Nearly nine in ten workers now use AI on the job, and most believe it makes them faster. But a new cross-university study finds that the hours employees spend correcting, supervising, and cleaning up after that AI, a task researchers have started calling “botsitting”, are quietly canceling out much of the productivity gain leaders were promised, and almost no organization is tracking it.
The Productivity Promise Meets a New Kind of Overhead
The Work AI Index, fielded by the Work AI Institute with researchers from Emory University, the University of Notre Dame, UC Santa Barbara, UC Berkeley, University College London, UNC Charlotte, and Stanford, surveyed 6,000 full time digital workers in the United States, United Kingdom, and Australia between December 2025 and January 2026. Eighty seven percent said they use AI at work, and 75% said it makes them more productive, with workers estimating it saves them roughly 11 hours a week through automation. Workers now expect 35% of their output to be automated within a year, a 30% jump from where they place automation today, and 57% say they want AI to automate even more than that.
Set against those numbers, one figure stands out: only 13% of workers say their organization is performing significantly better as a result. The gap between individual time savings and organizational performance is the productivity paradox the researchers set out to explain, and their answer is that a large share of the “saved” time is being spent somewhere leaders are not measuring.
What Botsitting Actually Looks Like
Researchers define botsitting as the work required to make AI usable: feeding it missing context, checking its outputs, debugging its mistakes, rerunning prompts, and cleaning up the confident but wrong answers it leaves behind. Across the surveyed workforce, that adds up to 6.4 hours a week per worker, split roughly between feeding context (2.3 hours), supervising outputs (2.2 hours), and debugging errors (1.7 hours). Of all the time workers spend interacting with AI, 37% goes to botsitting, versus 36% to actual production work and 27% to learning tools or building agents of their own.
Tool sprawl makes the problem worse. Seventy seven percent of workers juggle multiple AI tools in a given week, and a third use four or more. Heavy, multi tool users are the ones logging the most botsitting hours, which suggests that adding tools without consolidating workflows is adding supervision overhead faster than it is adding output.
The “Botshitting” Feedback Loop
A related behavior the researchers term “botshitting” compounds the cost. Sixty nine percent of AI users admit to it: passing along AI generated work they have not verified. Forty one percent say they have delivered work they could not explain if asked, 38% have used AI tools their employer has not approved, and 28% have blamed the AI when their own work was wrong. Heavy AI users are 64% more likely to botshit than light users, and workers who juggle multiple tools are 35% more likely to do it than those who stick to one. The pattern is self reinforcing: more tools produce more botsitting, more botsitting fatigue produces more unverified output, and more unverified output produces more rework, which shows up as still more botsitting the next week.
What It Means for the HR Leader
Six and a half hours a week per employee is not a rounding error once it is multiplied across a workforce; it is a labor cost hiding inside every “AI adoption” dashboard that only counts licenses purchased or prompts sent. Most performance management and workforce planning systems still have no field for it. That blind spot compounds a problem HR and finance teams are already confronting elsewhere: the AI spending built into next year’s budgets is outrunning the productivity returns finance can actually document, and uncounted botsitting hours are one reason the return keeps failing to show up in the numbers.
The wellbeing data raises the stakes further. Frequent botsitters are 73% more likely to be actively job hunting than infrequent ones, and workers who admit to botshitting are 3.8 times more likely to be looking for a new employer. Heavy AI users are also 3.4 times more likely to blame the tool when their own work fails, a dynamic that erodes both morale and accountability at the same time. Treated only as a productivity story, this looks like a rollout problem. Treated as an attrition and quality story, it looks like a design problem in how AI work is being assigned, measured, and reviewed.
What HR and People Leaders Should Do Now
First, measure time on task rather than task completion alone; a manager who only tracks whether AI-assisted work got done cannot see the hours it took to make that output trustworthy. Second, build verification time into role design and staffing plans explicitly, rather than treating “review the AI’s work” as free capacity that appears from nowhere. Third, train managers to treat botshitting, passing along unverified AI output, as a policy violation rather than a shortcut, since the data ties it directly to rework and to attrition risk. Fourth, consolidate the sanctioned AI toolset instead of letting adoption sprawl across tools, since multi tool users carry the heaviest botsitting and botshitting burden. Finally, add botsitting and botshitting indicators to the same dashboards that already track AI adoption and license spend, so a rising number reads as an early warning on burnout and turnover, not just a footnote to a productivity story that is not fully panning out.
Source: Glean Work AI Institute
