Botsitting is the work of feeding AI missing context, checking its output, debugging its mistakes and rerunning prompts until something usable comes out. Digital workers now spend an average of 6.4 hours a week on it —almost a full working day — according to Glean’s Work AI Index 2026, which coined the term.

If you’ve ever run a work task through AI to save time, only to spend just as long correcting its output, that’s botsitting.

The index, a survey of 6,000 full-time digital workers in the US, UK and Australia, finds nearly nine in 10 use AI at work. Three-quarters say it makes them more productive, saving roughly 11 hours a week. Yet only 13% say their organization is performing significantly better as a result. The report argues the missing gains have gone into a fresh, largely invisible form of labor it dubs botsitting.

Mark Hoffman, Work AI research lead at Glean’s Work AI Institute, says the data show a second, even more dubious cousin: companies that have had recent AI-attributed layoffs “are more likely to see ‘botshitting,’ meaning people using AI and passing along its output without checking it.”

Pressure to adopt AI, and workers detaching from employers that don’t support them, may both be at work, he suggests — hurting quality as well as productivity.

A direct report that is AI

Hoffman, who has studied technology adoption at Stanford University, says the researchers labeled this work botsitting because it looks more like management than administration.

“You have a direct report that is AI, and you have to monitor and supervise it, give it context and help it understand what it’s trying to achieve,” he says. It also grows with use: “The more you use agentic tools, the more that work expands.”

The index finds that most botsitting is grunt work. Workers reload context into different tools, catch hallucinations and verify outputs that sound authoritative or simply flatter them. A smaller share is productive, such as checking high-stakes outputs or adding domain knowledge a model couldn’t have had.

Hoffman does point out an upside: “They remain ultimately responsible for the quality of the work, and that keeps them engaged with the AI,” he says.

Whether workers see it that way depends. Hoffman says reactions are mixed, and craft professions feel it most. Designers and software engineers can feel the job has drifted from the one they chose. Managers and executives, by contrast, often see botsitting as a natural next step.

“A lot depends on where you come from and what you find great about your work,” he says.

The context tax

For every 10% more time workers spend feeding AI context, they are 25% more likely to report feeling worn out by it, the index finds. Frequent botsitters — those spending 40% or more of their AI time on it — are 73% more likely to be actively looking for another job.

One major reason is tool sprawl. About 77% of AI users switch between multiple tools every week, and a third use four or more. Workers who use multiple tools are 35% more likely to report frequent botsitting. Some 60% rerun the same prompt across several tools because the first output was too generic, too disconnected or plain wrong.

Standards such as APIs and the Model Context Protocol (MCP) are meant to ease this by letting tools share data. They help with connectivity, the index says, not context.

A well-built integration could pull every Q3 number from a company’s systems, yet still be unable to say which version was final, whether Q3 meant 2025 or 2026, or that finance had restated the figures two weeks earlier. So the worker becomes the go-between, explaining the same project to one tool after another.

Context is also what Glean sells. The company makes an enterprise platform that feeds AI tools company knowledge, and the index draws on its own usage data as well as the survey — so the diagnosis doubles as a pitch. The problem it describes, though, is one many companies recognize.

Could teamwork be the answer?

Fabien Potencier, chief product and technology officer at cloud platform provider Upsun, says his firm’s experience shows how personal botsitting becomes an organizational problem. Last year Potencier told staff to use coding agents with unlimited access. Individual output jumped, though the company’s capacity to build product did not.

“We had more productivity but not more output,” he says.

He believes it’s because agents on private laptops leave knowledge locked away from colleagues. “They’re more productive individually, but the team isn’t,” he says.

Platforms like Upsun are releasing products that move agents out of individual terminals and into shared cloud sandboxes so the supervision happens where the teams already works. “It’s about interacting with the agents instead of watching a screen and waiting for something to happen,” Potencier says.

Spending is not a proxy for results either. Potencier recounts a CTO’s story of an engineer who burned through $1,000 in tokens in two hours after months of building without reading or reviewing the code. The models, he says, “were creating slop.” His conclusion: “There’s no correlation between being productive and spending a lot of money on tokens.”

Hoffman says the same. Organizations that measure both productivity and quality get better outcomes from AI, he says, and “if your measure of AI success is the number of tokens used, all bets are off.”

From doing the work to architecting it

Sherif Mansour, head of AI at Atlassian, says managing an AI is part of a bigger change in what a job is. “People are moving away from doing the thing to architecting AI to do the thing for them,” he says.

AI magnifies whatever skill is already there, he says, giving the example of a team member who is weak at writing for executives and expects AI to fix it. “They just produce more bad communication,” he says. “AI amplifies your current level.”

Which is why expertise is so important. Hoffman says senior software engineers are getting more value from AI than junior engineers are, and the same goes for managers. “Domain expertise is still a huge part of whether people succeed,” he says. A junior who becomes a manager of agents may not know which problems to choose — a cocktail of capabilities that Hoffman says include taste, intuition and expertise.

Potencier agrees: “Experience isn’t something you can replace.” He still looks for curiosity above all when hiring, and argues that juniors must write some code themselves.

Atlassian’s juniors use AI more than its seniors, Mansour says, so the exchange runs both ways: “The juniors teach the seniors new tools and techniques, and the seniors teach the juniors domain expertise. You need both.”

To make room for experimentation, Atlassian runs an AI Builders Week four times a year, when “we tell the whole team to stop for a week, try something, build it and showcase it to everyone.”

Trust, transparency and fear

Part of managing AI well is knowing where its output came from, especially when agents work across a team.

Mansour says Atlassian’s agent collaboration protocol shows the human responsible for any agent’s action, just as it would for a human teammate. “It might be slop, or it might contain nuggets of gold, but either way you can find the human who created it, work with them and make it better,” he says. A team is also working full time on citations: “Anything that gives you an answer should say where it got it.”

To get employees past the feeling that they are babysitting a bot, Hoffman suggests leaders go first — and show their failures as well as their successes. Removing the sense of risk helps too. “We still see a sizable portion of the workforce that’s afraid of being seen as a fraud for using it,” he says.

Then train people on the basics. “People are being told to use AI rather than being given the tools to do it,” he says. His suggested starting point is the dull project-management work — pulling action items from meeting transcripts, keeping track of deadlines — because it “just makes everything around the work far less taxing.”

A craft in itself

So how can managing AI become a fulfilling role? Mansour points to the history of programming, where each layer of abstraction was dismissed as not being the craft, until it became the craft.

“The abstraction keeps rising,” he says. His 12-year-old son, who went straight to typing “AI, build me a game,” regards AI as normal.

The index offers a model for employees to follow. High AI achievers — users who say AI has improved both their productivity and the quality of their work, according to Glean — focus primarily on where they use it. A high-performing analyst might use AI to clean messy data, summarize interview notes and poke holes in her assumptions, but not where her own judgment is needed: which model to build, which variables to choose and what the results mean. They still do the grunt work of correcting the AI, and treat bad output as feedback to ultimately get better-quality results.

The gap between time saved and results delivered, the index concludes, won’t close with more tokens or tools. The organizations narrowing the gap treat AI as a work-design problem rather than a procurement one.

Mansour expects the work to become more satisfying along the way. “I think people will take more joy in designing [a team of] minions to work for them,” he says.

Personalized Feed
Personalized Feed