The people who sell AI keep telling the world the same bedtime story: the machines will not take your job, they will just do the boring parts so you can do the interesting parts. The Work AI Index 2026 just measured what happens when that story meets a real desk, and the punchline is not interesting at all. It is a second, unpaid job. Six thousand workers across the United States, the United Kingdom, and Australia reported saving eleven hours a week with AI — and burning 6.4 of those hours babysitting the very machines that were supposed to set them free, feeding them context, catching their mistakes, and mopping up the confident nonsense they leave behind. The machine did not replace the worker. It hired the worker as its custodian, and it is not paying.
Rebecca Hinds, who runs Glean’s Work AI Institute and led the study, gave this quiet catastrophe a name — because the first move of any exploitation is to make sure the thing has no name at all. She calls it botsitting: the invisible labor of making AI usable, the feeding, checking, debugging, rerunning, and cleaning that sits between the promise and the deliverable. And when that labor goes unbudgeted and unrewarded, it collapses into something worse — botshitting, the act of shipping AI-generated work you never actually reviewed, could not defend if asked, and quietly hope nobody notices. The report she co-authored is the fullest accounting yet of what AI actually costs the people who use it, and it reads less like a tech whitepaper than a labor grievance. Glean’s Work AI Index lays the whole thing out.
Rebecca Hinds — She Named the Second Job No One Puts on a Timesheet
Here is the number that should keep every manager awake tonight: workers say AI saves them eleven hours a week, and only 13 percent say their organization is actually performing significantly better because of it. Eleven hours saved on the individual’s side, a rounding error on the company’s side. Where did the gain go? Hinds and her co-authors — a heavyweight bench that includes Stanford’s Bob Sutton and UC Santa Barbara’s Paul Leonardi — tracked the hours and found they did not vanish. They got reassigned. Of the time a worker spends with AI, only 36 percent goes to actually using its output. Another 27 percent goes to learning the tools and building the agents. And the biggest slice of all — 37 percent, an average of 6.4 hours every single week — goes to botsitting. For every hour the model hands back, the human hands over most of another hour making it usable. The genie did not just escape the bottle. It put itself on your payroll as a dependent.
Robin the Junior Engineer — 41 Percent of Workers Now Ship Work They Can’t Explain
The report does not speak in abstractions. It gives the collapse a face. Robin, a junior engineer, pastes a thousand lines of AI-generated code into a pull request and goes to bed; the senior engineer, already drowning, spends half a morning untangling code no one on the team — including Robin — can explain. Robin is not a bad employee. He is a statistic: 41 percent of workers now admit to shipping AI output they cannot fully explain, and 28 percent admit they have blamed their own mistakes on the AI. That is the quiet arithmetic of botsitting run to the end of its rope. When the cleanup is nobody’s job, the cleanup stops happening. Sixty-nine percent of AI users now admit to botshitting — delivering work they have not reviewed, do not understand, and could not defend. The technology did not fail. The accountability did, and it failed in the direction that is cheapest for the company and most dangerous for everyone downstream.
The Customer-Service Rep — AI Didn’t Take Her Job, It Took the Part She Loved
The cruelest detail in the whole report is not about layoffs at all. In her interview with Cognitive Revolution, Hinds described a customer-service representative who genuinely liked talking to people — and is now asked to supervise an AI agent doing the talking instead. Her job was not eliminated. It was hollowed out. The part of the work that made it worth getting out of bed for was handed to a machine, and the part that remained is quality control over a bot that keeps getting things confidently wrong. That is the shape of this particular future: not a pink slip, but a desk where the human is the error-correction layer for software, expected to be grateful the machine “helps.” Nobody lost a job today. They lost the reason they wanted it.
Bob Sutton and the Bosses — “Productivity” Is Just Unpaid Overtime With a Better Name
The people running companies have a comfortable word for this arrangement. They call it productivity. Hinds, Sutton, and the rest of the academic team call it something else — the failure to build what they term the human infrastructure of AI, the training, support, and honest accounting that would make the tools actually pay off. The data shows the split is not random. Workers whose AI lacks access to their organization’s knowledge — the “context-poor” — are nearly three times as likely to feel worn out by it, and they are the ones stuck doing the heaviest botsitting, because the model does not know their job, so they have to carry the context to it one prompt at a time. And when that hidden labor goes unlogged and unpaid, people start cutting corners — which is how 36 percent of enterprise AI agent sessions end up failing outright, and how a productivity revolution becomes a giant game of telephone where everyone sounds confident and no one is responsible. SoftwareSeni’s breakdown of the index walks the numbers. The boss sees hours saved. The worker sees a second shift added to the first.
There is a comfortable way to read all this, and it is the wrong one. The comfortable reading says this is a rough patch, that the tools will get better and the babysitting will fade. But the report is clear that the problem is not the quality of the model — it is that nobody planned for the human work the model requires, and the people doing that work are being told their reward is eleven hours they then have to spend feeding context to the thing that ate them. The customer-service rep is not coming back to the conversation she loved. Robin’s thousand lines are still sitting in the pull request. And Rebecca Hinds has given the whole thing a name, which is the first thing any worker gets before a grievance, and the first thing any good system takes away. The machines did not rise up and take over. They did something far more bureaucratic: they hired their owners, one unlogged hour at a time.
Sources: Glean (Work AI Institute), Cognitive Revolution, SoftwareSeni.