Eighty-seven percent of digital workers use AI. Seventy-five percent say it makes them more productive. Workers claim to save roughly eleven hours a week. Yet only thirteen percent say their organization performs significantly better because of AI. That is not an early-adopter wobble. It is the AI productivity paradox, and it has a name: botsitting.
The Glean Work AI Index 2026 surveyed six thousand workers across the US, UK, and Australia with a nationally representative sample and logistic regression controls. The gap between individual savings and organizational performance is real, and it is large.
The 11-Hour Number Measures Perception, Not Net Gain
When a worker says AI saves them eleven hours, what exactly is counted? The minutes the model spent generating the draft? Or the time spent feeding context into the prompt, checking each paragraph against the source documents, fixing the hallucinated citation, and then rewriting the section that sounded plausible but was wrong? The Glean survey asked about all of it, and the breakdown tells a different story.
Of the total time workers spend engaging with AI tools, only 36% goes to what the survey calls "productive use" — actually reading or using the output. Another 27% goes to learning how to use the tools. The remaining 37% — more than a third — goes to botsitting: the work of making AI usable. That averages 6.4 hours per week per worker.
What Botsitting Actually Looks Like
The Glean data splits those 6.4 hours into three subcategories. Each one is worth sitting with:
- 2.3 hours per week feeding context — uploading documents, writing detailed prompts, explaining the background the model needs. This is the work of turning a general model into something that knows your specific domain.
- 2.2 hours per week supervising outputs — reading generated text, checking code for logic errors, verifying that the data table matches the source. This is the work of quality control.
- 1.7 hours per week debugging and correcting — fixing mistakes the model made, re-prompting to get a better result, cleaning up formatting disasters. This is the work of remediation.
I have watched teams adopt AI tools and, six months later, ask themselves why they are still too busy to take on strategic work. The answer is that they are not saving the hours they thought. The model saved five minutes generating the first draft. The human spent twenty minutes turning that draft into something that could be sent to a client. The net loss is fifteen minutes per task.
When Botsitting Turns Into Botshitting
If every hour of botsitting led to a solid, verified output, the overhead might be worth it. But the Glean data reveals a second problem: many workers stop verifying before the work is actually ready. Sixty-nine percent of AI users admit to "botshitting" — shipping AI-generated work they have not fully checked. Seventy-seven percent have corrected or redoe AI-assisted work in the past month.
The more telling number is the 41% who say they sometimes deliver AI-generated work they couldn't explain if asked. That is not a verification gap anymore. It is a competence gap. Heavy AI users are 64% more likely to botshit than light users, which suggests the problem scales with adoption — the more you rely on the model, the less you inspect its outputs.
This matches what I have seen in organizations that pushed AI adoption hard: the velocity of output increases, but the ratio of work that passes a genuine review decreases. The result is a surface-level productivity that masks a decline in quality and in the worker's own understanding of the material.
Three Reasons the Gains Don't Add Up
The botsitting and botshitting numbers describe the mechanism. Three deeper paradoxes explain why organizations see so little aggregate benefit.
The Productivity Paradox
Individual workers feel faster. But the organization does not get faster, because the rework, coordination, and quality checking absorb the savings. Analysis from Digital Applied estimates that unmeasured human rework consumes 22–38% of self-reported time savings in mature AI programs. Even when the worker perceives a gain, the gain is partly fictional — someone downstream is cleaning up.
The Judgment Paradox
AI output is polished. The grammar is correct, the formatting is clean, the tone is professional. That very polish removes the disfluency cues — awkward phrasing, mechanical sentence structure, plain wrong assumptions — that normally trigger a human reader to stop and verify. A report that a junior analyst would have written with obvious gaps alerts a senior reviewer that something needs attention. AI hides those gaps behind clean paragraphs.
I have seen a team spend a week developing a strategy based on an AI-generated market analysis. The analysis was internally consistent, well-structured, and perfectly wrong — the model had hallucinated an entire competitor's revenue figure. Nobody caught it because the output looked right. The judgment paradox is the most dangerous of the three.
