Botsitting Statistics: How Much Time Workers Spend Supervising AI (2026 Data)

Half of American workers who use AI at work say they’d get just as much done without it.

That isn’t a complaint from skeptics. It’s from the people who use AI regularly. In a new Founder Reports Botsitting Report, a survey of 2,052 US employed adults who use AI at work at least weekly, 53.2% said that if their AI tools disappeared tomorrow, they’d get the same amount of work done or more. Only 13.6% said they’d get much less done.

The gap between what workplace AI is supposed to deliver and what it actually delivers is known as “botsitting” (a name given by researchers at Glean’s Work AI Institute). Botsitting is the labor of feeding AI the context it needs, supervising what it produces, debugging its mistakes, and cleaning up after it. In their 2026 Work AI Index, workers spent 6.4 hours a week botsitting, accounting for about 37% of all time spent on AI-related work.

In the Founder Reports data, 82.5% of regular AI users spend some amount of time on this oversight work. This article compiles the strongest supporting primary research from Workday, Microsoft, KPMG, BetterUp/Stanford, METR, Stanford RegLab, and the University of Chicago, in addition to Founder Reports.

👉 Download the full Bottsing Report.

Half Would Get as Much Done Without AI

In the Founder Reports botsitting survey, 53.2% of weekly AI users said they’d get the same amount of work done or more without their AI tools. Only 13.6% said they’d get much less done. 8.9% said they’d actually get more done without AI.

The pattern varies by usage.

AI usage	Would get the same or more done without AI
Once a week	67.2%
2–6 times a week	55.3%
Every day	47.3%
Multiple times a day	37.6%

Even among workers using AI multiple times a day, nearly 4 in 10 say they’d get the same amount done or more without it. For 24.9% of those multi-daily users, AI is clearly essential (they’d get much less done without it). For the rest, the picture is murkier.

If half of a tool’s own users say they could do their jobs just as well without it, the standard story about AI productivity is incomplete without covering both sides.

How Much Time Botsitting Actually Takes

The reason so many workers feel AI isn’t delivering is visible in the hours.

In the Founder Reports survey, 82.5% of weekly AI users spend some amount of time overseeing, checking, or fixing AI’s work in a typical week. 31.6% spend 3 or more hours. 15.8% spend 6 or more hours.

Hours per week on botsitting	Share of weekly AI users
None	17.5%
Less than 1	27.4%
1–2	23.5%
3–5	15.7%
6–10	7.4%
10+	8.4%

As you would likely expect, the number of hours scales sharply with how often people use AI. Among workers who use AI multiple times a day, 52.7% spend 3 or more hours a week on oversight, and 32.1% spend 6 or more. Among weekly users, those numbers drop to 17.6% and 4.4%. More than one in three heavy AI users spends what amounts to most of a workday each week overseeing, checking, or fixing AI output.

Here is how our survey data compares to three independent teams using different methods:

Source	Finding	Method	Sample
Founder Reports	31.6% spend 3+ hrs/week; 15.8% spend 6+ hrs	Survey of weekly AI users	2,052 workers
Glean	6.4 hrs/week average botsitting	Survey + platform telemetry	6,000 workers
Workday	~40% of AI time savings lost to rework	Survey	3,200 workers
METR	~9% of working time reviewing/cleaning AI output	Screen recordings	16 devs, 246 tasks

Add one more number, from a University of Chicago paper. In Large Language Models, Small Labor Market Effects (NBER WP 33777), Humlum and Vestergaard asked workers to describe, in free text, the new tasks AI had created for them, then categorized the responses. Only 41% described productive AI use. The other 59% described “AI implementation and oversight.” Two survey rounds of roughly 25,000 workers each, linked to Danish administrative earnings records.

If most of what AI adds is oversight of AI, workers who notice they could do their jobs without it aren’t being contrarian.

The Uncounted Hours

Botsitting requires time, but it isn’t always recognized by employers.

Our botsitting survey asked the 1,693 workers who spend any time managing AI whether their employer factors that time into their workload. The answers:

Does your employer factor AI oversight time into your workload?	Share
Yes	38.2%
Partly	20.0%
No	34.5%
Not sure	7.3%

61.8% said their employer doesn’t fully count the time, or they aren’t sure. For 34.5%, oversight is work piled on top of regular responsibilities. It is unmeasured, unbudgeted, and almost always invisible in productivity reporting.

