs.⁷
Consultant research mostly measures total weekly hours.
Designer research measures the share of time lost to non-creative administrative work.
None of these are the same thing. A developer's "deep work" is the absence of synchronous interruption during an edit state. A salesperson's is uninterrupted time on prospect research. A CEO's is the protected agenda block that hasn't been hijacked by reactive demand. The fragmentation problem looks structurally different in each role.
This matters for a practical reason. When you read "knowledge workers only get two hours of real work done a day," that figure is built on aggregated telemetry from a vendor's customer base. It's not a measurement of your team. The number you see is shaped by who happened to be in the dataset.
The stats for each profession
Across the eight roles, the depth of available research varies by roughly an order of magnitude. Here's what each body of evidence actually shows.
Developers (well-researched)
The most studied profession in this entire report. The findings:
Parnin and Rugaber's analysis of 10,000 programming sessions across 86 programmers found it takes 10 to 15 minutes to begin editing after resumption. Only 10% of resumptions during an active edit recovered in under one minute.⁵
Iqbal and Horvitz found programmers need at least 7 minutes to transition from high to low memory state.⁸ This is the cognitive cost of just shifting attention, before any new work begins.
Amoroso d'Aragona and colleagues (2023) found that context switching and breaks correlate with measurable code quality degradation, more bugs, and technical debt accrual.⁹ This is the only major study that quantifies the downstream cost of fragmentation rather than just measuring fragmentation itself.
On AI tooling, the evidence cuts in different directions. A lab study of 95 developers found Copilot users completed an HTTP-server task 55.8% faster.¹⁰ Larger observational studies of nearly 2,000 developers at Microsoft and Accenture found smaller gains: 13–22% more pull requests per week at Microsoft, 7–9% at Accenture, with substantial noise.¹¹ A controlled Microsoft study of 200+ engineers found self-reported productivity gains but no telemetry change in lines of code, time spent coding, or PR activity.
DORA's 2025 State of AI-Assisted Software Development found 39% of developers report little or no trust in AI-generated code despite productivity gains.¹²
CEOs (moderately well-researched)
Two strong empirical studies exist:
Porter and Nohria tracked 27 CEOs of large companies (average revenue $13.1 billion) over 13 weeks, in 15-minute increments. The headline: 72% of time in meetings, average 37 meetings per week, 9.7 hours per weekday. But the more interesting finding is variance: CEOs spent between 14% and 80% of their time on agenda-furthering activities, with an average of 43%. The rest was reactive (36%), routine (11%), and crisis (1%).⁷
Bandiera and colleagues analyzed 1,114 CEOs across six countries and identified two CEO types using statistical modeling: "manager" types and "leader" types. The leader type correlated with higher firm productivity. A surprising finding: bottom-quartile CEOs spend about 40% of their time in meetings, while top-quartile CEOs spend 65%.¹³ More meetings, not fewer, correlated with better firm outcomes at the CEO level.
The Porter and Nohria data predates COVID and is based on large-company CEOs. Neither study tells you anything reliable about founders or SMB CEOs.
Sales (consistent industry data)
The most-cited finding: sales reps spend only 28–30% of their working week on actual selling. This comes from Salesforce's State of Sales Report, which surveyed 7,775 sales professionals across the Americas, APAC, and Europe using a double-anonymous third-party panel.⁶ The figure has been reaffirmed across multiple editions from 2022 to 2026.
Other significant findings:
HubSpot's 2024 Sales Trends Report found reps spend approximately 2 hours per day actively selling.
Salesroom's 2024 B2B Sales Performance research found 68% of sellers say note-taking and data input are their most time-consuming tasks, and 43% report 10 to 20 hours per week on administrative work.
Forrester's activity study of 3,031 sales reps across multiple industries found nearly two full days per week spent on admin.
On AI: Salesforce's 2026 data shows 54% of sellers have used AI agents, with users expecting a 34% reduction in prospect research time and 36% reduction in email drafting time. But 78% of sellers missed their quota in 2025, up from 69% in 2024, despite widespread AI adoption.⁶ AI changed the work without solving the quota problem.
Product managers (thin, vendor-dominated research)
The most-cited study is the Pragmatic Institute's 2019 Annual Product Management Survey of approximately 2,500 respondents. The findings:
PMs spend only 27% of their time on strategic activities. The remaining 73% is tactical.¹⁴
Individual contributors say they should be at 51% strategic; executives say 60%. Reality is roughly half that.
65% of PMs work more than 40 hours per week. 42% of PM executives work more than 50.
74% spend fewer than 5 hours per month with customers, against Pragmatic's recommendation of 8 hours per week.
