AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Workers Switch Apps 900 Times A Day. Where Does AI Fit? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

A Harvard Business School preprint analyzed 103 million app events from 1,017 workers at eight Indian organizations. Workers averaged 908.3 app switches per day; AI use was linked to fewer switches in the next ten app uses but about 200 more switches on AI-use days. The observational data cannot show whether AI caused the changes.

A Harvard Business School preprint tracking 1,017 office workers found they switched applications an average of 908.3 times per day, while AI use was associated with two different patterns: fewer switches in the next ten app uses, but about 200 more switches across AI-use days. The findings matter to employers considering AI as a way to improve focus, but the study was observational and cannot establish that AI caused either pattern.

Researchers Sumer S. Vaid and Ashley V. Whillans analyzed 103 million application events recorded from November 2025 through January 2026. The sample covered employees at eight organizations in India across software, accounting, financial services, e-publishing, law and marketing. Each employee was tracked for a median of 58 workdays.

The paper reports 56.9 million app switches across 62,693 person-days. The median employee switched apps 814 times on a workday, and the middle half of employees ranged from 547 to 1,130 switches. A period using one app lasted an average of 27.8 seconds, though the median was 3.3 seconds. A switch meant moving to a different application, not simply opening another document in the same one.

The AI results vary by time scale. In the ten app uses following an AI use, workers used fewer apps and switched less, staying about half a second longer in each app on average. Across full days, however, workers used AI on days with roughly 200 more switches; days with more minutes of AI use were also more fragmented. These are associations in the recorded data, not results from an experiment.

At a glance
reportWhen: Preprint first posted July 7, 2026; rev…
The developmentA Harvard Business School preprint reports that workers averaged 908.3 daily app switches, while AI use had different associations over short intervals and across whole workdays.

App switching and AI · one moment

What happens around one AI query?

The study zoomed in on every AI use: the ten app uses before it, and the ten after. Up close, work calms down a little. Zoom out to the whole day, and the picture changes.

10 app uses before AI median 7 seconds 10 app uses after Whole AI day: about 200 more switches
  1. 1 · A workday of switchingWorkers changed apps 908 times a day on average. One app held their focus for 27.8 seconds on average, 3.3 seconds at the median.
  2. 2 · AI comes in at a busy momentPeople used AI a median of 13 times a day, for a median of 7 seconds. Fragmentation peaked in the hour of AI use.
  3. 3 · The next ten app usesFewer different apps, fewer switches, about half a second longer per app, more predictable. Real, but small.
  4. 4 · Zoom out: the whole dayOn days with AI use, the same people made about 200 more switches. Whether AI calms work or busy moments bring AI: unknown.

Source: Sumer S. Vaid, Ashley V. Whillans (Harvard Business School), “Digital Fragmentation and Generative AI Use Across 103 Million Application Events”, arXiv:2607.06681v2, preprint under review. Blocks are schematic, not to scale.

Graphic: Powered by Thorsten Meyer AI

AI’s Mixed Link to Workday Focus

The results caution against treating an AI assistant as an automatic focus tool. A short stretch of activity after an AI interaction may look more concentrated, while the broader day on which AI is used may involve more app switching. The paper does not show that the tool itself produced either outcome.

For managers, the distinction matters when evaluating workplace AI. A simple count of app changes cannot tell whether someone is distracted, coordinating work, or completing tasks more quickly. The study did not measure whether work quality, productivity or employee well-being improved. Its findings support asking when and how tools fit into a workday, rather than assuming fewer switches alone mean better work.

The authors also found that fragmentation varied substantially from day to day for the same worker. In their analysis, 44.6% of variation came from changes within individuals across days, compared with 35.8% from differences between people and 19.6% between organizations. That pattern suggests daily workloads and communication demands may be relevant, though the data cannot identify the causes.

Products Worth Considering

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, White Smoke

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, White Smoke

As an affiliate, we earn on qualifying purchases.

