A new Glean / Work AI Institute report names the hidden labor eating the AI productivity dividend — "bot sitting" — and its uglier cousin, "bot shitting." NLW digs in because, unlike most 2026 studies, he thinks the agentic shift would amplify these findings, not erase them.
AI's individual gains keep disappearing into invisible work — and the fix isn't a better tool, it's transformation.
Workers save 11 hours a week with AI but burn 6.4 of them "bot sitting" — feeding context, checking outputs, debugging, cleaning up confident-but-wrong answers. The Glean / Work AI Institute report frames this as the reason only 13% of organizations say AI made them significantly better. NLW partly disagrees on the mechanism — individual gains never automatically become organizational gains — but argues bot sitting and its degenerate form, bot shitting, are real artifacts of the transition. The differentiator isn't the AI you buy; it's the human infrastructure you build across individuals, teams, and organizations.
NLW has covered fewer consulting and research-house studies this year because the paradigm shifted so far from non-agentic to agentic work that anything measuring non-agentic work feels largely irrelevant. His bias is toward opportunity AI — doing new things — not just efficiency AI doing the same work faster.
Per the report: 87% of digital workers now use AI, 75% say it makes them more productive, and they save an average of 11 hours per week — yet only 13% say their organization is performing significantly better as a result.
NLW disagrees with the report's core explanation. He argues individual productivity gains — wherever they come from — do not inherently translate into organizational gains unless there's an actual mechanism to facilitate the transformation, even if bot sitting disappeared entirely.
— Glean / Work AI Institute report. The report's definition of bot sitting: feeding AI missing context, checking its outputs, debugging its mistakes, rerunning prompts, and cleaning up the confident-but-wrong answers it leaves behind. Workers burn an average of 6.4 hours a week on it.
Bot sitting is a byproduct of AI succeeding individually. Surveyed workers said AI now automates about 27% of their output and expected that to climb to 35% — and 57% said they want AI to automate even more of their job than they think it ultimately will be able to.
The report's most striking chart breaks AI time into thirds: 27% learning and building agents, 36% actively using AI to complete work, and 37% bot sitting. Bot sitting splits into productive (verifying high-stakes outputs, iterating prompts, adding domain context) and unproductive (reloading context, comparing tools, cleaning up).
Of weekly bot-sitting time: 2.3 hours feeding AI context (14% of total AI time), 2.2 hours supervising outputs, 1.7 hours debugging, and roughly 10–15 minutes on cleanup or switching tools.
For every 10% more time workers spend feeding AI context, they're 25% more likely to report feeling worn out. Frequent bot sitters — those spending 40%+ of AI time on it — were 73% more likely to be actively hunting for another job.
Tool sprawl is a major driver: workers using multiple AI tools are 35% more likely to report frequent bot sitting, and 60% are rerunning the same prompt across multiple tools because the first output wasn't good enough — what the report calls the "AI toggle tax."
The report's bleeped term for cognitively offloading too much to AI — shipping the first output that looks good enough instead of one you can explain, defend, and stand behind. It's described as "a slow surrender of agency one shortcut at a time": workers stop understanding the output, stop interrogating it, then stop feeling responsible for it.
When AI-generated work fails, 40% of workers blame the AI and only 29% admit it was their own fault. Heavy AI users are 3.4 times more likely than light users to blame the tool — what the report calls moral disengagement.
The report's cycle: deploy AI → bot sitting rises → fatigue sets in → bot shitting (fatigued workers take shortcuts) → unverified outputs move upstream → cleanup and rework pile up downstream.
The 6,000 respondents answered in December 2025 and January 2026, before tools like Claude Code drove autonomous agent work. NLW argues that, unlike most 2026 studies, a rerun today wouldn't nullify these findings — it would amplify them.
The tools with the biggest reported productivity gains were also the ones whose users admitted to the most bot shitting. ChatGPT drove 67% productivity gains, Claude 59% — and their users reported the highest rates of shipping unverified work (71% and 92% admitting to it at least monthly).
As people use AI to do things previously beyond their abilities, a new category of bot shitting emerges — not laziness, but lacking the capability to verify outputs. NLW feels it himself judging new coding models impressionistically because he's never coded; democratization of skills is the upside, but the verification challenge is built in.
High AI achievers spend ~38% of their AI time on core job tasks versus ~48% for low achievers — keeping more core work themselves. NLW thinks this is a temporary state (Claude Code's creators barely code anymore), but the lasting lesson is being discerning and trusting your own judgment to lead the AI.
High AI achievers actually bot-sit more, but orient it toward improvement: they're more than twice as likely to rate AI as a valuable teacher, and they reinvest the saved hours into new skills rather than just more work.
A leader using AI makes the average employee 2.4× more likely to adopt; a direct teammate, 3.2×; but a cross-functional teammate, 5×. The reason: their workflows survive contact with real organizational messiness — silos, bottlenecks, dropped balls — not a tidy fantasy version of work.
— Glean / Work AI Institute report. High-achieving managers delegate 32% more of their coordination time to AI — drafting status updates, routing requests, summarizing meetings — and reclaim that time for coaching, developing, and inspiring people.
Workers with good managers are roughly twice as trusting of AI in sensitive calls: 53% are comfortable with AI in performance reviews versus 26% for those with bad or average managers — the same pattern holds for pay and termination decisions.
71% of workers at transformative organizations can see their own AI usage data versus 40% at non-transformative ones — turning AI into a feedback mechanism to improve rather than a surveillance tool for deciding who gets fired.
At transformative orgs, 93% of workers say their AI policy gets reviewed (vs. 55%), 91% say the rationale is explained (vs. 57%), and 93% trust the company's AI strategy (vs. 57%). Treating AI strategy as just a vendor-selection choice is itself a hallmark of non-transformative organizations.
84% of workers at transformative organizations say AI skills are formally rewarded (vs. 48%), and 90% say they get enough AI training and support (vs. 52%). The investment is in people, not just tools.
NLW's closing argument: AI transformation is about messy, complex change in what you do — not just implementation of new tools. Bot sitting and bot shitting are part and parcel of a transition we'll all live through for years. There are no short answers, only organizations willing to do the work and those that aren't.