Workers say AI saves them about 11 hours a week, which is more than a quarter of the workweek. The same survey of 6,000 digital workers found they spend 6.4 hours a week "botsitting": checking, correcting, re-prompting and cleaning up after the tools that were supposed to save them time.
That is more than half of the time saved. It also explains a gap most owners can feel but rarely measure. In Glean's Work AI Index, 87% of workers use AI and 75% say it makes them more productive, yet only 13% say it has significantly improved their organization's performance.
Where the hours go
Glean's researchers split AI-related time into three buckets. Botsitting takes 37% of it, producing actual work takes 36%, and learning tools or building agents takes the remaining 27%.
In other words, the average knowledge worker now spends more time managing AI than using it to do work.
The details will feel familiar. 36% of AI sessions require substantial rework or a full reset, 77% of users juggle multiple tools in a typical week, and 60% run the same prompt through a second tool when the first answer falls short. That isn't a productivity system. It's comparison shopping, done dozens of times a day.
The quieter cost: work nobody can defend
Lost time is the visible problem. The less visible one is what ships when people stop checking.
69% of AI users admit to "botshitting", Glean's term for passing along AI output they haven't properly vetted. 41% sometimes deliver work they couldn't explain if asked, and 28% have blamed AI for mistakes they made themselves.
For a small firm, that's a reputational risk, not a statistic. A proposal, a contract summary or a client report you can't explain is a liability with your name on it.
Heavier use makes it worse, not better. 82% of heavy AI users report at least one botshitting behavior, versus 50% of light users.
The bottleneck is context, not the model
The most useful finding in the research is about why this happens. 53% of workers say the information they need isn't accessible from their AI tools.
Those context-poor users ship work they can't explain at roughly twice the rate of everyone else (54% vs. 26%). They're also far more likely to feel worn out by AI (50% vs. 18%).
An AI tool that doesn't know your pricing, your client history or which version of the template is current will produce fluent guesses. Someone then has to spend an afternoon correcting them.
"Too many companies are treating AI adoption like a vanity metric — more seats, more prompts, more usage," said Dr. Rebecca Hinds, head of Glean's Work AI Institute. "But adoption alone doesn't equal transformation."
Agents will raise the stakes
This matters more now because the next wave of tools acts on its own. Littler's November report cites Gartner's 2026 CIO survey: AI agent deployment is expected to rise from 17% of organizations to more than 60% within two years, with at least 15% of day-to-day work decisions made autonomously.
Readiness hasn't kept pace. The same report says 74% of organizations expect moderate AI use within two years, but only 13% believe they have the right governance in place. It also flags forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027 over cost, unclear value and weak risk controls.
A chatbot that needs babysitting wastes an hour. An agent that needs babysitting and doesn't get it sends the wrong invoice.
Littler's separate labor survey found that only 9% of employers feel "very prepared" for the workplace issues AI will create, and just 17% communicate proactively with employees worried about it.
"Some employers treat AI as an IT project when it's really a workplace change project," said Jonathan Levine, co-chair of Littler's Labor Management Relations Practice Group.
What the 13% do differently
Glean's data on the organizations that do report real gains is more useful than the warnings. Those teams spend only 27% of their AI-related time directly using AI tools, compared with 49% at organizations seeing no impact. The rest goes into setup, measurement and redesign.
They measure more than activity. They track AI across about five dimensions instead of three, including quality, time saved and skills, not just logins and prompts. Where quality is measured alongside output, 83% of workers say AI improved their work, versus 68% where only productivity counts.
They also redesign the work instead of piling AI on top of it. 90% of workers at these organizations say their employer treats AI as a chance to redesign work, versus 54% elsewhere.
For a founder or a lean team, that becomes a one-week context audit:
- List your five most-repeated AI tasks and note how often each needs a redo.
- Give each one a source of truth: current pricing, approved templates, client notes. Feed it to the tool, not to the person fixing the output.
- Pick one tool per job. The prompt-shopping habit is where the hours leak out.
- Track quality alongside speed. If a task gets faster but needs more rework, it isn't a win.
- Name a human owner for every AI output that leaves the building.
What this means
As Stanford economist Erik Brynjolfsson put it in a World Economic Forum paper this year, "the real bottleneck in the digital economy is not invention – it is diffusion."
The tools are already good enough. What separates the 13% is not access to better models but the less exciting work of giving those models the right information and deciding who answers for the result.
As agents move from drafting to doing, botsitting time becomes a direct measure of how ready a business is. Teams that cut it now will be the ones able to hand real decisions to software next year. Everyone else will be paying people to supervise their own shortcuts.
Discussion