ActivTrak spent three years watching how 163,638 employees actually used their time. Not surveying them. Watching the behavioral data itself, across more than 443 million hours.
One finding sits above the rest. Employees who spend seven to ten percent of their work hours inside AI tools show the highest productivity of any usage band measured, ninety-five percent, ahead of every lighter and heavier user group in the study.
Only three percent of employees are actually in that band.
Fifty-seven percent, the largest single group, spend less than one percent of their hours in AI at all.
Read those two numbers together. The optimal zone exists, confirmed in the data. Almost nobody is standing in it.
Why This Isn't An Adoption Problem
Most AI conversations through 2026 still treat adoption as the finish line. Get people using the tool, and value follows. That framing made sense when AI use was rare. It stopped making sense the moment ActivTrak's data showed usage intensity matters as much as usage itself, and the two are not the same question.
Fifty-seven percent of employees barely touch AI. That group has an adoption problem, and the fix is straightforward: more exposure, more training, more reason to open the tool at all.
But a separate, smaller group is using AI heavily, well past the seven-to-ten percent band, and still not hitting the ninety-five percent productivity mark the sweet spot produces. That group doesn't have an adoption problem. It has a calibration problem, and no amount of additional training fixes overuse, because overuse was never about unfamiliarity with the tool.
Two different problems, requiring two different fixes, and almost every organization is currently running one program designed to solve only the first one.
What The Before-And-After Data Actually Shows
ActivTrak also tracked a smaller group of 10,584 employees across 376 companies for 180 days before and after they adopted AI. Every single measured category of work increased. Email rose 104 percent. Chat and messaging rose 145 percent. Time in business management tools rose 94 percent. Nothing went down.
That's consistent with the sweet spot finding, not separate from it. AI didn't just get adopted. It got adopted unevenly, with no organization-level signal telling anyone what a healthy dose actually looked like. Absent that signal, usage drifted in both directions at once: a majority barely engaging, a smaller group over-engaging, and a narrow band in the middle quietly outperforming everyone on either side of it without anyone noticing why.
The report's own language doesn't hedge on this. The data is unambiguous. AI does not reduce workloads. It also, evidently, does not automatically produce its best output just because usage is high. The relationship between AI time and productivity isn't a straight line. It's a curve, and most organizations are managing toward the wrong end of it without knowing the curve exists.
There's a second finding in the same dataset that explains why this drift goes unnoticed for so long. The same three years of data show disengagement risk rising twenty-three percent, even as burnout risk fell to five percent. Those two trends usually get treated as good news and bad news separately. Read together against the sweet spot finding, they describe the same underlying failure: organizations got better at preventing overload, and no better at redirecting the capacity that freed up. A workforce with unused capacity and no clear signal for where to point it doesn't sit idle. It drifts into whatever's nearest, which is exactly how a majority ends up under one percent AI usage while a smaller group drifts well past the point where more usage helps. Average focus session length fell nine percent over the same period, with focus efficiency overall dropping to a three-year low, evidence the same drift is showing up in attention as much as in AI usage specifically.
Who Should Actually Own This Number
Here's the part missing from almost every AI rollout. Nobody has been assigned to find, for their specific organization, where that seven-to-ten percent band actually sits, and then to notice who's underneath it and who's blown past it.
This isn't a training question, and it isn't an IT question. It's an ownership question, and for most companies the answer is currently nobody. Deloitte's own research on enterprise AI backs this up from a different angle: only a third of organizations report using AI to genuinely transform how work gets done, while thirty-seven percent describe their own usage as surface-level, with little real change underneath the adoption number. A high adoption rate and a well-calibrated one are not the same claim, and most companies are measuring only the first, if they are measuring anything about calibration at all.
The fix costs less than most AI initiatives already funded in 2026. Someone needs to own a single, specific number: what fraction of work hours, for each role or team, actually produces the best output, and how far is the real distribution from that band right now. That person doesn't need a new platform. The behavioral data already exists inside whatever tools track how people spend their day. What's missing is the assignment, not the evidence.
Consider what this ownership actually looks like in practice, because it's more concrete than another training rollout. It means pulling the usage distribution for a team, finding out how many people sit under the one percent floor and how many sit well past the ten percent ceiling, and treating both groups as needing a different intervention rather than the same generic push toward "more AI adoption." The under-users need a reason and a nudge. The over-users need a conversation about whether the extra hours in the tool are producing anything the data can actually show, or just accumulating as habit. Neither conversation happens if the only number anyone tracks is the adoption rate itself.
Most organizations will spend the coming year debating whether their AI adoption rate is high enough. The more useful question was never about the rate. It was about where the seven-to-ten percent band sits for their own workforce, and whether anyone has been given the job of finding out.
References
ActivTrak Productivity Lab. "2026 State of the Workplace: AI Adoption & Workforce Performance Benchmarks." activtrak.com, published March 2026, updated May 28, 2026.
Deloitte. "State of AI in the Enterprise 2026," as cited within the ActivTrak report above.
