If you want to know whether AI changed the work at your company, the comparison group matters more than the size of the dataset. One of the largest observed records of AI at work says adopters and non-adopters diverged: AI users' daily focus time fell 9% while non-users showed virtually no change. That contrast is worth more than any before-and-after number, because both groups worked through the same conditions. It is also not the end of the argument, because employees chose whether to adopt. Here is what the comparison supports, what it does not, and what to do about recovered capacity that nobody has claimed.
Two things are true at once. AI is saving people time on individual tasks, and the workday is not getting lighter. Both are visible in the same data.
What the 443M-hour dataset actually measured
ActivTrak's Productivity Lab published its fifth annual State of the Workplace report in March 2026, drawn from more than 443M hours of behavioral activity across 1,111 companies and 163,638 employees, recorded between January 1 2023 and December 31 2025. This is observed activity rather than survey response, which is why it is worth reading closely.
Adoption is settled. 80% of employees now use AI tools, up from 53%. Monthly retention has averaged 92% since January 2024 and never dropped below 88%, so this is habit rather than trial. Average time in AI tools rose 8x, from 0.1% of total hours to 0.8%.
Hold that last pair together. Adoption is near universal and depth is still under 1% of the working day.
The before-and-after numbers, and why they cannot stand alone
A second dataset covers 10,584 users across 376 companies, comparing behavior 180 days before and 180 days after AI adoption. Every measured work category increased. Email rose 104%. Chat and messaging rose 145%. Business management rose 94%. No category decreased.
That looks conclusive and it is not. A before-and-after comparison of one group is what Campbell and Stanley classified in 1963 as the one-group pretest-posttest design, and its first named weakness is history: events occurring between the two measurements produce differences that get credited to the treatment.
The history here is not hypothetical. Across the same three years, in the full dataset, collaboration rose 34% and multitasking rose 12% for everybody. Some of the post-adoption increase is the period, not the product, and a one-group comparison cannot say how much.
The comparison that answers the period objection
The interesting cut gets less attention. ActivTrak also reports two adopter versus non-adopter findings:
- AI users' daily focus time declined 9%, compared with virtually no change for non-users. Focus efficiency is ActivTrak's term for the share of work time spent in focused, uninterrupted activity.
- AI users spent 9% more days in healthy utilization patterns than non-users, meaning days inside the employer's own productive-hours band rather than above or below it.
Those point in opposite directions, and both belong in an honest summary. On workload balance, adopters look better than their colleagues. On sustained attention, they look worse, and the flat non-user line is what makes the adopter decline readable rather than ambient.
This is the same structure we use in our own measurement methodology: compare adopting teams to non-adopting teams, on observed behavior, and disclose the limits.
What the comparison still cannot rule out
Selection. Nobody assigned these employees to adopt. People who reach for AI early may already sit in more interrupt-driven roles, in which case the focus gap describes who adopts rather than what adoption does. That threat closes by controlling for tenure, role and tool stack, or better, by deciding rollout order at random where the order is genuinely discretionary. ActivTrak also does not publish the observation window for this cut, so treat it as a same-dataset comparison rather than a period-matched one.
State the claim, then the number, then the limit. The claim is that adopters and non-adopters diverged on focus. The number is 9%. The limit is that the groups formed themselves.
Where did the recovered capacity go?
The report's own conclusion is blunt: there is no evidence that AI frees capacity at this time.
The mechanics are visible. The average workday shortened by 9 minutes, from 8h 53m in 2023 to 8h 44m in 2025, while productive hours inside it rose 5% to 6h 36m and productive sessions rose 13%. Shorter and denser at once. Focus efficiency fell to 60%, a three-year low, and the average focus session now runs 13m 7s, down 9% since 2023.
The maintenance layer nobody budgeted
Self-report fills in part of the gap. The Work AI Index 2026, published by Glean's Work AI Institute with academic co-authors at seven universities, surveyed 6,000 full-time digital workers in the United States, the United Kingdom and Australia between December 2025 and January 2026. 87% use AI at work and 75% say it makes them more productive, saving roughly 11 hours a week through automation. Only 13% say their organization is performing significantly better as a result.
The same respondents report spending an average of 6.4 hours a week on what the authors call botsitting: supplying missing context, checking outputs, rerunning prompts and correcting confident errors. That is self-reported, and the report discloses that its sample skews toward heavier AI users while stating that its main findings hold after adjusting for role, industry, demographics, employment status and organization size. Note also that Glean sells enterprise AI software, as ActivTrak sells workforce analytics. Read the method, not the headline.
Two different instruments, one observed and one self-reported, both vendor-published, land in the same place. Hours are being recovered at the task level and consumed again before they reach the organization. That is the productivity gap we have written about before, seen from the behavioral side.
Burnout is down. Chronic underuse is up.
The wellbeing numbers are genuinely good. 75% of employees maintained healthy work patterns, up from 66% and a three-year high. Burnout risk fell 22%, to 5% of employees. Overutilization fell from 12% to 7%.