How the Harvard Study Tracked Apps

The study, titled Digital Fragmentation and Generative AI Use Across 103 Million Application Events, was written by Vaid of Harvard Business School’s Artificial Intelligence Institute and Whillans of its Negotiation, Organizations and Markets Unit. It was first posted on arXiv on July 7, 2026, then revised on August 7. The manuscript identifies itself as a preprint under review and warns that findings may change during peer review.

The records came from Flowace workforce-monitoring software, already used by the participating organizations. It logged when an employee changed the active application, along with information such as the app name and session times. Employers determined which apps counted as productive. Researchers received anonymized data under a data-use agreement, and the paper says the study protocol was approved by Harvard’s human-subjects committee.

The authors state that the work had no external funding and that Flowace provided no funding or commercial support and had no role in study design, analysis or interpretation. One author disclosed equity in a private company unrelated to Flowace that uses AI to create enterprise digital twins. The sample included five AI tools, with Anthropic’s Claude the least used among them.

“The manuscript is currently under review, and the findings may change based on the peer review process.”

— Sumer S. Vaid and Ashley V. Whillans, the paper’s authors

What App Logs Cannot Establish

The data cannot determine whether AI use led to more switching, whether workers turned to AI on days that were already fragmented, or whether both reflected another factor such as workload. The authors describe these explanations as possibilities; the study was not an experiment.

The sample was limited to eight organizations in India and did not cover phone use. The records show application activity, not whether workers completed tasks better, faster or with less strain. App-switch counts also do not reveal what the worker was doing or whether a switch interrupted concentrated work.

The paper remains a preprint under peer review, so its results and interpretation could change. The headline that app switching “drops after AI” describes only the short sequence after AI use; it does not capture the separate finding that AI-use days had more switches overall.

Peer Review and Workplace Evidence

The paper’s next formal milestone is peer review; the authors’ preprint notice says the findings could change through that process. No publication decision or further study timeline is specified in the source material.

Employers assessing AI and focus will need evidence that measures more than app activity, including task outcomes and worker experience. Until then, the study’s observations can help frame questions about daily schedules and communication demands, but they do not establish that buying or deploying an AI tool will reduce fragmentation.

Key Questions

Did the study find that AI reduced app switching?

Not across the whole day. The researchers found fewer switches in the ten app uses after an AI use, but about 200 more switches on days when workers used AI. The two findings describe different time windows.

How many app switches did workers make?

The average was 908.3 switches per day. The median was 814, meaning half the employees had a higher daily figure and half had a lower one.

Can the study show that AI caused more switching?

No. The data were observational, not experimental. The researchers say it is unclear whether AI changed the workday, whether workers used AI on already fragmented days, or whether both were true.

Who took part in the research?

The records covered 1,017 employees at eight organizations in India, spanning several sectors, including software, accounting, financial services, law and marketing. The study did not track phone use or workers in other countries.

Has the paper been peer-reviewed?

Not according to the paper’s stated status. It was posted on arXiv as a preprint under review, and its authors warn that findings may change during peer review.

Source: ThorstenMeyerAI.com

You May Also Like

Kill-Switch-Proof: How To Build So Washington Can’t Take Your AI Stack Down

Strategies for making AI stacks resistant to government shutdowns, including dependency mapping, abstraction layers, fallback tiers, and open-weight models.

Are ByteDance, Runway, And Luma Outperforming OpenAI In AI Video Development?

ByteDance, Runway, and Luma are reportedly dividing the AI video market, leaving OpenAI behind, though detailed data and measurements are not yet available.

Signal: Four Frontier-Class Open Models in Eight Weeks — China’s Release Cadence Is the Story

Chinese AI labs released four frontier-class open models from late April to mid-June 2026, signaling a fast-paced production line that challenges Western dominance.

Regex Chess: A 2-ply minimax chess engine in 84,688 regular expressions

A developer created a functional chess engine using 84,688 regular expressions, demonstrating an unconventional approach to programming chess logic.