87% of Workers Use AI. Only 13% Say the Company Got Better.

Rebecca Hinds, author of *Your Best Meeting Ever* and head of the Work AI Institute at Glean, joins Nathan Labenz on The Cognitive Revolution with the Work AI Index 2026. The survey covers 6,000 digital workers—3,000 in the United States, 1,500 each in the United Kingdom and Australia—fielded in December 2025 and January 2026, then cross-checked against anonymous telemetry from the Glean platform. Eight founding members of the institute designed the questionnaire, spanning psychology, org design, and technology transformation.

The headlines are strange on purpose. Eighty-seven percent of workers now use AI, 73% say it makes them more productive, and they report saving 13 hours a week—a third of a work week. Depending on the wording, time-saved answers ranged from 10 to 14 hours; fully automated work alone came in around 11. Flip the page and only 13% say their organization is performing significantly better because of the technology. People are cashing in. Companies are not. The rest of the episode is about where the hours went.

Glean started in enterprise AI search and now sells a work AI platform: an assistant plus agents, with context as the product. Every employee gets a teammate that is supposed to understand their work, the team’s work, and how the company actually runs. Hinds has been there about 11 months. She trained as an ethnographer, the kind of researcher who embeds in an organization for months or years, and earlier collaborated with Carol Dweck’s Stanford lab on mindsets. Her zoomed-out view is that AI is a human change first and a technology change second, and that organizations keep underestimating fear. Naming it “artificial intelligence,” she says, puts it etymologically in tension with human intelligence. If employees do not see it as something that can amplify their skill, they will resist—or use it only symbolically. What she sees on the ground are islands of excellence: a few people and teams transforming, most stuck, unable to turn personal speed into team and business results.

The report coins two terms for labor that used to have no name. Botsitting is the unglamorous work of making AI usable: feeding it context, watching it, debugging it, cleaning up after it. It is tedious, uncounted, and unrewarded, and it consumes 6.4 hours a week—roughly half of what AI supposedly gives back. Not all of it is bad; the report notes healthy forms. Most of it, Hinds says, is negative, exhausting, and invisible to the org chart. The exhaustion multiplier ranks feeding context first: where the documents live, which version is authoritative, whether this month’s report beats the one from two years ago. Debugging is second. LLMs are probabilistic, so you tweak one thing, rerun, and still do not know whether you fixed it or made it worse.

The second word is harsher. Botshitting is handing downstream AI output you cannot explain or defend. Forty to 41% of employees admit doing that; with a looser definition, 69% do. Labenz jokes that the number has two faces: models are now good enough to fake it, and a lot of the work people are asked to do may not have meant much. The visible form is polished nonsense—a complete-looking deliverable that is hollow when you look closer. Shadow AI sits next to it: unsanctioned tools whose output nobody can trace.

Labenz offers his own near miss. He has a daily agent that researches upcoming podcast guests from his calendar and Drive. For this episode it missed the new report, which a Glean teammate had emailed him rather than dropping into the Drive the agent can see. Maybe 10 to 20% of the brief was usable. He caught it because he knows the beat; someone less careful would have shipped textbook botshitting. Even after a corrected pass, he still had to internalize the material so he could show up as a host, not as someone reading an outline cold.

Hinds’s causal chain is simple. Organizations are under pressure to adopt. Employees spend exhausting, unrewarded hours botsitting. They hit a breaking point and start satisficing: a good-enough output becomes permission to ship. Thirty-six percent of AI sessions fail outright, forcing a scratch restart or major rework. Tools do not talk to each other, models leapfrog, and without a shared context layer the human becomes the integration layer, hauling work between Slack, email, docs, and a pile of AI apps. Token maxing and click counts make it worse. Years ago, when Labenz did AI consulting, he told leaders they could watch token consumption—but never tell the team that was the metric. People are still shooting themselves in the foot that way.

Labenz presses the arithmetic. Thirteen hours minus 6.4 still leaves almost a full workday. At an AI event he asked the room: if the team had to replace you-plus-AI, how many unaided yous would they need? The median answer was two. He puts himself there too. So have we quietly invented a four-day week that nobody wants to admit? His own stance is that botsitting is the price of getting much more done, learning more, and taking on more. He does not fully buy that babysitting models is uniquely alienating compared with ordinary computer work.

Hinds splits the answer. At the individual level, even after botsitting, the net is still positive. Two things go wrong next. First, people pocket the time. In organizations with no clear AI strategy, non-disclosure of usage is rampant. Tell your manager you got six hours back and you will probably get six hours more work. She is not against giving time back for innovation, creativity, or a better life. The problem is that companies cannot see where the hours go, so they cannot catch them at the team level.

