0
0
You ask the machine a question. It answers in two seconds, sounds completely sure of itself, and hands you something that looks finished.
Then you read it properly. A name is wrong. A number came from nowhere. It missed the one thing about your job that would have changed the whole answer. So you go back, explain what it should have known, and ask again.
That second part now has a name. It is called botsitting, and a large survey of office workers has put a number on it: about 6.4 hours a week.
What botsitting actually is
Botsitting is all the work you do to make an AI answer usable. Feeding it the context it was missing. Reading its output closely enough to catch what is wrong. Fixing the mistakes. Running the prompt again, worded better this time. Quietly cleaning up after a confident answer that turned out to be nonsense.
None of it appears on a task list. Nobody was hired to do it. There is no job title for it. But it is now a real part of the working week for a lot of people, and the 2026 Work AI Index — a survey of around 6,000 full-time desk workers in three countries, carried out over December 2025 and January 2026 — is the first thing to size it properly.
Their headline figure: 6.4 hours a week. Slightly more of a worker’s total AI time goes into botsitting than into actually producing anything with it — 37 percent against 36 percent, with the remaining 27 percent spent learning the tools. Those first two numbers are close enough that the honest way to say it is simply this: about as much time goes into managing the machine as into making things with it.
Where the 6.4 hours actually go
This is the part of the report that is easy to read past, and it is the most interesting thing in it. The 6.4 hours break down roughly like this:
- 2.3 hours feeding the AI context it was missing
- 2.2 hours supervising and checking its output
- 1.7 hours debugging its mistakes
Look at the biggest slice. The single largest cost is not fixing wrong answers. It is telling the machine the things it needed to know before it answered — and you are telling it afterwards, because it did not stop to ask.
That is the whole problem in one line. The tool does not pause. It does not say “I do not know enough about your situation yet.” It produces something fluent and certain from whatever fragment it was given, and the gap between what it knew and what it needed gets paid for later, by you, in hours.
A wrong answer delivered with total confidence is more expensive than no answer at all. No answer costs you the time to go and find one. A confident wrong answer costs you that, plus the time to discover it was wrong, plus whatever went out the door before anyone noticed.
Everyone feels faster. Almost nobody looks faster.
Here is the strange gap the same survey turned up.
Eighty-seven percent of desk workers now use AI at work. Seventy-five percent say it makes them more productive. Workers estimate it saves them somewhere around 11 hours a week.
And only 13 percent say their organisation is performing significantly better because of it.
Worth being precise about what that is and is not. These are people’s own estimates of their own time, not a stopwatch. Self-reported hours saved are famously generous. But the gap is big enough to be interesting even if you discount the numbers hard: a great many people feel faster, and very few workplaces can point at anything that got better.
Some of that gap is botsitting eating the savings. Some of it is that time freed up in one person’s afternoon does not automatically become anything a company can measure. It is a similar shape to what happens when employers are asked what they actually did with AI rather than what they planned to do — the story on the ground is quieter and messier than the story in the headlines.
What happens when people get tired of checking
This is the finding that should worry people most, and it was not in any of the headlines.
The researchers gave a name to what happens when botsitting wears you down: shipping AI work you have not properly reviewed, do not fully understand, or could not defend if someone asked you to. Sixty-nine percent of AI users admitted to at least one behaviour like that.
And it is not the beginners. Among heavy AI users it was 82 percent. Among light users, 50 percent.
Read that twice, because it turns the obvious advice inside out. The natural conclusion from “AI creates cleanup work” is “get better at AI, use it more, learn the tools.” But the people using it most are the most likely to have stopped checking. Not because they care less. Because checking everything, all day, forever, is exhausting in a way that nobody planned for or staffed for.
Vigilance is a limited resource. You cannot be the last line of defence for eight hours a day indefinitely. Eventually something goes out unread — and the people it happens to are usually the conscientious ones who have simply run out of road. If that is starting to sound familiar, it may be worth taking ten minutes with our free Am I Burned Out at Work? check, which asks about exactly this kind of invisible, unending load.
It is the same reason some people feel tired all the time while every test comes back normal. The effort is real. It just never shows up anywhere that counts it.
Four things that cut the cleanup
None of these cost money, and none of them need permission from anyone.
- Front-load the context. The biggest slice of those 6.4 hours is context you supplied after the fact. Supplying it first is the same work, done once instead of twice.
- Make it ask you first. Before requesting the answer, ask what it would need to know to answer well. You get a list of your own blind spots, and it takes about twenty seconds.
- Decide the check before you see the output. Name the one thing that must be true for this to be usable, and check that. A plan made in advance survives tiredness much better than good intentions made at five in the afternoon.