The Ownership Paradox
Forty-four percent of workers say AI is fairer than their human manager. That percentage climbs to 50% when the manager oversees ten or more direct reports. When workers feel that their manager is unfair or unavailable, they turn to AI for decisions. The result is that the hardest, most ambiguous tasks — the ones that require human judgment and experience — are increasingly outsourced to a model that cannot be held accountable. The worker who fears replacement clings hardest to the tool that might replace them.
What High Achievers Actually Do
If the story ended here, it would be a cautionary tale with no way out. But the Glean data identifies a group that breaks the pattern: high AI achievers. These are workers who report both high productivity and high quality from AI use. They do not use AI less. They use it differently.
High achievers spend 40% of their AI time on botsitting — versus 33% for low achievers. They feed more context, supervise more carefully, debug more thoroughly. The extra hour or two per week they invest in verification pays back in lower error rates and less rework. They are also 18% more likely to deliberately limit AI use on certain tasks. They know when the tool is not the answer.
The behavior that sets them apart is not technical skill. It is judgment about when to trust and when to override. And that judgment depends on having a deep understanding of the domain — which means protecting time for skill-building, not just for more tool use.
The Sweet Spot and the Toggle Tax
There is also a quantitative upper bound on useful AI engagement. The ActivTrak 2026 State of the Workplace report found that productivity peaks when workers spend 7–10% of their total work hours in AI tools. Beyond that window, productivity plateaus or drops. This is a single-study finding — I would not treat it as a formula — but it aligns with what the Glean data shows about diminishing returns from heavy use.
Part of the drop is what I call the "toggle tax": every switch between an AI tool and the application where the work actually lives creates a context-switching cost. The more tools you use, the more time you lose just moving between them. If your organization has adopted five different AI assistants, the tax multiplies.
For a deeper look at which tools actually deliver measurable time savings and which ones inflate the overhead, see our separate analysis on AI productivity tools that save time vs. those that are just hype.
What to Do
The evidence points to a clear set of actions. None of them involves buying a new tool.
Do the Thinking Yourself
If the task requires your unique judgment — strategy, nuanced client communication, creative problem-solving — consider doing it without AI first. Use the model as a reviewer after you have a rough draft, not as the generator. The time you invest in your own thinking builds the domain expertise that makes your AI use more effective elsewhere.
Build a Context-Rich Environment
The Glean data shows that workers in context-rich organizations are dramatically less likely to burn out or botshit. Making documentation, meeting notes, and project history easily accessible through AI tools — rather than requiring workers to chase down information across ten systems — reduces the feeding-context overhead and improves output quality. It is an infrastructure investment, not a tool choice.
Measure Quality, Not Speed
Seventy-three percent of enterprise leaders feel pressure to show AI ROI that does not yet exist (Zapier). That pressure leads to adoption targets and output metrics that incentivize volume over quality. Instead, measure time to verified completion, error rates in AI-assisted work, and worker confidence in their outputs. Only 41% of agent deployments cross positive ROI within 12 months (Digital Applied). Let the data speak honestly.
For a decision framework on which AI tools to adopt first and how to sequence them, see our structured adoption guide for knowledge workers.
The Best AI Productivity Strategy Is Sometimes Not Using AI
This has been a long path to a counterintuitive conclusion. But the evidence supports it: high achievers limit their AI use, the 7–10% window suggests diminishing returns, and rework consumes a large fraction of claimed savings. The binding constraint on productivity is not the model's capability. It is the human infrastructure around it — the time for context feeding, the patience for supervising, the discipline to verify, the courage to turn the tool off when it does not help.
If you take one thing from the Glean data, let it be this: spending six hours a week botsitting and then shipping work you cannot defend is not productivity. It is busywork in a new wrapper. The real way to win with AI is to know when to walk away from the keyboard and do the thinking yourself.
Comments
Join the discussion with an anonymous comment.