Where Did the Eleven Hours Go? The Hidden Work Behind AI at Work
Ask people who use AI at work whether it saves them time, and the answer is emphatic. Ask their organisations whether performance has improved, and the answer gets very quiet.
That gap is the central finding of the Work AI Index 2026, a survey of 6,000 full-time digital workers across the United States, the United Kingdom and Australia. And the researchers’ explanation for where the missing time goes is uncomfortably recognisable to anyone who has spent an afternoon arguing with a chatbot.
The numbers that don’t add up
The adoption figures are what you’d expect by now. Eighty-seven percent of digital workers in the sample use AI at work. Seventy-five percent say it makes them more productive. Workers estimate that automation alone saves them around eleven hours a week – close to a third of a working week.
Then comes the number that stops you. Only thirteen percent say their organisation is performing significantly better as a result.
Eleven hours a week, multiplied across an entire workforce, should be visible somewhere. It should show up in output, in margins, in something. For the overwhelming majority of organisations surveyed, it doesn’t.
The work nobody budgeted for
The researchers argue the hours are being absorbed by a category of labour that didn’t exist five years ago and still doesn’t appear on anyone’s job description: the work of making AI usable.
They give it a name – botsitting – and define it as feeding the system context it should already have, supervising its output, debugging its mistakes, and cleaning up after it. Workers in the survey spend an average of 6.4 hours a week on this. That is more time than they spend actually using AI to produce work.
Broken down, the time workers spend interacting with AI splits roughly three ways: 37 percent goes to botsitting, 36 percent to genuinely producing work with the tool, and 27 percent to learning the tools and building workflows. Part of the reason so much time evaporates is failure rate – respondents reported that more than a third of their AI sessions fail outright, requiring a restart or substantial rework.
Tool sprawl makes it worse. Seventy-seven percent of AI users bounce between multiple tools in a given week, and a third use four or more. Sixty percent report running the same prompt through several different tools because the first result wasn’t good enough. Every switch means re-explaining the same context to a new system. The worker, in effect, becomes the integration layer between tools that don’t talk to each other.
Underneath all of it sits a context problem. More than half of workers said critical information they need to do their jobs isn’t accessible through their AI systems at all. Giving a model access to company data, the report argues, is not the same as giving it the understanding of which version is current, which source is authoritative, or what a particular internal term means in this specific organisation.
When cutting corners becomes the norm
The second half of the finding is the part worth taking seriously.
When the cleanup work is invisible, unbudgeted and unrewarded, people stop doing it. The researchers call the result botshitting: shipping AI-generated work you haven’t verified, don’t fully understand, and couldn’t defend if questioned. Sixty-nine percent of AI users in the survey admitted to at least one such behaviour.
The specifics are sobering. Forty-one percent said they sometimes deliver AI-generated work they couldn’t explain if asked. Twenty-eight percent have blamed AI for a mistake that was actually their own.
And it scales with use. Among the heaviest AI users, the share who said they couldn’t explain their own outputs reached 54 percent, against 24 percent among light users. The more someone relies on the tool, the less able they are to account for what it produced on their behalf.
There is a counterintuitive wrinkle here too. The report found that the tools whose users reported the biggest productivity gains were also the tools whose users admitted to the most unverified shipping. Better models don’t automatically produce more careful users. They can produce less careful ones, because polished, confident output removes the small friction signals – the typo, the clumsy sentence – that used to make a reader slow down and check.
What separates the people getting real value
The most useful section of the report is its profile of what it calls high AI achievers: people who report gains in both productivity and quality.
What distinguishes them isn’t volume of use. It’s where they point it.
They spend a larger share of their AI time on botsitting – 40 percent, against 33 percent for lower achievers. They’re 18 percent more likely to deliberately hold back from using AI on certain tasks. They’re markedly better at catching errors: 79 percent caught and fixed an AI mistake in the past month, against 64 percent of low achievers. And they tend to keep the core of their craft for themselves, using AI on the surrounding work rather than the judgment calls at the centre.
The hardest capability, by the report’s own measure, is knowing when not to use the tool. Only a third of workers said they were extremely confident in that judgment.
One group is flagged as particularly exposed: workers early in their careers were more likely than older colleagues to ship unverified output. The suggested reason is that they haven’t done the work slowly enough, for long enough, to recognise when something is missing. Confident language reads as competence when you don’t yet have the experience to know otherwise.
What to do with this
If the study is right, the productivity problem with AI isn’t that the tools are bad. It’s that the work of supervising them is real, substantial, and almost entirely unaccounted for – so it gets skipped.
At an individual level, that suggests a few practical things. Notice how much of your AI time goes to fixing rather than producing. Treat the ability to spot a wrong answer as a skill that needs deliberate maintenance, not something that comes free with tool fluency. And be honest about the pieces of your job where the supervision cost exceeds the time saved, because those are the pieces worth keeping.
A note on the source: the Work AI Index is published by the Work AI Institute, backed by Glean, a company that sells enterprise AI software. The findings on context gaps and tool sprawl happen to align with what that company sells. The academic contributors are named and the methodology is published, including the caveat that the data is self-reported and the sample skews towards tech-heavy, high-adoption workplaces. Worth reading with that in mind – and worth reading anyway, because the pattern it describes will feel familiar to almost anyone who has used these tools under deadline.
Sources
- Work AI Institute, Work AI Index 2026: Botsitting, Botshitting, and the Hidden Human Labor of AI at Work — survey of 6,000 full-time digital workers in the US, UK and Australia, conducted December 2025 to January 2026 – glean.com/work-ai-institute