The draft arrives fast. The checking takes the afternoon. That is the real shape of AI work in most offices now. Nobody planned the checking. Nobody assigned it. It just lands on the most senior person near the task. The held detail is how that checking load compares to a cost every owner knows. It compares to training a new hire.
The direct answer: in an October 2026 survey of 2,052 workers who use AI at work, 60% of those who have trained an employee said checking AI takes at least as much effort as training one. The work has a name. It is called botsitting. The fix is not less AI. The fix is a named checker on every task.
- What the survey counted
- Why managers carry the load
- The unnamed job in your office
- One rule to set this week
What the survey counted
The survey covered US workers who use AI at least once a week. Among the 1,856 who have trained an employee, six in ten put AI oversight at or above the effort of training a person. One in three said it takes more effort. These are US figures. Treat them as a shape, not a Canadian count.
The hours are concrete. 31.6% of all respondents spend at least three hours a week training, overseeing, checking and fixing AI work. 8.4% spend more than ten hours a week. That second group gives the tool a full working day, every week, just to keep its output safe.
The checking load, from the October 2026 survey of 2,052 workers
| Measure | Share |
|---|---|
| Say checking AI takes at least as much effort as training a new hire | 60% |
| Spend at least three hours a week checking and fixing AI work | 31.6% |
| Spend more than ten hours a week on that checking | 8.4% |
| Managers spending at least three hours a week on AI oversight | 40.7% |
| Individual contributors spending at least three hours a week | 25% |
The work itself is unglamorous. It is reading a draft for the claim that sounds right and is not. It is fixing
the tone before a client sees it. It is re-running the prompt because the first answer missed the point. None of it appears in a job description. All of it appears in a senior person’s week.
Why managers carry the load
The load does not spread evenly. 40.7% of managers and above spend at least three hours a week overseeing AI and fixing errors. Only 25% of individual contributors do. Seniority attracts the checking. The person who can spot the error becomes the person who must.
There is a second bill. 57% of managers and above have had to fix a colleague’s work that leaned too hard on AI. So the manager checks the tool’s output. Then the manager checks the human output built on the tool. The tool saved drafting time. It created two review rounds in exchange.
Vendor numbers rarely subtract this time. Beacon’s 9.6 AI hours only count after you subtract review time makes that point on a vendor claim. Run the same subtraction on your own office. The saved hours shrink. They do not vanish. Honest numbers let you staff the work properly.
The unnamed job in your office
Your office has this job already. It just has no name, no owner and no hours in the plan. The work drifts to whoever cares most about quality. That person burns out first. Their own work slips while they guard everyone else’s.
Tools make it worse when their memory goes unmanaged. Notion agent Memories are private to each user. Decide what they may keep shows how stored context shapes later drafts. A checker who knows what the tool remembers catches errors a stranger would miss. That knowledge lives in one head today. Write it down before it walks out the door.
The heaviest users feel it most. Among staff who use AI several times a day, 56.1% said the oversight takes at least as much effort as training a person. More use does not retire the checking. More use hires more of it. A firm that doubles its AI use without naming checkers doubles an invisible payroll line.
The prompt layer adds quiet hours too. You already paid for your AI tools. In hours, rewriting the same prompts counts the time staff lose re-deriving instructions. A named checker owns the standard prompts as well. One owner ends the daily rewrite.
One rule to set this week
Use the sign-off method that already runs on this site. Review AI-generated work with one signed line before your name goes on it sets one person to sign each piece of AI work before it ships. Extend that rule from posts to every AI task in the office.
Do it on Friday, in one short meeting. List the five AI tasks your team runs most. Put one name beside each task. That name checks the output, owns the prompts and logs the fixes. Give the role two hours a week in the plan, not in the gaps. The survey says the hours exist whether you plan them or not. Plan them.