Recognition rises for workers who spend more hours with AI. Only 24.4% of workers spending less than an hour a week on AI say that time is counted, versus 69.4% among those spending more than 10 hours. Workers doing light, routine AI oversight are more than twice as likely to say the time goes uncounted as workers doing it heavily. Heavy AI work is visible enough to end up in a job description. Light daily oversight rarely is.

A structural fact from Workday’s 2026 research explains why: in 89% of organizations, fewer than half of roles have been updated to reflect AI capabilities. Workers are, as Workday puts it, “using 2025 tools inside 2015 job structures.” If the job description hasn’t changed, the oversight work it added doesn’t exist on paper.

Managing AI Takes as Much Effort as Training a New Hire

The clearest challenge in our survey data for the “AI replaces entry-level work” narrative is this: workers who have trained new employees and now also manage AI say the two take comparable effort.

Among the 90.4% of respondents (n=1,856) who had trained an entry-level employee, Founder Reports asked how the effort of getting good work from AI compares to the effort of training a hire.

Effort of training AI compared to training a new hire	Share
Much more	9.9%
A little more	23.2%
About the same	26.9%
A little less	23.3%
Much less	16.7%

60.0% said getting good work from AI takes at least as much effort as training an entry-level employee. 33.1% said it takes more.

The finding is remarkably consistent across groups. Among individual contributors, 62.3% said training AI takes as much or more time. Among managers, 58.4%. Gen Z (61.0%), Millennials (60.8%), and Gen X (58.4%) are within three points of each other.

The number has a direct implication. Companies cutting entry-level hiring on the assumption that AI replaces junior work are substituting one supervision task for another. Based on worker reports, the two take comparable effort. But AI, unlike a junior hire, never becomes a senior.

That matters for how organizations plan around agents too. In KPMG’s Q1 2026 AI Quarterly Pulse, 64% of leaders said agents had already changed entry-level hiring, and 71% said the same for experienced hires. If training AI carries the same as training a hire, the headcount math on that transition deserves scrutiny.

The Dependence Divide

The most eye-opening pattern in our survey data is how sharply attitudes toward AI split by how much time workers put into it. The more they use AI, the better they rate it, and the more dependent they feel on it.

Asked what they would do if they managed an employee who performed like their main AI tool:

• 17.1% would promote it.
• 66.4% would keep it as is.
• 12.5% would warn it.
• 4.0% would fire it.

For every worker who sees their AI as high-performing enough to promote, another sees it as underperforming enough to need corrective action.

But those tails are populated by two very different groups:

AI usage	Would promote	Would warn or fire	Would get much less done without AI
Once a week	8.0%	22.2%	6.9%
2–6 times a week	14.9%	16.8%	9.9%
Every day	20.6%	13.2%	16.4%
Multiple times a day	28.5%	12.0%	24.9%

The same split shows up when sorted by hours of oversight. Among workers who spend more than 10 hours a week on AI oversight, 32.9% would promote their AI, and 41.0% say they’d get much less done without it. This is the group that has committed real time to making AI work, and most are convinced the time was well spent.

Looking directly at the groups makes it clearer. Among the 350 respondents who said they’d promote their AI, 59.7% use it daily or more, 42.9% spend 3 or more hours a week on oversight, and 32.3% say they’d get much less done without it. Among the 339 respondents who’d warn or fire theirs, 36.9% only use it once a week, 29.2% spend 3 or more hours on it, and just 5.0% say they’d get much less done without it.

The data shows that AI users are divided into two populations. One has built workflows around AI and defends it strongly. The other uses it occasionally, finds it adds roughly as much work as it saves, and would like it to perform better.

Who Carries the Load

Botsitting is not evenly distributed. Three breakdowns in our survey data stand out: managers versus individual contributors, women versus men, and sector.

Managers Carry the Load

Managers and senior leaders spend substantially more time managing AI than individual contributors do. They’re also more invested in it, lean on it more heavily, and rate it more generously.

Measure	Managers and above (n=884)	Individual contributors (n=1,014)
Spends 3+ hours a week on oversight	40.7%	25.0%
Would get the same or more done without AI	47.3%	57.5%
Would promote AI if it were an employee	20.1%	14.2%
Would warn or fire AI if it were an employee	16.0%	16.8%
Says employer accounts for time overseeing AI	40.7%	35.8%

Discipline rates are nearly identical across the two groups. Managers also get slightly more credit for the time (five points more likely to say their employer counts it). For individual contributors spending any time on AI oversight, more say the work goes uncounted than counted.

