Short answer: AI saves people time, then much of it goes into checking, fixing and forwarding output that only looks finished. In Glean’s Work AI Index 2026, workers say AI saves them about 11 hours a week, yet they spend 6.4 hours a week botsitting. The fix keeps the AI and redesigns the workflow around it.
AI workslop is output made with AI that looks finished but isn’t, so someone else has to check it, fix it or redo it. Botsitting is the time people spend making AI usable: pasting in context, checking drafts, rerunning prompts. Together they explain why AI is not saving your team the time you expected. In Glean’s Work AI Index 2026, workers say AI saves them about 11 hours a week, yet they spend 6.4 hours a week botsitting. The fix keeps the AI and redesigns the workflow around it: the context arrives with the task, “done” is written down before anyone prompts, and checking happens once, at a designed approval point where the reviewer sees the sources.
That is our argument. Below: what the two terms mean, where the hours actually go, a five-day audit you can run in your own team, and the design changes that remove the leak.
What are workslop and botsitting?
Workslop comes from researchers at BetterUp Labs and the Stanford Social Media Lab, who introduced it in Harvard Business Review in September 2025. In our words: work produced with AI that has the look of a finished product, such as tidy slides, long memos or confident summaries, but does not move the task forward, so the person who receives it ends up doing the thinking. In their survey of 1,150 full-time US desk workers, 40 percent had received workslop in the previous month. Each incident took about two hours to resolve, which BetterUp puts at $186 per employee per month and $9 million a year for a 10,000-person company. HBR’s own summary gives 41 percent. Either way, these are self-reported figures from one survey. The researchers also report that receiving workslop changes how people see the colleague who sent it.
Botsitting is newer. Glean’s Work AI Institute introduced the term in its Work AI Index 2026, published in June 2026, and defines it as the largely unrecognized, unbudgeted and untracked labor of making AI usable: feeding it context, supervising its output, debugging its mistakes and cleaning up after it. The word is Glean’s, not ours.
The two terms describe one ledger from both ends. Botsitting is what the person using AI pays to make the output good. Workslop is what the next person pays when the first one skips that work. Glean reports that 69 percent of AI users admit to shipping AI-generated work they have not verified, do not fully understand or could not defend.
Why isn’t AI saving your team time?
Because the hours saved at the first step get spent at the next ones. Glean surveyed 6,000 full-time digital workers in the US, UK and Australia between December 2025 and January 2026:
Of all the time workers spend with AI each week, 37 percent goes to botsitting, 36 percent to producing work and 27 percent to learning tools and building agents.
Feeding the AI context takes 2.3 hours a week, supervising its outputs 2.2 hours and debugging 1.7 hours.
Workers say more than a third (36 percent) of AI sessions fail outright and need a restart or substantial rework.
75 percent say AI makes them more productive. Only 13 percent say their organization performs significantly better because of it.
Workday found the same pattern separately. In its survey of 3,200 full-time employees at companies with $100 million or more in revenue, fielded in November 2025, 85 percent save one to seven hours a week with AI, but nearly 40 percent of that saved time is lost to rework: correcting errors, rewriting content and verifying outputs. Only 14 percent of employees consistently get clear, positive net outcomes. Developers describe the same thing. In Stack Overflow’s 2025 Developer Survey, the top frustration, named by 66 percent, was AI solutions that are “almost right, but not quite,” and 45.2 percent said debugging AI-generated code is more time-consuming.
A caution on all of this. These are surveys of what people believe about their own time, and Glean and Workday both sell AI software whose pitch fits their conclusions. We use them because they agree with each other and with the measured data further down, not because any single number is precise.
Where do the hours go?
Follow an AI-assisted output from request to final use and the lost time tends to leak in one of three places. The three leaks are our framing, not a research finding.
The context leak
People paste in what a system should already know: the client file, the current price list, the standard operating procedure, last quarter’s version of the report. In Glean’s survey, 53 percent of workers say critical information they need to do their jobs is not accessible through their AI systems. So the person becomes the integration layer. Every paste costs minutes, and every paste is a chance to paste the wrong version.
