AI Didn't Remove the Work, It Swapped Doing for Watching: Botsitting and the Productivity Paradox

Summary

Glean’s Work AI Institute, working with researchers from universities including Notre Dame, Stanford, and UC Berkeley, surveyed 6,000 full-time workers in the US, UK, and Australia who primarily work on computers or digital tools, between December 2025 and January 2026. The report introduces a new term, botsitting, for the often-overlooked work required to make AI useful: feeding it context, checking outputs, debugging mistakes, cleaning up errors. By its math, white-collar workers spend an average of 6.4 hours a week on it, close to a full working day.

The report also lines up a set of numbers that don’t sit comfortably together. 87% of respondents say they use AI at work and 75% say it makes them more productive, yet only 13% say their organization performs significantly better because of it. The individual view looks great while the organizational view barely moves. That is the productivity paradox it sets out to name.

My read: the signal here deserves to be taken seriously, but don’t treat it as a verdict that AI doesn’t work. Its real value is offering a widely ignored explanatory variable. The saved time did not vanish, it got absorbed by a category of hidden work that no one counts. For teams evaluating AI ROI, what needs to change is not whether you use AI, but which metric you use to keep score and which tasks you bring into the open.

The debate

On the surface the fight is an old one: does AI raise productivity or not. The report’s contribution is to split that vague question into three layers, because the argument usually stalls when people are really talking about different layers.

The first layer is a metric dispute: individual versus organizational. 75% feel more productive, and that is true. Only 13% of organizations perform significantly better, and that is also true. The two numbers do not contradict each other because they measure fundamentally different things. The individual metric asks “am I finishing a given task faster,” the organizational metric asks “is the company’s output, revenue, and delivery better.” Optimists grab the former to claim AI has paid off, skeptics grab the latter to say the spend was wasted. The real question is where the transmission between them breaks.

The second layer is what botsitting answers: where did the saved time go. The report’s explanation is that it gets digested by newly created supervision work. Employees now spend large amounts of time moving information between disconnected AI systems, fixing mistakes, and supplying context the tools should already have, becoming, in the report’s words, human glue between technologies that don’t work well together. Every minute saved at the task level can get handed back at the system level.

The third layer is whether it should be counted: does botsitting qualify as real work. Rebecca Hinds, head of the Work AI Institute and one of the report’s authors, described this work on a podcast as “often tedious” and “exhausting,” and as something that is “not rewarded and it’s not appreciated or tracked or measured and certainly not incentivized within the organization.” That is the sharpest cut. A body of labor consuming nearly a full working day is invisible to every management metric. It does not enter performance reviews, schedules, or cost accounting, so on paper AI both saves time and adds no cost while organizational performance somehow stays flat. The paradox follows from there.

Who’s right

My judgment: the report is more right about where the gap went, but what it offers is still a mechanism hypothesis, not a proven causal verdict.

First, why it is more right. The productivity paradox is not new, economists argued over it in the 1980s and 1990s when IT investment failed to show up in corporate output. This report’s contribution is a concrete, actionable explanatory variable for the current AI version: hidden supervision labor. It does not lean on an emotional “AI is useless” conclusion, it points to one systematically unmeasured link in the value chain. That sits closer to a builder’s reality than either cheerleading or doom. Many teams genuinely feel that everyone reports AI gains while the quarterly numbers stay still, and botsitting gives that confusion something to land on.

Two details make the mechanism more credible. One is that the report points at attrition: workers who spend an unusually large share of their AI time botsitting are 73% more likely to be actively looking for another job. Hidden labor does not just eat efficiency, it quietly drives people out. The other is an irony Hinds raises, that some roles are being asked to automate the part of the job they enjoy most, like customer-service staff who find meaning in building relationships but are increasingly expected to supervise AI agents instead. If that holds, AI is not just failing to lift output, it is eroding the part of the work that keeps people around.

Now, why I hold back. This is a survey led by an institute owned by an AI company, and the remedy it offers, do more work around AI, connect employees to the right context, define what good looks like, build judgment, maps neatly onto the enterprise AI platform Glean itself sells. The conflict of interest is worth stating plainly. The 6.4-hour figure is estimated from a self-report survey, not objective timekeeping, and 87%, 75%, and 13% all come from subjective answers in the same sample. Treating it as a strong signal of what is happening in the industry is fair, treating it as hard proof that AI does not raise organizational productivity is too much. A single report cannot establish causation, it can only name a suspect worth investigating.

Why it matters

For founders and team leads evaluating AI ROI, the practical use of this report is not the percentages, it is that it forces you to switch the metric you keep score with.

If you measure AI gains only by the individual metric, like “engineers say coding is faster” or “support says replies are faster,” you will likely overstate the return, because you have not subtracted the share of saved time that botsitting claws back at the system level. What you actually want to measure is the organizational metric: did delivery cycles, unit output, or revenue move because of AI. The gap between the two metrics is the part being swallowed by hidden labor.

A more concrete step is to make botsitting visible. Right now it enters no metric, so you can neither see its scale nor judge whether a given AI tool is a net time saver or a net burden. Logging the time spent feeding context, checking, and fixing errors for a tool, even as a rough estimate, gives you a basis for deciding whether the spend pays off. The counterintuitive finding is worth copying: the organizations gaining most are often not the ones using AI the most, but the ones doing the most work around it, especially the ones willing to decide which tasks should never be handed to a model.

There is also a cost that gets overlooked: people. If supervision burden really correlates with intent to leave, it is not a one-time efficiency loss but an ongoing attrition risk. A team quietly loaded with cleanup work that goes uncounted and unrewarded will eventually vote with its feet. That cost never appears on the AI tool’s invoice, but it shows up on the hiring and handover bill.

What to ignore

Ignore the “AI is useless” conclusion this report invites but does not support. The report itself acknowledges that 75% feel more productive and 87% use it. What it questions is the transmission link, not AI itself. Using it to argue against adopting AI is a misreading.

Ignore the precision of the 6.4-hour number. It is a survey-estimated population average with no direct bearing on your team, and it will vary widely by role, tool, and maturity. What to keep is the phenomenon it points to, that supervision cost is real and routinely uncounted, not the specific figure.

Also ignore the novelty of the “botsitting” label itself. Naming a category of hidden labor that has always existed has communication value, but the name does not change the substance. What matters is not whether the word exists, but whether you have counted that work inside your own metric. When an AI company defines both the problem (botsitting) and the cure (do more work around AI on a better context platform), accept the diagnosis if you like, but keep extra independent judgment about the cure.

Sources

1. The rise of the botsitters: workers spend 6.4 hours a week babysitting AI (Business Insider) / news

No official primary source available; this analysis is based on reliable secondary reporting (named outlets, cross-confirmed).