AI saves your team about 11 hours a week. It also costs them 6.4 hours babysitting it. That is the finding from a new survey of 6,000 full-time digital workers published June 10 by Glean’s Work AI Institute, co-authored by researchers from Stanford, UC Berkeley, Emory, and five other universities. The net gain is real. But it is smaller than anyone expected, and the reason has nothing to do with the models.
The researchers call the lost time “botsitting.” It includes feeding context into AI tools that do not understand the business, supervising outputs that look polished but might be wrong, debugging errors that surface three or four steps after the AI made them, and cleaning up work that was generated but never verified. Workers are doing this across multiple disconnected tools, often repeating the same prompt because none of the systems talk to each other.
Only 13 percent of the workers surveyed said AI has significantly improved their company’s performance. Eighty-seven percent are using it. The gap between those two numbers is the story.
The real cost is not time. It is trust.
Here is where it gets worse. Sixty-nine percent of workers admit to shipping AI-generated work they have not verified. The researchers coined a term for it: “botshitting.” Forty-one percent say they sometimes deliver work they could not explain if someone asked. Twenty-eight percent have blamed AI for mistakes they caused themselves.
This is not laziness. It is exhaustion. When people spend hours feeding context, checking outputs, and switching between tools that do not share information, they eventually stop checking. The supervision burden becomes unsustainable and the quality of oversight drops. The survey found that workers using multiple AI agents are significantly more likely to ship unverified work because the volume of output simply outruns their ability to review it.
The problem compounds. Unverified work creates downstream errors, which create cleanup work, which pulls people away from the tasks AI was supposed to free them from. That is the loop nobody budgeted for.
This is an operations failure, not a technology failure
Most organizations responded to AI the same way: buy the tools, roll them out, tell people to use them. The implicit assumption was that better AI would produce better results. It did not. It produced more output that required more human labor to manage.
The organizations pulling ahead, according to the report, are not spending more time using AI. They are spending more time on the work around it. Setting context. Defining what “good” looks like. Building judgment about what should and should not go to a model in the first place.
That is process design. That is operations work. It is the boring, unglamorous part of AI adoption that nobody wants to talk about at a keynote, and it is the only part that actually determines whether the investment pays off.
The survey reinforces what keeps showing up in every serious study of enterprise AI: the technology works. The workflows do not. A team using three disconnected AI tools with no shared context and no quality standards will produce worse results than a team using one simple tool embedded in a workflow that was designed around it.
What the successful companies are doing
The report is specific about what separates the organizations getting real value from everyone else. They are not buying more tools.
They treat AI adoption as a reason to redesign work, not just accelerate it. They ask what changes in how a task gets done, not just whether AI can do it faster. And they invest in judgment — the hardest skill to build, according to the researchers, is knowing when not to use AI. That is not something you learn from a vendor demo. It comes from practice, from making mistakes, and from organizations that create space for people to share what went wrong alongside what went right.
They also consolidate. Fewer tools, more context. When workers repeat the same prompt across four different systems because none of them understand the business, the time cost adds up fast. The companies getting results are putting their AI where their data already lives and connecting the pieces so people do not have to be the integration layer.
The 6.4-hour number is a diagnostic
If your team is spending a third of their AI time savings on supervision, the answer is not better supervision. It is fewer tools, clearer standards, and workflows that were actually built for how AI works rather than workflows where AI was dropped into a process designed for humans alone.
The organizations that figured this out are not using more advanced models. They are not running more agents. They are running simpler systems that their teams can actually use without spending half the day managing them.
The AI works. It was always going to work. The question is whether your operations are built to absorb it. For most companies, the honest answer is still no. And every week that stays true, the 6.4 hours keep compounding.