Work-tracking data shows AI use and work changes rising together — but does that prove AI caused the shift?

This asks whether workplace monitoring data, such as app-usage logs, hours and focus time, can prove that AI caused changes in how people work, or only that the two happened at the same time. On its own, the corpus suggests the answer is no. It also suggests that monitoring data becomes much more convincing once other kinds of evidence explain the same pattern. The clearest example is ActivTrak's trace data. As AI adoption rose across monitored organizations, people spent more time in work apps and worked more weekend hours, and daily focus time fell to a three-year low Does AI adoption actually reduce the work that employees do?. The note's own title says 'accompanied by,' and that wording matters. Logs record that two trends moved together. They can't rule out other explanations, such as layoffs leaving fewer people to do the same work, new reporting demands, or a general speed-up that would have happened anyway.

What moves the story from correlation toward explanation is a different kind of evidence. A Berkeley Haas ethnography watched people work instead of logging their clicks, and it named three mechanisms. AI made more tasks feel doable, so people took more on. It removed natural stopping points. And it let people run several threads at once Does generative AI actually save workers time or intensify it?. Those mechanisms predict exactly the pattern ActivTrak found: more activity, less focus, and work spilling into breaks. Neither source settles the question alone, but a log pattern combined with an observed mechanism makes a stronger case than either one separately. This is how causal claims about work usually get built, by combining sources rather than reading off a dashboard.

There's a less obvious problem. Monitoring may be measuring the wrong thing. Research on how AI reshapes tasks finds that it doesn't so much cut time as move it, away from doing the work and toward writing prompts and checking outputs Does AI really save time, or just change how we spend it?. A tracker that counts 'time in work apps' can't tell drafting a report apart from fixing what an AI drafted. Zapier's survey makes this concrete: 92% of enterprise AI users report productivity gains, yet the average worker spends about 4.5 hours a week cleaning up AI output, and the heaviest users spend the most How much time do workers really spend fixing AI mistakes?. Self-reports and activity logs can disagree with each other. Each one captures a different part of what's happening.

Averages also hide a lot of variation. AI effects depend on whether people are applying skills they already have or learning new ones When does AI actually boost worker productivity?. Real delegation to AI is concentrated in information-heavy jobs Where have workers actually delegated tasks to AI?. Even firms with similar exposure substitute AI for labor at different rates Do firms substitute labor for AI at different rates?. A single organization-wide trend line in monitoring data can blend very different stories into one. The hiring 'doom loop' note is a useful warning here. Its survey data supports every step of an AI arms race between applicants and recruiters, but the note explicitly says the data can't establish which direction the causation runs Are job applicants and employers locked in an escalating AI arms race?.

The takeaway is this. Monitoring data is good at showing that work changed. It is weak at showing why, and it can count AI-related labor such as prompting, checking and cleanup as if it were the original work. To get from a dashboard to a causal claim, you need the mechanism, for example from ethnography, and you need to break the averages down by task type and skill level.