Marty Cagan of Silicon Valley Product Group has argued qualitatively that PMs need "on the order of four solid hours a day" of product discovery work.¹⁵ Combined with the calendar reality the Pragmatic data describes, this is why PMs routinely report 60-hour weeks. There is no telemetry study of PM focus time.
Marketers (thin, vendor-dominated research)
The CMO Council's 2023 survey of approximately 200 CMOs found:
45% spend most of their time reviewing and approving plans, budgets, and campaigns.
42% spend significant time in internal meetings.
Only 1 in 6 spend significant time on global business strategy with leadership.
66% would prefer to spend their time on global business strategy. Only 1 in 6 prefer their current allocation.¹⁶
The Microsoft Work Trend Index 2023 found marketers are particularly meeting-affected: 32.9% cite meetings as their largest distraction, compared to roughly 10% across all roles. The Duke Fuqua CMO Survey found 56.4% of purchased martech tools are actually used, suggesting tool sprawl is part of the focus problem.
Slack's Workforce Index found marketers among the earliest GenAI adopters, with daily AI users self-rating 41% "very good" productivity versus 25% for non-AI users. Self-reported, not telemetry-confirmed.
Designers (thin, vendor-dominated research)
The InSource and inMotionNow 2019 In-House Creative Management Report (500+ respondents) found:
48% of in-house creatives spend about one day per week or more on non-creative administrative work, up 14% year-over-year.
22% spend 10 or more hours per week chasing information, feedback, and approvals.
72% say "obtaining the necessary information just to get started" is the largest admin time sink.
Figma's 2025 Shifting Roles report (1,199 survey respondents plus 51 qualitative interviews) found 72% cite AI tools as the primary driver of role shift, though this measures role boundary blurring rather than focus time. No academic peer-reviewed dataset on designer focus time exists.
Consultants (working-hours data, no focus-time data)
The available data is dominated by hours rather than focus:
MBB consultants typically work 50–80 hours per week, with a typical week of 65–75 and bad weeks reaching 100.¹⁷
Memtime's 2023 data found consultants average 9.3 extra hours per week beyond their contracted hours. Strategy consultants average 20 extra hours.
Daily breakdowns suggest 8.1–8.7 hours per day for consultants, with partners working 12 or more.
On AI: McKinsey has deployed an internal tool called Lilli with over 100,000 documents and interviews indexed. BCG has invested heavily in similar tooling. None of the firms have published focus-time impact data. No academic peer-reviewed dataset on consultant focus time, interruption frequency, or flow exists.
Operations, Chief of Staff, and Founders (no data)
This is the empty quarter. For Chiefs of Staff, compensation surveys exist (Chief of Staff Association, Chief of Staff Collective, Chief of Staff Network all publish annually). Time-use surveys do not.
The only major analytical work is McKinsey's 2024 Anatomy of the role in eight charts, which analyzed public records of approximately 250 Chiefs of Staff. It contains no hours data. The most relevant findings:
Median CoS tenure is 2.3 years, compared to 4.8 years for large-company CEOs.
Only 9% of Chiefs of Staff formally hold other positions in addition to their CoS duties.
McKinsey describes the role verbatim as "typically all-absorbing."⁴
For operations managers and business operations leaders, no role-specific quantitative data on meeting load, focus time, or context switching exists in published research.
For founders specifically, every "founder time" claim you've read is either anecdotal, drawn from large-cap CEO research that doesn't apply to early-stage operators, or invented. The role most likely to be reading this report is the role with the least empirical research behind it.
What the research shows
If you stitch together the credible findings, the picture looks like this.
Across every profession with credible data, meeting load has risen since 2020 and after-hours work has normalized. The headline numbers:
Microsoft's 2023 Work Trend Index, fielded across roughly 31,000 workers in 31 countries, found that 68% of people say they don't have enough uninterrupted focus time during the workday.¹⁸
RescueTime telemetry covering more than 50,000 knowledge workers found that 40% of them never get a single 30-minute uninterrupted block in a working day.¹⁹
Slack's Workforce Index, surveying 10,333 desk workers across the US, Australia, France, Germany, Japan, and the UK in 2023, found workers want around four hours of focus time per day, and consider more than two hours of meetings the point at which they feel overburdened.²⁰
Most working days don't deliver either.
Microsoft's 2025 Breaking Down the Infinite Workday report makes the much-quoted claim that workers are interrupted every 2 minutes by a meeting, email, or notification. This figure has a methodological caveat the report itself notes: it reflects the top 20% of users by ping volume received, and is calculated across the full 24-hour day rather than the 8-hour workday. It does not describe all knowledge workers, despite being cited as if it did.
Has AI fixed this?