Second, and more fatal, individual gains evaporate in coordination neglect. Her concrete example: you expand a bullet into a five-page report; a colleague finds it too long and compresses it back to a bullet. Each person looks productive. Together it is a hamster wheel of AI slop that does almost nothing for the organization. Because leaders cannot map X to a business KPI, the knee-jerk is more tokens, more clicks, more-is-more. The 13% question was a Likert scale, so some people do say things got worse. What worries her more than a slow start is the volume of botsitting, botshitting, shadow AI, and hidden usage—none of which gets better if nothing else changes. The best organizations bake failure in: executives who expect 80% of AI initiatives to fail, because failure is part of innovation. A lack of transformation in some areas can be a healthy sign that people are taking well-intentioned risks and learning.

One of the report’s brain-bending findings is that the employees most fearful of being displaced are often the most eager adopters, going so far as to automate work they would rather keep. Hinds’s chain is psychological: you feel threatened, you do not fully understand the tech, you lack organizational support, and you want to look AI-native. When you automate in a hurry, you point the tool at what you know best. What you know best is often what gives the job meaning.

Customer service is the sketch she reaches for. Representatives who spent years, even decades, building long-term human relationships are suddenly asked to supervise an agent that talks to people for them. Labenz can feel both sides. A people-person did not sign up to sit in an agent-builder UI and mop up failures. His own company uses Intercom’s Fin; customers often leave in a couple of minutes instead of a 30-minute human back-and-forth, which is real value for the customer too. A Stanford study in the report found that 41% of Y Combinator AI startups are automating things people would prefer to keep human. Just because the technology can do it does not mean it should. Friction is sometimes the point. Building the thing yourself creates ownership, judgment, purpose, and pride—the IKEA effect, not a soft extra. Those are hard drivers of performance.

Turnover is already in the data. People who botsit more are more likely to be actively looking; people who botshit more are too. Botsitting correlates with using AI well, because more use means more babysitting, but exhaustion and exit sit next to that. Co-author Paul Leonardi, who studies digital exhaustion, argues that the digital employee experience *is* the employee experience. If you spend your days manually stuffing context into a model while leadership calls the technology transformative, you stop believing the strategy. Losing ownership of AI output is often one piece of a larger disengagement. Berkeley co-author Aruna hypothesizes another subset: they got so good that their market value is higher outside. Labenz adds a colder reading. Some of that exit may be healthy. A lot of people did not love the original job. The real question is whether the new role is one they can, and want to, pivot into. Often, from this data, the answer is no.

Hinds’s technical bet is Glean’s Enterprise Graph: mission, goals, projects, tasks, people, documents, and systems on one map. If the AI works from that, it knows this report shipped this month, that one is two years old, and whose word counts in which domain, so people stop hand-feeding context. Labenz built a personal “Deep Context” system on the same idea and says his own botsitting time dropped sharply once the plumbing existed.

Two scenes are already showing up at customers. One is a smarter split in customer service. An AI that has seen the full interaction history can tell which conversations want instant response and full automation, and which are long-term relationship building that should stay human—and, in the best case, replace the meaning you took away with equally meaningful new work. The second is dynamic staffing. Yesterday you pulled a project team off a static org chart. Tomorrow the model can, across a thousand or ten thousand employees, mix skill, expertise, career ambition, and bandwidth in a calculus no human staffing meeting can run. That, she says, is the North Star of enterprise AI, and Glean customers already have early versions of it.

A smaller finding delighted her and Berkeley’s Aruna. The 13% of organizations that employees say are actually getting better do not only measure more. They put that data disproportionately into employees’ own hands. Hinds has watched the same pattern with collaboration tools and with remote and hybrid work: when the numbers are transparent to the people doing the work, everyone is better off. Another benefit of a company-wide graph is that everyone can query the state of play, not just leadership staring at a dashboard.

Sixty-nine percent admit to some botshitting, which Labenz reads, almost against himself, as a compliment to the models: if two-thirds of people are already passing AI work down the line and the wheels have not fully come off, the tech is further along than the discourse. As a leader he still wants to know who is doing it well. He used to tell teachers not to trust AI detectors. Pangram Labs is getting better press now. Should a company score intermediate work that way?

Hinds says it is too early for a universal answer, and it depends on psychological safety. She has long followed Harvard’s Amy Edmondson: the ideal is people raising their hands—“this output is bad, and here is why.” Detectors will false-flag, and they will push the faking further underground. Context-aware tools that know Rebecca’s default writing style will be more useful than a generic “was this generated” score. The why behind botshitting matters as much as the act. Shadow AI is the tell. The people coloring outside the lines are often the highest performers, because the approved stack cannot keep up with what they can see the technology doing. Crack down blindly and you punish your best. The gold standard is to make the safe path the efficient path.