The check leak
Checking is either duplicated or missing, and rarely designed. The author skims the draft, the manager rechecks the numbers, finance rechecks them again, and nobody has the source open. Or nobody checks at all. In KPMG and the University of Melbourne’s global study of more than 48,000 people in 47 countries, the findings on AI at work show 66 percent relying on AI output without evaluating its accuracy and 56 percent making mistakes in their work because of AI. Trust in the tool makes it worse. A Microsoft Research survey of 319 knowledge workers found that higher confidence in generative AI is associated with less critical thinking.
The forward leak
Unchecked output travels, and workslop is what arrives. Matthias Holweg and Thomas Davenport argue in HBR that when workslop happens in sequence across a process, errors compound, trust erodes and the gains disappear. Their summary puts it plainly: “Generative AI’s gifts come with a hidden danger: decay in the accuracy and quality of organizational knowledge.”
Software teams show where the forward leak lands, because their tools record it. Faros AI’s telemetry from more than 10,000 developers across 1,255 teams found that developers on teams with high AI adoption complete 21 percent more tasks and merge 98 percent more pull requests. Review time rises 91 percent and average pull request size by 154 percent, and Faros found no significant correlation between AI adoption and improvement at the company level. The work did not disappear. It moved to the reviewer.
Is the AI productivity paradox real?
So far, yes, and it is less mysterious than the name suggests. Individual speed is real. But a company’s output is a chain of handoffs, and the chain moves at the pace of its slowest link, which is now often the check.
The best large-scale evidence points the same way. Linking adoption surveys to Danish payroll records, economists Anders Humlum and Emilie Vestergaard estimate precise null effects on earnings and recorded hours, ruling out effects larger than 2 percent two years after ChatGPT’s launch. What did change was the structure of work, including new tasks in AI oversight and integration. The hours saved at one desk are partly spent at another desk, checking.
That is why freed hours are not savings until someone decides what happens to them. We cover how to count them in how to measure AI ROI.
Is the answer less AI?
No, for two reasons.
First, some botsitting is the job. Glean separates productive botsitting, such as verifying high-stakes outputs or adding domain context the model could not have known, from unproductive botsitting, such as reloading the same context into several tools or cleaning up after a handoff. Its “high AI achievers,” people who report both productivity and quality gains, spend more of their AI time botsitting than low achievers: 40 percent against 33 percent. The goal is one deliberate check in place of several accidental ones.
Second, the cause sits upstream of the person. Gerrit Kazmaier, Workday’s president of product and technology, put it this way: “Too many AI tools push the hard questions of trust, accuracy, and repeatability back onto individual users.” When every employee prompts a general chatbot on their own, every employee becomes an integration layer and a one-person quality department. The workslop researchers point the same way: their HBR summary asks leaders to consider how they may be encouraging indiscriminate mandates with too little guidance on quality standards. That is a design choice, and a CEO can change it.
Holweg and Davenport recommend restricting generative AI and defining what value it adds. We agree with defining the value. We would make restriction the exception, not the default: drop AI where it adds nothing, and redesign the workflows where it does. It is the same argument we made in the model is no longer the advantage: the gap between companies opens in the workflow built around the model.
How do you find workslop in your own team this week?
Run a five-day audit. It takes about three hours of your time and 15 minutes a day from the people involved. Glean’s categories informed the time log. The audit itself is ours.
Day 1: pick three outputs. Choose recurring outputs where people already use AI and that someone else depends on: a client memo, a weekly sales report, a proposal, a customer reply, a deviation summary. Write down who produces each one, who receives it and what happens next.
Days 1 to 5: log the time. Ask everyone who touches those outputs to note minutes per instance in five buckets:
Bucket | What counts |
|---|---|
Context | Finding and pasting background, files, previous versions |
Producing | Prompting and generating the draft |
Checking | Reading against sources, verifying numbers and claims |
Fixing | Rewriting, rerunning, correcting |
Downstream | Time recipients spend asking, rechecking or redoing |
The downstream bucket matters most and nobody records it, so ask the recipients directly.
Day 5: trace three instances of each output. Sit with the owner and walk each one from request to final use. Ask five questions:
How many people checked it after the AI step, and did any of them have the source open?
What did the author paste in that a system already holds?
What changed between the draft and the final version, and who changed it?
Could the sender point to a source for every number and claim?
Is there a written definition of “done” for this output?