The AI question is the most genuinely unresolved part of the picture. The evidence:
Lab studies of GitHub Copilot show developers completing tasks 55% faster.¹⁰
Larger observational studies show smaller gains: 7% to 22% more pull requests per week, with substantial noise.¹¹
A controlled Microsoft study of 200+ engineers found self-reported productivity gains but no telemetry change.
In sales, AI users expect significant time savings on research and email drafting, but quota miss rates climbed from 69% to 78% between 2024 and 2025.⁶
In marketing, daily AI users self-rate higher productivity, but the data is self-reported and the comparison isn't controlled.
The summary: AI has changed the work without resolving the time problem. The workday got longer, focus didn't improve, and the tools that were supposed to fix this have produced mixed results.
Our objective definition of deep work
If you want to actually measure deep work, the literature points toward a specific threshold.
Not 30 minutes. Too short to be useful. Developer research shows it takes 10 to 15 minutes just to begin editing again after a context switch, and 15 to 30 minutes to reach flow.⁵ Iqbal and Horvitz's 7-minute memory-state transition adds to this. A 30-minute block can be entirely consumed by re-entry.
Not 90 minutes. This is the ultradian-rhythm number from popular productivity writing, without empirical grounding in the focus literature. We've examined the 90-minute claim in more detail, tracing it back to its origins in sleep research that doesn't actually support the way it's applied to daytime focus.
The defensible objective definition, drawn from the convergence of the developer interruption research⁵,⁸ and the working-sphere data,³ is this:
A continuous block of at least 45 minutes during which:
The primary active application is in a role-defined cognitively demanding category
No synchronous communication interruption is responded to
No application switch exceeding two minutes of dwell-time in a non-task category occurs
Three things this definition explicitly excludes:
Self-reported flow state. Correlates weakly with measured output. The Microsoft "Dear Diary" study found self-reported productivity gains with no telemetry change.
Physiological signals. Züger and Fritz's work on heart rate and eye-tracking for flow detection is rigorous, but not scalable to cross-profession measurement.
Output metrics. Lines of code, pull requests, deliverables. These conflate focus availability with productivity, and punish thinking time that doesn't produce immediate artifacts.
It's a measure of opportunity for focus, not its quality.
The "role-defined cognitively demanding category" is the part most popular definitions skip, and the part that makes the metric comparable across roles. Coding for a developer. Editing-mode Figma for a designer. Writing and analysis for a marketer. Prospect research for a salesperson, not CRM data entry. Agenda-furthering work for a CEO.
Managing time inside a session
If you're running a business with 5 to 200 people, the implication isn't the standard productivity-industry advice about blocking your calendar. The implication is two things:
The research available to guide you is thinner than the confident claims suggest.
The structural focus problem in your team is probably specific to the roles you employ, rather than generic to "knowledge work."
A developer-heavy team has an interruption problem during edit states. A sales-heavy team has an admin and CRM problem. A marketing team has an approval and meeting load problem. A PM-led team has a cross-functional integrator problem. A founder-led team has a problem nobody has measured. These aren't the same problem, and the same intervention won't solve them.
The honest summary of months of research is this. Deep work is a useful concept. The data behind it is weaker than the productivity industry pretends. And until the literature converges on a shared, measurable definition (something like the 45-minute threshold above), most confident claims about focus availability are essentially incomparable.
We'd rather tell you that than tell you what you've already heard.
FAQs
Is deep work the same as flow?
No. Flow is a psychological state defined by Csikszentmihalyi: complete absorption, loss of self-consciousness, distorted sense of time. Deep work, as Newport defined it, is cognitively demanding work performed in distraction-free concentration. The two overlap heavily but aren't identical. You can do deep work without being in flow, and you can be in flow doing something that isn't deep work by Newport's definition. The two terms get used interchangeably in productivity writing, which is part of the measurement problem this report describes.
Why did you exclude self-reported flow state from your definition?
Self-report correlates weakly with measured output. The clearest example is the Microsoft "Dear Diary" study of 200+ engineers using GitHub Copilot. Engineers self-reported significant productivity gains, but telemetry showed no measurable change in lines of code, time spent coding, or pull request activity. Self-report data is useful for wellbeing research, but it shouldn't be the basis of an objective focus metric.
Doesn't 45 minutes feel arbitrary? Why not 60 or 90?
It's not arbitrary. It's the smallest block that allows for full cognitive re-entry plus meaningful sustained work, based on the developer research. Parnin and Rugaber found 10 to 15 minutes to begin editing after a switch, and 15 to 30 minutes to reach flow. A 30-minute block can be entirely consumed by re-entry. Forty-five minutes is the threshold where the entry cost is paid and useful work happens. Sixty minutes is fine but more demanding. Ninety minutes is the popular ultradian-rhythm number, which has no empirical grounding in the focus literature.
