AI may be changing how employees work, but new behavioral data raises an uncomfortable question for managers: what happens when saved time turns into extra work instead of better output?
ActivTrak’s latest research found that heavy AI users spent more time at work than non-users, even though an earlier study had linked AI use with shorter workdays. The contradiction does not settle whether AI improves productivity. It exposes a deeper management problem: leaders need to define what productivity means before time savings become a source of additional demand.
At a glance:
- The latest study covered 9,140 knowledge workers at 467 enterprises over five months.
- Heavy AI users worked between 8.42 and 9.08 hours per day, compared with 7.58 hours for non-users.
- Heavy users had focus sessions 22% shorter than light users.
- Web-based AI users switched apps more frequently than users of embedded AI tools.
- The earlier study, based on a much larger sample, found that AI adoption was associated with roughly nine minutes less work per day.
The numbers point in opposite directions
The contrast matters because it puts a crack in the neat claim that AI automatically produces productivity gains.
The latest ActivTrak research used behavioral data collected from April 1 to September 25, 2026. Heavy users recorded longer days than non-users, while their focus sessions averaged 37.8 minutes rather than the 48.2 minutes recorded for light users. They also switched applications 15% more frequently.
That does not prove AI caused the longer days or the interruptions. The data shows association, not clean causation.
The previous research had a broader base: 443 million hours of work across 163,638 employees and 1,111 organizations, collected from January 2023 through December 2025. It reported that average workdays fell by roughly nine minutes, or 2%, with AI use. It also found that daily productive time increased by 19 minutes, even as focus efficiency declined.
The studies differed in duration, sample size and measurement context, which may partly explain the discrepancy. But the contradiction gives leaders something useful: AI productivity claims need better boundaries, definitions and controls. Measuring productivity is less like counting units produced than judging whether attention became more valuable.
Why focus time deserves a management decision
ActivTrak’s data suggests that frequent AI use can interrupt sustained attention. Sarah Altemus, manager of the ActivTrak Productivity Lab, argued that repeated interruptions add friction to work requiring concentration, critical thinking and sustained attention.
The app-switching figures sharpen that point. Employees using AI embedded in workplace applications switched apps 0.6 times per hour. Those using web-based AI tools switched between 1.7 and 4.2 times per hour. ActivTrak calls the additional movement between applications a “swivel-chair tax.”
Light users also spent more active work time in core, job-specific applications: 81.4%, compared with 78% for heavier users. In other words, more frequent AI engagement does not necessarily translate into more focused work.
There is another wrinkle: AI may work in the background, through an API or agent, without producing a visible desktop session. An employee may use AI on an unmanaged device and still appear to be a non-user. Screen activity, paid hours, meeting time and completed work are not interchangeable measures.
Sanchit Vir Gogia, chief analyst at Greyhound Research, noted the obvious conflict: ActivTrak sells the platform that produces these measurements. That does not make the findings useless. It does mean leaders should treat behavioral data as evidence rather than an automatic verdict.
AI productivity measurement needs a capacity policy
Altemus raised the management issue: when tasks finish faster, managers may assume people can absorb more responsibilities. Without explicit workload boundaries, saved time can disappear into additional requests.
That is where AI productivity measurement becomes a leadership practice rather than a reporting exercise. The key question is not simply whether a task took less time. Leaders also need to know what employees did with the released capacity: whether it reduced low-value work, created room for deeper work or enlarged the backlog.
The embedded-versus-web comparison offers a practical clue. Embedded-AI users switched applications less often, while web-based AI users recorded substantially more switching. That suggests tool placement may affect workflow friction. It does not prove that every embedded tool performs better, nor does it make app switching a complete productivity score. It does mean that tool selection should consider how work flows, not just model capability or purchase price.
There is also a fairness issue. If behavioral monitoring shows that one group appears less productive because work happens off-screen or through agents, managers risk treating visibility as performance. Instrumentation has limits, and leaders should distinguish measurable activity from actual contribution.
Our analysis: the benefit may be capacity, not a shorter day
The most useful finding is not that AI users work longer. It is that organizations must decide what longer or shorter working time means.
Heavy users appear to carry more of the fragmentation visible in the latest dataset: shorter focus sessions, more app switching and longer recorded days. That does not automatically make them less productive, but it does make a simple hours-based comparison unreliable. A shorter workday could conceal work that never reaches the timesheet, while a longer one could include added checking, coordination or demand.
The same problem applies to the earlier report. It found nine-minute shorter workdays, a 19-minute increase in daily productive time and longer productive sessions, alongside falling focus efficiency and a three-year low of 60%. Those results point in different directions within the same study as well.
Organizations that treat productivity only as “hours returned” risk encouraging the wrong behavior. The better frame is capacity released for consequential work. If the organization cannot define that destination, efficiency will likely be absorbed by demand.
Decision checklist
Before treating AI time savings as a productivity gain, leaders should be able to answer:
- What improved?** Identify a defined task or outcome rather than relying on AI adoption or screen activity alone.
- Where did the capacity go?** Decide whether released time should remove low-value work, support judgment-heavy work or absorb additional responsibilities.
- What was the cognitive cost?** Review focus patterns, interruptions and the effort required to verify or reconstruct work.
- How is invisible work counted?** Account for activity outside managed devices, through APIs, background agents and other systems.
- Who benefits?** Test whether the gain improves the employee’s day or simply moves more work into it.
- Does the measurement fit the context?** Compare behavioral signals with output quality, workload and the employees’ experience of the change.
What to watch next
Watch whether organizations publish outcomes beyond adoption. Token use, active users and seat growth do not show whether work improved or teams made better decisions.
Also watch for clearer capacity policies. Teams need to say what happens when automation saves time: fewer low-value tasks, more deliberate work or simply more work. Without that choice, saved time becomes an invisible transfer from employees to the backlog.
The research does not prove that AI makes people more productive. It shows that leaders need better measurement and firmer boundaries. Treat AI as capacity to be allocated, not time saved to be harvested.
Key takeaways
- Measure delivery quality and cognitive load alongside hours worked.
- Set explicit rules for what teams should do with time released by AI.
- Compare tools by workflow friction and context switching, not only model quality.
- Treat behavioral tracking as partial evidence rather than a complete productivity score.
Frequently asked questions
Does ActivTrak’s research show that AI makes employees work longer?
It shows that heavy AI users worked longer in the latest dataset. It does not establish that AI caused the extra time or that those employees were less productive.
Why did ActivTrak’s two studies reach different conclusions?
The studies used different sample sizes, time periods and datasets. Those differences may partly explain the contradiction, while productivity remains difficult to measure.
What should a project manager track when introducing AI?
Track a defined workflow’s delivery time, output quality, rework, focus patterns and resulting workload. Avoid using app activity alone as the success measure.
Does less app switching make embedded AI better than web-based AI?
The latest data associates web-based AI with more frequent switching. It does not prove that every embedded tool improves performance or that switching is the only relevant factor.