We’re being bombarded by tech companies hyping the transformation of the workforce, which they contend no longer consists solely of human-to-human collaboration but also of human-to-AI agent collaboration. It’s the era of the agentic enterprise or the proliferation of Frontier Firms—whichever marketing label you prefer. Yet, despite all the pronouncements that AI agents will accelerate productivity and free humans to do creative and strategic work, a good chunk of that freed-up time goes right back into managing the agents.
According to this year’s Work AI Index from Glean’s Work AI Institute, of all the time a worker spends interacting with AI each week, 6.4 hours are spent babysitting “botsitting” the agents—the equivalent of nearly a whole workday. In other words, they’re devoting time to filling in missing context, verifying outputs, debugging errors, rerunning prompts, and cleaning up “confident-but-wrong” answers generated by AI. So, while AI may have turned workers into prompt and software engineers, it’s also made them quality assurance and code reviewers.
But that’s not the most significant finding: Workers are becoming fed up with this administrative work, so they’re cutting corners. They’re running fewer checks and taking the agents’ output on faith. They’ve thrown up their hands and no longer want to be the adult. Shouldn’t these agents generate accurate work? After all, that’s what software vendors have promised. It’s led to a trend Glean terms “botshitting”: shipping AI-generated work that workers haven’t verified, don’t fully understand, or couldn’t confidently defend. It’s more common than you might think: 69 percent of AI users said they’re doing it at work.
Not All Botsitting Is Bad, But It Comes With a Cost
Glean’s research showed that botsitting accounts for 37 percent of the time workers spend on AI-related work each week—not 37 percent of their week. The remaining time is split between producing work with AI (36 percent) and learning the tools and building agents (27 percent).
And the 6.4 hours spent on botsitting? Glean doesn’t count all of it as waste. Verifying high-stakes outputs, reworking a prompt until the result is meaningfully better, and supplying domain knowledge a model can’t know are what the researchers called productive botsitting—work anyone accountable for the output would do regardless of the software.
The rest is unproductive botsitting, which is the friction the tools create: loading the same context into one system after another, running the same question through a second and third tool hoping for a better answer, and cleaning up AI-generated work after it breaks downstream.
Still, Glean said that the 6.4 hours per week spent on botsitting is more time than workers actually spend using AI to create the work. And the productive tasks still come with a cost—one that Glean calls ”invisible, unbudgeted, and unsupported.” “Workers who absorb it without recognition or reward grow exhausted,” the researchers wrote. “Then they grow resentful. Then they start polishing their resumes.”
To add a finer point to that fact: those who frequently botsit are 73 percent more likely to be actively job hunting.
The AI Toggle Tax
Two factors drive how much botsitting workers do: how much time they spend with AI, and how many tools they spread it across. Of those who spend at least half their work time with AI, nearly three-quarters (74 percent) reported “frequent” botsitting, Glean’s research showed. That drops to 65 percent among moderate users, who spend 20 to 49 percent of their time with AI. And it’s 35 percent among light users, those who use it less than 20 percent of the time. Glean doesn’t define what counts as “frequent.”
And no single tool will do everything for workers. Glean found that tool sprawl is a major contributor to botsitting—those who use multiple AI tools are 35 percent more likely to report botsitting frequently. The company doesn’t define “multiple,” but its own numbers suggest the bar is low: 77 percent of AI users bounce between tools each week, and a third use four or more tools. If two is enough to qualify, then sprawl isn’t a fringe problem. It’s the default condition of working with AI.
Workers are left to juggle tools and agents, all in the pursuit of completing an assignment. And sometimes, these systems don’t play well with one another. Sure, standards like Model Context Protocol, Agent-to-Agent (A2A), and APIs help to improve integrations, but that’s only structural. What about knowing the task and how the organization actually works? Glean argued that a context gap remains.
All of this creates what Glean calls the “AI toggle tax.” It’s the cumulative cost in time, attention, and sanity of switching between disconnected AI tools, apps, and systems, with the worker left to carry the context, data, and intent from one to the next.
From Botsitting to Botshitting
Someone has to pay the toggle tax, and workers are increasingly deciding it won’t be them. Glean found that as the tax climbs, workers cognitively offload, handing more of their thinking and judgment to the machines. That means a reshuffling of duties. Instead of verifying outputs, checking sources, and asking whether the AI’s recommendations make sense, workers ship the first thing that looks good enough.