Why are there no estimates for founders?
Because the data doesn't exist. There is no published time-use study of founders as a distinct population. The CEO research everyone cites is based on large-company CEOs (Porter and Nohria's sample averaged $13.1 billion in revenue) and tells you nothing reliable about someone running a 20-person startup. We could have invented a number based on intuition. We chose not to.
Hasn't AI fixed the focus problem?
The evidence is genuinely mixed. Lab studies show large productivity gains. Larger observational studies show smaller gains with substantial noise. A controlled Microsoft study showed self-reported gains but no telemetry change. In sales, AI adoption has been widespread but quota miss rates climbed from 69% to 78% between 2024 and 2025. The honest summary is that AI has changed the work without resolving the time problem.
What should I actually do about this in my business?
The next Original in this series audits the interventions the way this one audited the research. The short answer is that the structural focus problem in your team is probably specific to the roles you employ, and the same generic intervention won't work across all of them. A developer-heavy team has an interruption problem. A sales-heavy team has an admin problem. A marketing team has an approval problem. The intervention should match the structural cause.
Related reading:
Attention residue — The cognitive cost of leaving one task incomplete when you start another. The mechanism behind why uninterrupted blocks matter more than total hours worked.
Interruption recovery cost — The shorter, sourced version of the Mark research this report corrects. If you want the original studies properly summarised without the misattributions, start here.
Flow state conditions — What Csikszentmihalyi actually defined, and how flow differs from "deep work" in the popular usage. Necessary context for the definitions section above.
The 90-minute myth (ultradian rhythms) — Why the popular 90-minute focus block has weaker empirical grounding than its cultural footprint suggests. Same source-checking discipline applied to a different number.
Pseudo-productivity — Cal Newport's term for the practice of using visible activity as a proxy for useful output. Useful framing for understanding why so much of the productivity literature defaults to easily-measured-but-unhelpful metrics.
References
Mark G, Gudith D, Klocke U. The cost of interrupted work: more speed and stress. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '08). New York: ACM; 2008. p. 107–110.
Mark G, Gudith D, Klocke U. The cost of interrupted work: more speed and stress. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '08). New York: ACM; 2008.
Mark G, Gonzalez VM, Harris J. No task left behind? Examining the nature of fragmented work. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '05). New York: ACM; 2005. p. 321–330.
Goodman A, Rochford C, Ramsay Gray L, Johnson R. Chief of staff: anatomy of the role in eight charts. McKinsey & Company; 2024 Aug 29.
Parnin C, Rugaber S. Resumption strategies for interrupted programming tasks. Software Quality Journal. 2011;19(1):5–34.
Salesforce. State of Sales Report (5th edition, 2022 fielding; 6th edition, 2024; 7th edition, 2026). San Francisco: Salesforce; 2023–2026.
Porter ME, Nohria N. How CEOs manage time. Harvard Business Review. 2018 Jul–Aug;96(4):42–51.
Iqbal ST, Horvitz E. Disruption and recovery of computing tasks: field study, analysis, and directions. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '07). New York: ACM; 2007.
Amoroso d'Aragona D, et al. Context switching impact on software quality. Empirical Software Engineering; 2023.
Peng S, Kalliamvakou E, Cihon P, Demirer M. The impact of AI on developer productivity: evidence from GitHub Copilot. arXiv:2302.06590; 2023.
Cui Z, Demirer M, Jaffe S, Musolff L, Peng S, Salz T. The productivity effects of generative AI: evidence from a field experiment with GitHub Copilot. MIT Generative AI Symposium; 2024.
DORA. Accelerate State of DevOps Report 2024 and State of AI-Assisted Software Development 2025. Google Cloud; 2024–2025.
Bandiera O, Hansen S, Prat A, Sadun R. CEO behavior and firm performance. Journal of Political Economy. 2020;128(4):1325–1369.
Pragmatic Institute. Annual Product Management Survey 2019. Phoenix: Pragmatic Institute; 2019.
Cagan M. Coaching: managing time. Silicon Valley Product Group; published online.
CMO Council. State of marketing leadership survey. San Jose: CMO Council; 2023.
Industry surveys aggregated from Strategycase, Casecoach, Memtime, and MyConsultingOffer (2020–2025). Note: practitioner publications, not peer-reviewed.
Microsoft. Will AI fix work? 2023 Work Trend Index annual report. Redmond: Microsoft; 2023.
RescueTime. The state of work life balance in 2019 (and ongoing telemetry analyses). Seattle: RescueTime; 2019.
Slack Workforce Lab. Workforce Index winter 2023 report. San Francisco: Slack Technologies; 2023.