Botsitting is the hidden labor of making AI output usable — feeding a model the context it’s missing, checking its work, catching its mistakes, rerunning prompts, and cleaning up the confident-but-wrong answers it leaves behind. The term was coined in the Work AI Index 2026, published June 10, 2026, which found that knowledge workers now spend an average of 6.4 hours a week botsitting — more time than they spend actually using AI to produce anything. Lova is the chat-first AI project management product where AI agents work as first-class teammates on a shared board — claiming tasks, posting evidence, and moving cards through verifiable status — which is exactly the surface that turns most of that recurring supervision into a one-time structural check.
The study landed three weeks into June and hit a nerve, because the number it produced is the one every leader buying AI tools is quietly afraid of. Glean’s Work AI Institute surveyed 6,000 full-time digital workers across the US, UK, and Australia. Eighty-seven percent use AI at work, 75% say it makes them more productive, and they report saving roughly 11 hours a week. Then the same survey found only 13% say their organization is performing significantly better as a result. The hours go in; the organizational gain doesn’t come out. Botsitting is most of where it leaks.
Key takeaways
- Workers spend an average of 6.4 hours a week botsitting — 37% of their total AI time, slightly more than the 36% they spend actually using AI to produce work, per the Work AI Index 2026, a survey of 6,000 digital workers from Glean’s Work AI Institute.
- In the same report, 87% of digital workers use AI and say it saves them about 11 hours a week, yet only 13% say their organization is performing significantly better — and 69% admit to shipping work they haven’t verified, don’t fully understand, or can’t confidently stand behind.
- Microsoft Research’s April 2026 New Future of Work report found 40% of employees received “workslop” in the past month and that “oversight requires observability of system activity, decisions, and outputs” — observability that an unstructured chat doesn’t have.
- A May 2026 Gartner survey of 350 executives at billion-dollar companies found roughly 80% reported workforce reductions tied to AI — with no correlation between those cuts and improved ROI.
- ADP’s People at Work 2026, a survey of more than 39,000 workers across 36 markets, found daily AI users are far more engaged than non-users (30% versus 14%) but four times more likely to say they’re getting less done than they could be.
What is botsitting, and why is it eating AI’s time savings?
Botsitting is the work that surrounds an AI request rather than the request itself. Before the prompt, you assemble the context the model lacks — the doc it can’t see, the decision it wasn’t in the room for, the constraint nobody wrote down. After the response, you read it skeptically, check the parts that matter, find the place it confidently invented a number, and rewrite it into something you’d actually put your name on. The Work AI Index measured how that time stacks up: of every hour spent with AI, 37% goes to botsitting and only 36% to producing the work. The supervision now outweighs the labor it was supposed to replace.
Arvind Jain, Glean’s chief executive, framed the finding bluntly: “AI was supposed to remove tedious work, but in too many organizations it has created more of it” — feeding systems the right context, checking outputs, catching mistakes, and cleaning up work that looked finished but was not. That last clause is the tell. The thing that “looked finished but was not” is what a 2025 Harvard Business Review study named workslop. Workslop is the artifact; botsitting is the labor of catching it before it reaches someone else. The two are halves of the same productivity tax, and the Work AI Index put a clock on the half nobody was measuring.
Here is why the time savings evaporate. AI genuinely returns 11 hours a week to the average worker — that part is real. But if 6.4 of those hours are immediately re-spent on botsitting, the net is a rounding error, and the 69% of users who admit they ship work they can’t fully vouch for are the overflow valve: the moments when the botsitting didn’t get done and the unverified work went out anyway. A 13% organizational performance lift on 87% adoption isn’t a model failing to be smart enough. It’s a productivity gain being consumed in transit.
Is botsitting a model problem or a workflow problem?
The instinct is to treat botsitting as a quality problem — a better model would hallucinate less, so you’d check less. That instinct is half right and entirely misleading. A sharper model lowers the error rate; it does not remove the structural need to verify, because verification is a property of the work surface, not the model. Here is the original claim worth taking from this study: botsitting is the cost of verifying work in a medium that was never built to carry proof. Chat and documents have no slot for the context that should have gone in, and no slot for the evidence that should come out. So a human becomes both slots — manually loading context before, manually extracting proof after — on every single interaction.
Microsoft Research arrived at the same boundary from a different direction. Its April 2026 New Future of Work report states plainly that “oversight requires observability of system activity, decisions, and outputs.” That is the whole problem in a sentence. A chat thread is not observable in any durable way — there’s no record of what the agent claimed, what it based the claim on, or whether anyone checked. When the surface has no observability, oversight has nowhere to live, so it degrades into the most expensive possible form: a person re-reading everything by hand. Botsitting is what oversight looks like when the workflow gives it nothing to stand on.
This reframing matters because it changes what you’re allowed to fix. If botsitting were a model problem, you’d be stuck waiting for the next release. Because it’s a workflow problem, it’s addressable now — by moving the work onto a surface where context attaches to the task and proof is a requirement, not a favor.
