Botsitting Is a Full-Time Job. You're Just Not Paying Anyone for It.

Six point four hours a week. That's the average time a worker now spends re-pasting documents into prompts, watching an agent chew through a task, catching what it got wrong, and cleaning up the parts that were confidently wrong instead of obviously wrong. Glean's Work AI Institute gave this labor a name in its first Work AI Index, published this month: botsitting. Six thousand workers surveyed, researchers from Stanford, Berkeley, and five other universities on the byline, and the number holds up across sectors.

Botsitting is a real job. It just isn't on anyone's calendar or performance review.

That gap is the whole story. BCG's fourth annual AI at Work survey, out in June and covering nearly 12,000 respondents across a dozen-plus markets, found frontline AI use jumped to 74 percent of employees using it daily or several times a week — up 23 points from last year. People report saving real hours. Glean puts the average at 11 a week. And yet only 13 percent of organizations say AI has meaningfully improved company performance. Three-quarters of workers feel more productive. One in eight companies can point to a result.

Azeem Azhar has been tracking this same gap from the strategy side. In a May essay on his newsletter Exponential View, he and Nathan Warren described a tech company where individual engineers using Claude Code produced more code and more pull requests, but the organization's output barely moved. An executive there summed it up better than any survey stat could: "One plus one plus one plus one equals one-and-a-half."

Where the hours actually went.

This isn't a mystery once you look at what botsitting involves. It's supervising outputs, re-explaining context an agent didn't retain, and quietly fixing errors before anyone downstream notices. None of that shows up in a usage dashboard. A tool adoption report will happily tell you that 87 percent of your staff used AI this month. It will not tell you that a third of those sessions failed outright and had to be redone, which is what Glean also found.

Sixty-nine percent of AI users admit to shipping work they never verified. At companies that have already cited AI in layoffs, that number jumps to 94 percent. Sit with that pairing for a second: the organizations under the most pressure to prove AI is paying off are the ones producing the least verified work. Pressure to show ROI is manufacturing the exact outcome it's meant to prevent.

The guidance that never showed up.

BCG's data explains the mechanism. Sixty-six percent of workers get little or no direction on what to do with time AI frees up. More than half never reinvest it into anything strategic. People were handed a tool and told to be more productive with it, and in the absence of a plan, the freed hours defaulted to babysitting the thing that freed them. Strategic clarity, in BCG's own framing, outperforms tool access — employees who get clear direction from leadership do more with less access than employees drowning in tools but no direction. That finding should reorder how a lot of rollout budgets get spent, but it rarely does, because buying licenses is easier than redesigning a workflow.

The standards moved too. Sixty percent of workers say the bar for "good enough" has quietly risen since AI entered their workflow. Forty-one percent now spend more time on decisions, not less, because there's more output to evaluate and more judgment calls about whether it's actually right. AI didn't remove a layer of work. It added a review layer, and no one adjusted the role underneath it to make room.

Ethan Mollick has been describing the mechanism behind that shift for a while now: when a task that used to take weeks starts taking minutes, the bottleneck doesn't disappear. It moves upstream, from doing the work to deciding what work is worth doing in the first place. Botsitting is what that upstream bottleneck looks like when it arrives unplanned: judgment work dressed up as maintenance work, with no time budgeted for either.

What's actually being measured.

Most companies are still tracking adoption: seats activated, prompts sent, sessions logged. It's the number that's easy to pull. It also tells you nothing about whether the work got better once you account for the hours spent supervising it. Glean's head of research put it plainly: too many companies are treating AI rollout like a vanity metric, more seats and more prompts standing in for actual value. A program manager would recognize this instantly in any other context. You wouldn't call a project on schedule because the team logged more hours. You'd ask what got shipped.

The fix isn't complicated, but it is work. It means treating the redesign of a role as the actual deliverable, not just the introduction of a tool. It means someone owning what a specific job does with its freed-up time, the same way someone owns a budget or a timeline. And it means measuring the thing that matters: whether the output got better, or the labor just moved somewhere less visible.

I'd rather see a team report honestly that AI saved four hours and two of them went to babysitting a chatbot than see a dashboard claiming five hours saved with no accounting for where they went. The first number is uncomfortable and usable. The second is a story leadership tells itself.

The tell isn't adoption. It's whether anyone can say, specifically, what a freed hour turned into. Most organizations right now cannot answer that question, and that's the actual AI maturity gap — not how many people are using the tool, but whether anyone designed for what happens after they do.

Further Reading: BCG — AI at Work: Why Strategy Matters More Than Tools — the fourth annual survey behind the strategic-clarity findings · CIO Dive — Employees spend more time managing AI than producing work — the Glean Work AI Institute data on maintenance overhead · Product Impact — Botsitting: Work AI Index 2026 — where the term originates, with the full breakdown of unverified-output rates · Gallup — AI at work puts new pressure on managers — companion data on why managers, not just tools, decide whether AI adoption sticks · Azeem Azhar & Nathan Warren, Exponential View — Why AI isn't showing up on your bottom line — the "one-and-a-half" line, and the strategy-side version of this same gap · Ethan Mollick — Making AI Work: Leadership, Lab, and Crowd — the bottleneck-shift argument this piece leans on, from the researcher most cited on this topic.