Are companies paying real money to keep the people who use AI best? Hinds’s answer is blunt: not in a meaningful way, and they should. Colleague Rob Cross, who has studied collaboration networks for years, found high-performing organizations are up to 5.5 times more likely than weaker ones to measure and reward effective collaboration. In the AI era that looks like a tragedy of the commons: everyone optimizing personal output. The better hackathons and “agentathons” she sees do not only prize biggest business impact. They prize best before-and-after prompt, co-creation, and peer feedback—value created together. A context graph, she hopes, will eventually feed fairer performance reviews that weight the collective, not only the individual. Employees already report that AI is informing performance management, hiring, and firing; their expectations for more of that are even higher.

Culture advice has to survive a brutal messaging environment: CEOs prompting in public while cutting 10 to 20% of headcount. Employees, Hinds says, can smell nonsense from a mile away. If the talk track and the practice diverge, trust collapses into company-wide symbolic use. She sits in too many executive rooms where someone decides on a 15% cut, then argues 14 versus 16, as if there were a template. There is not. Flattening and hierarchy experiments fail when they are short-term patches. You have to know what your culture rewards and punishes, and you have to say what AI is in relation to humans.

In the organizations that are actually transforming, employees are much more likely to view AI as a teammate—but not as an employee on the same level. Paul Leonardi’s research is the warning label. When a human assistant errs, we blame the assistant. When an AI errs, we still blame the human who used it. Responsibility rests with people. Once an organization lets staff dump the blame on the model, accountability starts to dissolve. The teammate frame is still useful: unlike a hammer or a calculator, you do not pick it up, use it, and put it down expecting a perfect answer. The interaction *is* the collaboration.

Glean is her test kitchen. The assistant knows how each executive likes to consume information, so before an upward report she asks how this leader reads. On joining, she asked what makes someone successful at Glean; it pointed to team mentality, long-term thinking, and a culture that rewards raising a weird idea. She changed her behavior. It flags unanswered email and action items left in docs. Mission, she keeps repeating, is what you put in when you take hierarchy out. She and Stanford’s Bob Sutton are writing about how to flatten well: hierarchy tells people what to do under uncertainty; remove it and mission has to replace it. A slogan is not enough. If people cannot connect their actual tasks to the mission, they slide into botshitting. Labenz asks whether Elon’s Twitter cuts—from roughly 7,500 toward 1,000 or 1,500, site still up—are tempting copycats. The same answer covers Jensen Huang canceling one-on-ones at NVIDIA: mission-as-boss can replace management motion in one company and fail in 99.9% of others.

Meetings are the amplifier. Hinds has no patience for the new slop of sending a digital twin or a notetaker bot in your place. If a robot can attend for you, the meeting did not need to happen. Meetings are expensive because they are deep human interaction; if they are worth having, a person should show up. How you use AI is becoming a character signal. Botshit a meeting and colleagues read that you do not value their time; design it carefully and use AI with discipline, and they read that too. She is enthusiastic about the other side: measuring meeting health, catching executives who dominate airtime, auto-deleting meetings with no agenda.

Looking a few years out, Hinds separates the vision people love from the path she actually expects. Jobs that never existed will appear. Costs of using the technology will not fall in the short term; she meets executives every week who have blown an annual AI budget in months, sometimes a year of tokens in a month. She does not expect a single-model platform. The future is multi-model and multi-tool, with AI routing work by efficiency, complexity, and cost. Teams will get smaller because coordination tax is what managers used to eat, and AI can take some of that. Roles will rebundle: work that lived in two, three, or four jobs collapses into one, specialists becoming generalists. A CHRO at a large healthcare organization mapped tasks across all roles and found 70% overlap. Once you have that map, a human-versus-agent split becomes a real conversation.

AI itself is neither good nor bad. It will not automatically make organizations better or worse; that depends on how people enact it. What works at an AI-native company will not paste onto a legacy one, and trying is a failure mode. For her own research, the Index is meant to become a pulse every six months. Ethnography—embedding and watching for months and years—will not go away. Survey plus telemetry gets you a piece; lived experience gets you the rest. Her dream embed is a legacy company that has decided to rewire the org chart for real, so she can watch it on the ground.

The practical close is already on the table. Log a week of feeding context, watching outputs, and fixing errors; that number is the start of every optimization. Before you send AI work, ask whether you could defend it face to face. If you run a team, kill metrics based on usage, clicks, and tokens, and look at business results. Hand AI-usage data to employees. Before you automate, ask the people who do the job which parts they want to keep. Every new meeting on the calendar has to earn synchronous time or it should not exist. Individuals have already run the 87% calculation. The next year belongs to the companies that can run the organizational one.