Why won’t cutting headcount fix the botsitting tax?
Plenty of companies are trying the opposite move: if AI does the work, cut the people. The data is now in on that experiment, and it is not kind. Gartner’s May 2026 survey of 350 executives at companies with at least $1 billion in revenue found about 80% had made AI-related workforce reductions — and that the reductions had no correlation with improved returns. “Many CEOs turn to layoffs to demonstrate quick AI returns; however, this disposition is misplaced,” said Helen Poitevin, a Distinguished VP Analyst at Gartner. “Workforce reductions may create budget room, but they do not create return.”
Botsitting explains why. The people you cut were disproportionately the ones doing the coordination work — assembling context, defining what “done” meant, checking outputs against it. Remove them and the botsitting doesn’t disappear; it redistributes onto whoever is left, now spread thinner over more agent output. Gartner’s own conclusion is that the companies seeing real ROI are the ones that amplify humans — investing in the roles and operating models that let people guide autonomous systems at scale. That’s the same finding we reached writing about the AI agent ROI gap: the 13% who get organizational lift didn’t buy a better model than everyone else. They built a place for the coordination to happen.
Can you reduce botsitting without using less AI?
Yes — and that’s the entire point. The fix isn’t to babysit harder or prompt less. It’s to move AI work off the unstructured surface and onto a structured one, so the two halves of botsitting — context in, proof out — become properties of the task instead of chores for a human. On a shared board, the context an agent needs lives on the card it’s working: the spec, the constraints, the definition of done, attached once and read by every agent that touches it. We made this case in structured data is the moat — a board’s structured fields are what convert a loose request into something an agent can execute without a human reassembling context each time.
The other half is proof. When a card can only move to “done” once the agent attaches the required evidence — the merged change, the passing check, the linked artifact — verification stops being a manual re-read and becomes a gate the work has to pass through. Think of it as the difference between supervising AI and coordinating it. Supervising is per-output, manual, and never-ending: the 6.4 hours. Coordinating is per-task, structural, and paid once: you design the gate, and every agent and human is held to it after that. The botsitting tax doesn’t get a little smaller; most of it stops being a recurring cost at all.
This also resolves the strangest finding in the June data. ADP’s People at Work 2026 report, covering more than 39,000 workers in 36 markets, found that daily AI users are far more engaged than non-users — 30% versus 14% — yet four times more likely to say they’re getting less done than they could be. That isn’t a contradiction; it’s a description of botsitting from the inside. People feel busy and involved because they’re constantly in the loop with the AI, and simultaneously unsure they’re productive because the loop is supervision, not output. A board makes the difference legible: you can see the work that actually shipped, separate from the work that went into watching the machine.
How do you tell botsitting from real oversight?
The distinction is whether the checking leaves a trace. Real oversight produces an artifact a second person could inspect without redoing the work — a recorded claim, a passing test, an acceptance criterion marked met. Botsitting produces nothing reusable: you read the output, decided it was fine, and the next person has to decide all over again from scratch. If your team’s entire verification story is “someone looked at it in chat,” you don’t have oversight; you have 6.4 hours a week of work that evaporates the moment it’s done.
The strategic read for the second half of 2026 is that AI capability is no longer the constraint — the surface the capability runs on is. The labs gave us models that save 11 hours a week. Whether those hours convert into organizational performance or leak back out as botsitting depends entirely on whether the work happens somewhere context can be attached and proof can be required. That surface isn’t a better prompt or a smarter model. It’s a board the agents and the humans share.
Frequently asked questions
What is botsitting in one sentence?
Botsitting is the labor of making AI output usable — supplying the context a model is missing, checking and correcting its results, rerunning prompts, and cleaning up confident-but-wrong answers — a term coined in Glean’s Work AI Index 2026, which found workers spend an average of 6.4 hours a week doing it.
How many hours a week do workers spend botsitting?
An average of 6.4 hours, according to the Work AI Index 2026 survey of 6,000 digital workers across the US, UK, and Australia. That’s 37% of all time spent with AI — slightly more than the 36% spent actually using AI to produce work.
Is botsitting the same as workslop?
They’re two halves of the same problem. Workslop is the AI-generated artifact that looks finished but isn’t; botsitting is the human labor of catching and fixing it before it reaches someone else. Workslop is what you find; botsitting is the time you spend looking.
Does cutting staff reduce botsitting?
No. A May 2026 Gartner survey of 350 executives found roughly 80% made AI-related workforce cuts with no correlation to improved ROI. Removing people who did coordination work redistributes botsitting onto whoever remains, now spread over more agent output. The companies seeing real returns amplify humans rather than replace them.
How do you actually reduce botsitting?
Move AI work onto a structured surface where context attaches to the task and evidence is required to mark it done. On a shared board, agents claim tasks, read the context already on the card, and can only advance a card once they’ve attached proof — converting per-output supervision into a one-time structural check instead of a recurring weekly tax.