- Keep one thing you never hand over. Pick the piece of your work that has to be yours — the judgement, the final read, the thing with your name on it. Being able to stand behind your work is not a luxury. It is the part that makes the rest of it mean anything.
This is not a new problem
Long before anyone built a machine that could answer, people were writing down what they had learned about living well, and they kept circling the same warning: whoever answers before they have listened has not saved anybody time. They have only moved the cost somewhere else, usually onto a person who now has to clean it up. They did not think of it as inefficiency. They thought of it as a kind of shame — because rushing to answer is almost always about how you want to look, not about what the other person needs.
The same old writings treat the willingness to say “I do not know yet” as something given rather than something manufactured — wisdom as a thing you ask God for, not a skill you grind out by working harder. Which is a peculiar idea to run into inside a report about software. It also happens to describe the only move that reliably helps: stopping, before you answer, long enough to find out what you actually know.
The hour you get back
The tools are not going away, and most of the people in that survey did not want them to. Something genuinely does get faster.
But the hours come back in a different place than anyone expected. Not from getting the answer sooner. From asking the question better — which is slower at the start, and then stops costing you the afternoon.
Six point four hours a week is most of a working day. That is a real thing to want back. And the way to get it, it turns out, is the least futuristic advice available: find out what you are dealing with before you open your mouth.
Over to You
Do you think the cleanup work will shrink as these tools get better — or is a confident wrong answer just something we are going to be living with from now on? Tell us what you have actually noticed in your own week. We read every reply.
Share This
“Turns out the biggest slice of time people lose to AI isn’t fixing its mistakes. It’s telling it the things it needed to know before it answered. 2.3 hours a week explaining yourself to something that never asked.” https://bgodinspired.com/index.php/bgodinspired-news/bgodinspired-technology-news/botsitting-6-4-hours-fixing-ai/
“69% of AI users admit to shipping work they haven’t really checked. Among heavy users it’s 82%. It’s not carelessness — nobody can stay vigilant for eight hours a day forever. Worth a read.” https://bgodinspired.com/index.php/bgodinspired-news/bgodinspired-technology-news/botsitting-6-4-hours-fixing-ai/
“87% of desk workers use AI. 75% feel more productive. 13% say their company is actually performing better. That gap has a name now, and it costs about 6.4 hours a week.” https://bgodinspired.com/index.php/bgodinspired-news/bgodinspired-technology-news/botsitting-6-4-hours-fixing-ai/
Questions People Are Asking About Botsitting
What is botsitting?
Botsitting is the unplanned work of making an AI tool’s output usable: giving it context it was missing, checking what it produced, debugging its errors, re-running prompts, and correcting confident answers that turned out to be wrong. The term was popularised by the 2026 Work AI Index, a survey of roughly 6,000 full-time desk workers in the United States, the United Kingdom and Australia. The survey found workers spend an average of 6.4 hours a week on it.
How many hours a week do workers spend fixing AI output?
About 6.4 hours a week, according to the 2026 Work AI Index survey of around 6,000 full-time desk workers. That breaks down into roughly 2.3 hours feeding the tool missing context, 2.2 hours supervising its output, and 1.7 hours debugging mistakes. It is slightly more of workers’ total AI time than they spend producing actual work with the tools — 37 percent against 36 percent.
If AI saves 11 hours a week, why do companies not perform better?
The 2026 Work AI Index found 87 percent of desk workers use AI and 75 percent feel more productive, with workers estimating around 11 hours saved a week — yet only 13 percent said their organisation was performing significantly better as a result. Two things explain much of the gap. Cleanup work absorbs a large share of the time saved, and hours freed up in an individual’s day do not automatically turn into anything an organisation can measure. The hours saved are also self-reported estimates rather than measured time.
Why do experienced AI users ship unchecked work more often than beginners?
The 2026 Work AI Index found 69 percent of AI users admitted to shipping work they had not adequately reviewed, did not fully understand, or could not confidently defend. The rate was 82 percent among heavy users and 50 percent among light users. The likeliest explanation is not carelessness but fatigue: constant verification is draining, and people who use the tools all day face far more of it, so their attention runs out sooner. The finding is correlational and does not prove that heavier use causes the behaviour.
How can I reduce the time I spend cleaning up after AI?
Four practical steps, none of which cost anything. First, give the tool the relevant context before asking rather than after, since supplying missing context is the single largest slice of cleanup time. Second, ask the tool what it would need to know to answer well before requesting the answer itself. Third, decide in advance what single thing must be true for the output to be usable, and check that specifically. Fourth, keep one part of the work — usually the final judgement — that you never hand over, so you can always stand behind what you produce.