This parallels the findings from our 2026 AI in the Workplace Report that managers and leaders are significantly more likely than individual contributors to have to spend time cleaning up a co-workers low-quality AI output.

Women Report Less Dependence Than Men

Women and men differ meaningfully in how dependent they feel on AI, but not in how they rate it. 59.0% of women say they’d get as much done without AI, versus 47.4% of men (an 11.6-point gap). The gap holds within usage levels: among workers using AI multiple times a day, 44.3% of women said they’d get as much done without it, compared with 31.8% of men. Promotion and discipline rates are nearly identical. Women aren’t thinking worse of their AI, but they do report less dependence on it.

Sector Differences

How workers experience AI varies meaningfully by employer type.

Measure	For-profit (n=1,584)	Government (n=275)	Non-profit (n=193)
Spends 3+ hrs/week on AI oversight	35.0%	20.4%	19.7%
Would get the same or more done without AI	51.3%	54.2%	66.8%
Would promote AI	18.7%	14.5%	7.3%
Would warn or fire	16.0%	14.9%	22.8%
Says employer doesn’t count AI time	30.5%	46.4%	53.6%

Nonprofit workers would discipline their AI more than three times as often as they’d promote it (22.8% vs. 7.3%). At for-profit companies, those numbers flip (18.7% promote vs. 16.0% discipline). The recognition gap is substantial, too. More than half of nonprofit workers who spend any time managing AI say their employer doesn’t count it, versus 30.5% of for-profit workers.

Why the Oversight Work Is Necessary

Professional-grade tools still fail at rates that require review. In Hallucination‐Free? Assessing the Reliability of Leading AI Legal Research Tools, published in the Journal of Empirical Legal Studies (vol. 22, 2025), Stanford RegLab researchers ran 202 preregistered legal queries against the major commercial tools. Lexis+ AI hallucinated more than 17% of the time and was accurate on 65% of queries. Westlaw AI-Assisted Research hallucinated on roughly one-third of its responses and was accurate on 41–42%. Ask Practical Law AI declined to answer 62% of queries and, when it did answer, was accurate on 19%. General-purpose LLMs hallucinate on legal queries between 58% and 82% of the time.

Unsupervised use degrades expert judgment. In the 2023 Jagged Technological Frontier field experiment with 758 BCG consultants, on a task deliberately placed outside AI’s capability frontier, 84.5% of the control group got the correct answer, versus 60% and 70% of the two AI-assisted groups. Buried in Appendix C: consultants’ “retainment” scores clustered near 0.87, which the authors say meant subjects were “essentially copying and pasting ChatGPT responses.” Consultants given prompt-engineering training retained more raw output, not less. The experiment ran on GPT-4 as it existed in 2023, so cite it for the mechanism rather than as a current error rate.

And the oversight is often skipped anyway. In the 2025 Melbourne/KPMG global AI study (48,340 respondents across 47 countries), two in three employees rely on AI output without evaluating it, only 55% agree they can evaluate its accuracy, and 61% have had no AI training of any kind. BetterUp Labs and Stanford’s Social Media Lab found 40% of US desk workers received AI “workslop” in the past month, costing about 2 hours per incident and roughly $9 million a year for a 10,000-person company.

Botsitting Is Becoming the Job Description

The oversight work isn’t a phase. It’s being written into job descriptions and hiring plans.

In Microsoft’s 2026 Work Trend Index, 20,000 full-time knowledge workers were asked which skills become more important as AI takes on more work. They ranked quality control of AI output first (50%), ahead of critical thinking (46%). The same report notes that 86% of AI users treat AI output as a starting point, not a final answer. Supervision is already default behavior.

Leaders are planning around it. In KPMG’s Q1 2026 pulse, 57% of leaders expected humans to primarily manage and direct AI agents within two to three years. Adaptability and continuous learning (83%) outweighed technical or programming abilities (71%) among the skills they wanted in entry-level employees working with agents.

The agent layer is where botsitting is about to scale sharply. 57% of organizations run a “human-in-the-loop” model where a human validates outputs but does not oversee each individual agentic action or decision. The gap between validating what came out and reviewing what the agent did to get there is where the next wave of unmeasured work lives.

The returning question is the one the Founder Reports Botsitting Report set out to answer: do the hours going in produce the work expected to come out? For slightly more than half of US workers who use AI weekly, the current answer is “no.” 53.2% say they’d get the same amount of work done or more without their AI tools.

👉 Download the full Bottsing Report.