Unsurprisingly, researchers found that the more workers use AI, the more likely they are to botshit. According to Glean, among self-reported heavy users, 82 percent admitted to at least one of three types of botshitting behavior—offloading understanding, offloading judgment, or offloading responsibility. That drops to 73 percent for moderate users and 50 percent for light users.
But who is botshitting inside an organization can vary. The report identifies managers as six percent more likely to engage in this behavior than individual contributors or executives. Glean attributes this to managers being between a rock and a hard place—senior leaders want answers and action now, and their direct reports need managers to review their work. So managers sacrifice verification.
Gen Z workers are 19 percent more likely to botshit compared to older colleagues. The report stated that “many haven’t done the work the slow way long enough to know what’s missing or wrong.” These workers have become vulnerable to “fluency bias,” mistaking polished, confident language for accurate information. Just because it looks good doesn’t mean that it’s correct.
Men are also eight percent more likely to botshit than women. Researchers said that women in professional settings have “long paid a steeper price than men for visible errors,” and are likely to double-check their work before it ships. On the other hand, men have a higher tendency to “wave through an AI output that looks ‘good enough’.”
Botshitting isn’t something that happens overnight. Rather, Glean described it as a “slow surrender of agency, one shortcut at a time.” Workers will first stop fully understanding an AI’s output, then stop questioning it. At some point, they won’t feel any responsibility for it whatsoever.
Consequently, botshitting produces what Stanford and BetterUp researchers have called “workslop.” Yet Glean says that’s only scratching the surface. All of this leads to what researchers call moral disengagement: the gradual process by which people stop holding themselves accountable for careless work. When it goes unnoticed, employees take the credit. When it fails, they blame the tool. Glean found that 40 percent of workers blamed AI when AI-generated work fell short, while 29 percent admitted fault.
Here’s what the botsitting-to-botshitting cycle looks like: It starts when an organization deploys AI, not always because it will solve a real problem, but because deploying it signals that the company is keeping up. Workers begin using AI, accepting the additional workload required to make it usable. At some point, fatigue sets in, with workers running out of time, attention, and patience. This is where botshitting occurs: worn-out workers take shortcuts just to get something out the door. However, others will see those unverified outputs, many of whom don’t fully understand them and are now tasked with cleaning them up. Now, teams are spending more time cleaning up poor AI work, with a growing backlog to clear. And organizations will add more AI to speed things up, only for the cycle to repeat itself faster and with higher stakes.
How to Break the Cycle
Breaking this cycle isn’t about introducing better AI tools and more capable models. Glean found those correlate with more botshitting, not less. Doing so could initiate several cognitive shortcuts “that make trust feel earned before it’s justified.” That includes trust through capability, where workers stop paying attention once the system starts performing well; trust through helpfulness, in which workers trust AI when it agrees with them; and trust through humanness, where workers trust AI once it starts to feel more human (e.g., warm and helpful).
Glean’s researchers said the one way to escape the botsitting-to-botshitting cycle is to invest in the “human infrastructure of AI.” In other words, organizations have to build everything around the AI, not the AI itself. That includes the management, judgment, and design work that determines whether the tools produce anything of value. It’s an effort that impacts all levels, from individuals to teams to the company itself.
Individuals
The workers who are getting the most value from AI are the ones pointing the technology at less of their actual job. They’re protecting their core job. High achievers—those AI users who report that AI has improved both their productivity and the quality of their work—are aiming AI at the edges, using it to clean messy data, summarize notes, and poke holes in assumptions while still choosing the model and deciding what the results mean themselves.
There are three reasons why these high AI achievers won’t surrender their core job to AI, according to Glean: First, AI doesn’t always make experts work faster, especially on tasks workers already know how to do. Second, passing skills off to AI can weaken cognitive muscles. Third, it’s a source of pride—these high AI achievers are 4.4 times more likely to be proud of their AI-assisted work than low-achievers because they’ve put more of their effort into it.
By using AI at the edges, high AI achievers are spending more time botsitting. But rather than moping about their situation, these workers are turning it into educational moments. After enough cycles, the worker can develop a theory of what the AI tool can be trusted with and when human intervention is needed. Glean said that high AI achievers are more than twice as likely to consider AI a valuable teacher, learning from it in real time and on real company tasks.
Another consequence of having high AI achievers is that they do it more; they’re much better at quality control. Glean reported that 79 percent of these workers identified and fixed an AI error in the month prior to the report’s publication, compared with 64 percent of low AI achievers.
