Some workdays feel like ten tabs opening inside your head. The next day, with the same job and the same person, the noise settles.
A revised preprint suggests that swing deserves more attention than a fixed label like “focused worker” or “distracted worker.”
What changed
The authors analyzed about 103 million foreground-application events from 1,017 employees at eight India-based organizations, largely in knowledge-work fields. Their Fragmentation Index measured the breadth, brevity, and unpredictability of application switching.
Day-to-day variation within the same employee accounted for 44.6% of the Index variance. Stable differences between employees accounted for 35.8%. Differences between organizations accounted for 19.6%.
That result puts the workday ahead of the worker or organization as a source of variation. A person’s own changing baseline may tell you more than a single snapshot or a company average.
Why it matters
This day-level view changes how the AI finding should be read. AI use showed up more often on days when employees were more fragmented than usual. Across 880,168 episodes, the ten application uses after an AI event involved fewer unique apps, less switching, longer dwell times, and more predictable movement than the ten uses before it.
But the trace cannot explain why. Work may already have been converging when someone opened an AI tool. A finished meeting, a clarified decision, or another condition could help explain both events. The study does not establish that AI improved attention, task quality, wellbeing, or productivity.
Because this is a study of software behavior, its measurement boundaries matter. Foreground apps are limited proxies for human attention. Employer-supplied productivity labels carry their own limits, and the detection rules covered five named AI tools. The revised preprint remains under review and has not been independently reproduced for this card. Its granular records are unavailable publicly because of data-use agreements and reidentification risk.
Measurement can become surveillance.
So these results cannot support invasive worker monitoring. Any follow-up needs consent, proportionality, privacy safeguards, and a narrow purpose that workers can understand.
Watch next
Independent and cross-regional studies could test whether the pattern survives outside this India-based knowledge-work sample. Better studies would pair consented application traces with a specific task outcome or a worker’s own report of effort and focus.
A bounded audit can start smaller. Pick one recurring workflow. Compare each person with their own baseline on fragmented and consolidated days. Inspect a short window before and after AI use, then pair that trace with the actual task result.
The useful question is whether a tool changes the shape of work in a way people can feel and verify. The current paper gives that question a measurement surface. It does not supply the verdict.