ActivTrak's 120,620-worker telemetry says peak AI adoption is not peak work health

ActivTrak's Productivity Lab followed the same 120,620 employees across 1,009 organizations for three consecutive quarters, from Q4 2025 through Q2 2026, and reported that work-health and productivity measures climb as people move from little or no AI use to regular task-level use, then give the gain back once AI is embedded in the workflow ^[1][2][3]. Writing in Fortune, ActivTrak's chief executive says healthy utilization peaks at 75% at task-level adoption and falls about five percentage points at the deepest maturity stage, to a level the Lab describes as statistically indistinguishable from employees who barely use AI ^[4][3][5].

The distribution matters more than the curve. The Lab sorts behavioral data into three stages of maturity ^[21]: 27% of tracked employees used AI like a search engine to answer questions and summarize ^[6], 14% used it to draft content, generate ideas and complete routine tasks they then validated ^[7], and 2% reached workflow integration ^[8]. Users total 43% of the population studied ^[9], which leaves 57% outside the three stages entirely ^[10], and the stage counts add up exactly to the reported total ^[11]. So the cohort sitting at the far end of the maturity ladder is about one in fifty employees overall and roughly one in twenty-two AI users ^[12], on the order of 2,400 people at this study's headcount ^[13]. Their healthy utilization lands near 70% ^[14]. The stage the CEO argues most organizations should target is the middle one, illustrated by a sales rep generating a quote from five systems with a single prompt instead of compiling it by hand ^[23][16].

The cost anecdote is the most concrete thing in the column. ActivTrak's own operations team watched Anthropic spend rise, traced it to employees routinely using the newest and most powerful model to rewrite customer emails, and built an internal reference to match model to task ^[17]. That is a procurement failure, not a capability story, and it follows from how maturity is priced: deeper usage means more powerful models, more tokens, more infrastructure ^[18]. The second named risk is operational disconnect, where individuals assemble sophisticated workflows that optimize their own tasks without improving the broader process, producing what the CEO calls more AI slop, faster ^[19].

Note who is selling the remedy. The answer offered in both cases is more visibility into how AI changes work ^[20], and visibility built on behavioral data is what the Lab itself runs on ^[21]; the same column argues that license and login counts measure deployment rather than impact ^[15]. The critique of consumption metrics is fair. It also happens to describe the vendor's product.

Watch whether the five-point dip survives as Stage 3 grows past 2%. With roughly 2,400 people in that bucket ^[13], self-selection among the earliest deep adopters is at least as plausible as workflow integration causing the decline, and the published piece does not set out how healthy utilization is composed or report stage-level sample detail ^[24]. Watch whether token spend per head starts appearing next to seat counts in board packs, since model-to-task guidance was the fix ActivTrak reached for internally ^[17]. And watch for a replication outside a monitoring vendor's customer base: the sample averages about 120 tracked employees per organization ^[22], which is not the shape of the enterprise-wide rollout most AI strategies are written for.