The Research · Capacity Intelligence™
The evidence that drag is built into the work, not the worker.
Below is the research Capacity Intelligence rests on — graded by how strong each source actually is, from measured behavioral data down to self-report. We show the soft spots on purpose. An evidence base you can't poke holes in isn't evidence; it's marketing.
What all of it points to
Capacity drift is structural, not dispositional — a property of how work is designed and routed, not of who is doing it.
The single cleanest anchor is a 2026 Science Advances study: day-to-day swings in cognitive precision strongly predict same-day output, and the effect shows no moderation by trait-level conscientiousness or self-control. In plain terms, the variation is in the conditions, not the character. Bain's organizational research points the same way from the opposite direction — high- and low-performing companies carry nearly identical shares of star talent, so the performance gap lives in the system, not the staffing.
That is the whole basis for measuring the work demand rather than rating the people. Everything below either supports that claim or sizes the cost of ignoring it.
The evidence, graded by strength
Not all research is equal. Measured behavioral data beats large surveys, which beat self-report. We sort every source into one of three tiers and name its limitation, so you can weight it yourself.
Tier 1 · strongest
Measured behavioral data
Instrumentation and peer-reviewed observation — what people actually did, not what they said they did.
Wilson & Hutcherson, Science Advances (2026)
A 12-week longitudinal study (N=184): a one-standard-deviation swing in daily cognitive precision was worth roughly 40 minutes of work, with no moderation by conscientiousness or self-control. The structural-not-dispositional anchor.
Limitation: sample is university students and the outcome is self-reported goal achievement; applying it to enterprise output is a reasonable inference, not a measured fact.
Microsoft Work Trend Index (2025)
Telemetry across millions of Microsoft 365 users: knowledge workers interrupted roughly every two minutes during core hours, 57% of meetings unscheduled, and a steady spill of work into the evening.
Limitation: the telemetry is real behavioral data, but the accompanying survey is self-report and Microsoft sells the AI tools it positions as the remedy.
Harvard Business Review "toggle tax" study (2022)
Software instrumentation of 137 workers across three Fortune 500 companies: roughly 1,200 app switches per day, about four hours a week lost to reorientation — close to 9% of the work year.
Limitation: small sample, but measured directly rather than self-reported.
Gloria Mark, UC Irvine (attention research)
Peer-reviewed observational work: after a significant interruption it takes about 23 minutes to return to the same depth of focus — the multiplier that turns frequent interruptions into lost days.
Limitation: often quoted loosely; the figure is an average from specific observed settings.
Tier 2 · strong
Large structured research
Big samples and established methods — robust in direction, though not behavioral measurement.
A survey of 300+ large companies: the average company loses 20–25% of its productive capacity to "organizational drag," and high and low performers carry nearly identical shares of star talent.
Limitation: consultancy research tied to a book and methodology, not peer-reviewed.
Rob Cross — Beyond Collaboration Overload
Two decades of organizational network analysis across 300+ organizations: collaborative load has grown about 50% and now fills roughly 80% of the week, and just 3–5% of people absorb 20–35% of value-adding collaboration.
Limitation: the network data is solid; the well-known "18–24% reclaimable" figure is a practitioner estimate from interviews, and the book uses composite illustrative cases.
Probability-based panel data: the average span of control rose to 12.1 direct reports (from 10.9 a year earlier, a 50% increase since 2013), with 45% of middle managers reporting burnout — the highest of any group.
Limitation: burnout is self-reported, though the sampling is rigorous and transparent.
CUNY / Johns Hopkins, Am J Prev Med (2025)
A peer-reviewed cost model of employer burnout that yields a per-employee burden in the low-thousands-to-low-tens-of-thousands range — the independent benchmark our calculator's floor is cross-checked against.
Limitation: a model built on input assumptions; a benchmark to triangulate against, not a measurement of any one firm.
Tier 3 · directional
Self-report & vendor-commissioned
Useful for direction and scale, but built on what people report and often funded by a party with a remedy to sell. We never rest a number on these alone.
BCG Henderson Institute — "AI brain fry" (HBR, 2026)
A survey of 1,488 workers: productivity rose with up to three concurrent AI tools and dropped at four or more, with affected workers reporting 39% more major errors and a jump in intent to quit.
Limitation: productivity is self-reported; the specific oversight-load multipliers in our calculator are our modeling assumption anchored on this direction, not a BCG measurement.
McKinsey decision-making survey (1,200+ leaders)
Executives spend roughly 40% of their time making decisions and say 61% of that time is used ineffectively; only about 20% rate their organizations as good at it.
Limitation: self-report; the widely cited "$250M per Fortune 500 company" is McKinsey's own labeled thought experiment.
Knowledge workers estimate about 60% of the day goes to "work about work" — coordination, status-chasing, and tool-switching — rather than skilled work.
Limitation: self-estimated time allocation (notoriously unreliable) and vendor-commissioned.
Stanford / BetterUp "workslop" (HBR, 2025) & MIT Media Lab "cognitive debt" (2025, preprint)
Stanford/BetterUp put the downstream cost of decoding AI-generated work at roughly $186 per affected worker per month. MIT used EEG to measure up to 55% lower neural connectivity in AI-assisted writers.
Limitation: the workslop figure is survey-based; the MIT result is an early, small preprint — we flag it as emerging and deliberately leave its long-horizon effect out of our numbers.
Why fixing the people only gets you so far
Cross's work produces the most useful single number for understanding where the leverage is. Individuals who manage their own collaboration well reclaim only so much time before they hit a wall built by the system around them.
of collaborative time is the most an individual can win back through personal behavior change.
Past that ceiling sit constraints no one can move alone: meeting-calling norms, response-time expectations, decision rights, and one's own position as a network hub. That residual is the demand-design problem — and it is why the durable gains come from redesigning how work is routed, not from asking depleted people to try harder. It is also why a tool that only coaches individuals leaves most of the value on the table.
How we use this research
We treat everything above as corroboration of the pattern — evidence that execution drag is real, costly, and structural. We do not treat it as a measurement of your organization.
The numbers in our calculator and audit come from your operating data and your inputs, not from these studies. The research tells you the pattern exists and roughly how big it tends to be; only your data tells you what it costs you. Where a source has a limitation — a student sample, self-report, a preprint — we say so here rather than hoping you don't check.
That is the whole posture: measure it, don't assert it.
See what the pattern is costing your operation.
The research says the drag is structural and expensive. The next step is finding out whether it's showing up in your team — and what it's worth to fix.