Teams
It may seem like semantics, but high AI achievers are opting to call AI a “teammate,” not a “tool.” The latter sounds transactional, while the former signals a shift in dynamics. Workers are pushing back on first drafts, explaining what fell short, and telling AI to try again, which could lead to better outcomes. Glean’s study found that 75 percent of high AI achievers trust AI as a teammate, compared with 32 percent of low AI achievers.
Researchers caution that the point isn’t to embrace agents in the same way as humans. It’s about treating them as teammates while ensuring humans remain accountable. “Don’t put a bot on the org chart until you’ve built the guardrails that keep humans responsible for whatever it produces,” Glean warned.
Workers are more likely to adopt AI if they see colleagues doing it, not because of an edict from the C-suite. A top-down mandate tells workers what’s allowed and what’s not, as well as the behaviors that are rewarded. But that’s it. It doesn’t change behaviors. Glean argued that cross-functional teammates have much more influence. “They’re painfully aware of the coordination tax of work: the bottlenecks, the silos, the duplicated efforts, the dropped balls,” its researchers wrote in the report. “When they build an AI workflow or an agent, they aren’t designing for some tidy fantasy version of the work. They’re designing for the messy version that they have to deal with, where the marketer needs the data the analyst hasn’t pulled yet, and the engineer needs the spec the product manager hasn’t written yet. Their workflows spread because they survive contact with real work.”
Glean found that the average worker is 5.6 times more likely to adopt AI by following a cross-functional teammate’s lead. They’re 3.2 times more likely when guided by a direct teammate, and 2.4 times more likely when they see their leader using it.
And remember how managers are most likely to botshit? Researchers found that the best high-AI-achieving managers aren’t trying to compete with AI on coordination work. Instead, they delegate coordination to AI so the technology can draft status updates, route requests, and summarize meetings. These managers can then devote the reclaimed time to management tasks, such as coaching, developing, and inspiring their people.
Organizations
For the company, better management discipline starts with measuring what matters. Instead of evaluating AI’s value by tokens consumed, lines of code generated, login rates, and fun-looking dashboards, focus on quality alongside speed. Whatever is measured is what people will produce.
Measuring the wrong thing can be costly. Glean’s data showed that 74 percent of workers in organizations that measured only productivity admitted they were botshitting. However, in organizations that tracked both productivity and quality, that figure dropped to 64 percent. Organizations that track quality alongside productivity see better results: 83 percent of workers there say AI has improved the quality of their work, compared with 68 percent at companies that measure productivity alone.
Organizations also need to implement an AI governance system that dictates how the technology is used. It’s different from the AI policy, which is a document that codifies which tools are approved, what data can be used with them, and what’s prohibited. AI governance is the operating protocol, not a document. It’s regularly reviewed, revised, and enforced.
Ensuring that AI tools have access to the data in the way human workers actually use it is highly recommended, Glean wrote. Without the right context, workers turn to unsanctioned methods, importing work into tools that IT departments can’t see, concealing their use, and running the same prompt across different platforms in the hope of finding an insightful answer. Workers who say the information they need isn’t accessible through their AI tools are far more likely to feel worn out by it—50 percent, compared with 18 percent of those in context-rich organizations.
Glean said the goal isn’t just about compliance—it’s confidence. Giving workers confidence drives retention. Workers at so-called “transformative” organizations are far more likely to trust their company’s AI strategy (93 percent versus 57 percent), and those who are confident in that strategy are 28 percent less likely to be actively job hunting.
The report concludes that AI automation has saved workers about 11 hours each week, though only 13 percent think their organizations are “significantly” better off as a result. Glean reiterated that those organizations aren’t doing it by adding more AI tools, tokenmaxxing, or creating adoption dashboards. Instead, they’re treating AI as a work-design problem rather than a procurement one. And for those organizations that still haven’t done this, researchers offered this reminder: “AI’s time savings aren’t free.” Any time saved will become botsitting, and any judgment surrendered to AI will become botshitting.
“Build the human infrastructure that makes AI worth using,” the report’s authors wrote. “Or keep paying the bill—in botsitting, in botshitting, and in the steady departure of the people who got tired of cleaning up after the bots.”
The 2026 Work AI Index is based on a survey conducted between December 2025 and January 2026 of 6,000 full-time digital workers across the United States, the United Kingdom, and Australia. It features respondents who mostly worked on a computer or with digital tools, a group Glean said it targeted because AI is most embedded in “digitally mediated work.”