The capacity numbers move in two directions at once, and the definitions explain why. Point-in-time underutilization fell from 22% to 18%. Disengagement risk, which ActivTrak defines as spending more than 75% of the work year in an underutilized state, rose from 19% to 23%, a 4 point increase across the window with most of it falling between 2024 and 2025.
Fewer people are underutilized on any given day. More people are underutilized chronically. ActivTrak connects the two directly: as overload decreases, freed capacity is not automatically redirected toward higher-value work.
What each kind of evidence can actually support
| Design | Example in 2026 | What it supports | What it cannot support |
|---|---|---|---|
| Self-reported time savings | Work AI Index 2026, n=6,000, fielded Dec 2025 to Jan 2026 | Perceived benefit, friction points, where maintenance labor accumulates | Any claim about output or financial return |
| Before and after, one group | ActivTrak Data Set B, 10,584 users, 180 days either side | That behavior changed around adoption | Attributing the change to the tool, because history is uncontrolled |
| Company versus company | PwC 2026 Global AI Jobs Barometer, over 1 billion job ads | A market-level pattern across firms | What would happen at your company, because the groups are self-formed |
| Adopters versus non-adopters | ActivTrak AI user and non-user cuts | A readable divergence between two groups in one dataset | Certainty, because selection remains open |
PwC's June 15 2026 Barometer illustrates the third row, and it also shows how easily a row gets misread. Companies in the most AI-exposed sectors recorded 34% productivity growth in 2025 relative to 2018, against 24% for the companies least able to use AI. Within that group, the top 20% averaged 163% labor productivity growth relative to 2018, which PwC describes as nearly 5 times the most AI-exposed companies overall. Real, large, and still not an instruction. The 163% belongs to a top quintile, not to AI adopters generally, and those firms differed from the comparison group in capital, sector and management long before AI entered the picture.
Four things to do with this
- Pick the comparison group before the rollout, not after. Choosing who to compare against once you have seen the results is how organizations find the answer they wanted.
- Measure work categories, not tool time. AI time averages 0.8% of the working day. The movement showed up in email, messaging and coordination.
- Count the maintenance labor as labor. Review, correction and re-prompting are work. If they sit outside the capacity model, every AI business case overstates its return.
- Give recovered capacity an owner. Freed hours default to absorption. Without an accountable owner for redeploying them, the day refills on its own, which is what a shorter and denser workday looks like.
That last point is the human capital operating system argument in one line. Capacity you cannot see is capacity you cannot reallocate, and the signals that would show it sit in different systems owned by different functions.
Levos sits above that stack as the intelligence layer, deriving measurement from observed behavior in the tools people already use rather than from surveys, and attaching a confidence rating to every score. Adoption counts stay on the input side, where they belong. AI Impact treats AI adoption as a measurable signal rather than a proxy for return, and the tool sprawl visible here, 7 AI platforms per organization on average, is the same fragmentation that shows up in SaaS stack intelligence. The longer version sits in how to measure AI ROI across your workforce.
Frequently asked questions
Does AI free up capacity?
Not so far, on the evidence that exists. ActivTrak states plainly that there is no evidence that AI frees capacity at this time. In its before-and-after subset, every measured work category increased and none decreased. Individual hours may well be saved on individual tasks. What the observed data does not show is those hours leaving the workday.
Why does comparing AI users to non-users matter more than comparing before and after?
Because a before-and-after comparison of one group cannot separate the tool from everything else that changed alongside it. Collaboration rose 34% and multitasking rose 12% across all employees between 2023 and 2025. A comparison group lives through the same conditions, so whatever it also shows is not attributable to the tool.
What did the AI user versus non-user comparison actually show?
Two findings pointing in opposite directions. AI users' daily focus time declined 9% while non-users showed virtually no change. AI users also spent 9% more days in healthy utilization patterns than non-users. Neither is settled, because employees chose whether to adopt.
What is the AI usage sweet spot, and should we set it as a target?
ActivTrak reports that users at 7% to 10% of total hours in AI tools have the highest productivity rate of any tier, 95%, and that only 3% of users are in that band. Read as a description of tiers it is sound. Read as a target it is not, because users sorted themselves into those tiers rather than being assigned.
How should an organization measure whether AI changed the work?
Compare adopting teams to comparable non-adopting teams over the same period, on outcome measures rather than tool time, controlling for tenure, role and tool stack, and publish a confidence level with every number. Aggregate before interpreting: team-level views require 5 or more people.
Start with a comparison, not a dashboard
If your AI program currently reports adoption and spend, the fastest upgrade available is not a better dashboard. It is a comparison group.
Request a Demo to see how controlled cohort analysis with confidence scoring is constructed, or read the measurement methodology first.
Levos is accepting design partner applications from US organizations of 150 employees or more with an active AI rollout. Large organizations typically start with one function or division, measured against comparable teams that have not adopted yet. That is a stronger attribution design than a company-wide before-and-after.