Then read the results. These red flags are our rules of thumb, not research thresholds:
Context, checking and fixing together take longer than producing. The AI is drafting, and people are doing the work.
Two or more people check the same output without defined roles. That is the check leak.
The downstream bucket is bigger than the author’s checking. That is the forward leak, in other words workslop.
Someone pastes the same background more than once a week. That is the context leak.
Nobody can answer question 5. Then nobody can tell finished work from finished-looking work.
Finally, compare total minutes per instance with how long the output took before AI. If nobody knows, that is your first finding. Our ROI guide explains how to take a two-week baseline.
What design changes remove workslop?
Each change closes one of the leaks. None of them requires banning anything.
Write “done” down before anyone prompts. One page per output: what it must contain, which sources it must cite, what it must never say, and who approves it. That page becomes the model’s instructions, the author’s checklist and the approver’s standard. Without it, every check is a matter of taste.
Bring the context to the model, not to the person. Connect the workflow to where the facts live, such as the CRM, the document system, the price list or the SOP library, read-only where possible, so every draft starts from the current version. This targets the largest single bucket in Glean’s data and the most common source of wrong-version errors.
Let rules do the mechanical checks. Totals match the source, required sections exist, every claim carries a citation, names and dates match the file. Rules do this in seconds and do not get tired.
Check once, at a designed approval point, with sources shown. One named person approves, on a screen that shows each claim next to its source and highlights what changed. Everyone downstream relies on that approval instead of redoing it. The Work Management Institute makes a similar case for treating review as a designed stage with an owner. Where to place that point, and how to keep reviewers from rubber-stamping, is in our guide to human in the loop approval points.
Make the sender own the output. One rule for everyone: nothing AI-assisted leaves your desk unless you would defend it in your own name. With the approval point in place, this closes the forward leak.
Measure the workflow, not the usage. Track cycle time, downstream rework and edits at the approval point. Prompts sent and licenses activated tell you nothing about whether the work got better.
Worked example: a weekly retail report
This example is hypothetical. A 60-person consumer-health brand produces a weekly sell-through report for its retail accounts. Before AI, it took about 150 minutes across three people. With a general chatbot, it looks like this:
Step | Minutes |
|---|---|
Analyst exports data and pastes it in with last week’s report | 25 |
Chatbot drafts the report | 5 |
Analyst checks the draft | 20 |
Head of sales rechecks the numbers and sends questions | 30 |
Finance rechecks the margins | 20 |
Fixes and a second round | 25 |
Total | 125 |
On paper, AI saved 25 minutes. In practice, 50 of the 125 minutes went into two rechecks by more senior, more expensive people, and another 25 into fixes.
After the redesign, the draft is built from the sales system against a written definition of done. A rule check confirms every figure against its source row. The analyst reviews the narrative for 15 minutes, and the head of sales approves once, in about 15 minutes, on a screen that links each number to its source. Finance reads the approved version and no longer rechecks it. Total: about 35 minutes. Same model, different workflow.
Where to start
This week, run the audit on three outputs. Next week, take the output with the most hands on it and the largest downstream bucket, write its definition of done, and decide who approves it and what they see on screen. Then log the time again and compare.
Is one report or handoff eating your team’s week? Book a fit check. We’ll tell you whether it’s worth a diagnostic. To put numbers on the result, read how to measure AI ROI. To place the approval point, read human in the loop: how to design the approval point.
Frequently asked questions
What is AI workslop?
Work produced with AI that looks finished but does not move the task forward, so the person who receives it ends up doing the thinking. Researchers at BetterUp Labs and the Stanford Social Media Lab introduced the term in Harvard Business Review in September 2025.
What is botsitting?
The time people spend making AI usable: feeding it context, supervising its output, debugging its mistakes and cleaning up after it. Glean’s Work AI Institute introduced the term in its Work AI Index 2026. Botsitting is what the user pays. Workslop is what the next person pays when that work is skipped.
How do you stop workslop?
Write down what done means before anyone prompts, and connect the workflow to where the facts live. Let rules do the mechanical checks, and check once, at a designed approval point with the sources shown. Then make the sender own the output, and measure the workflow instead of usage.
Is the answer to workslop less AI?
Usually not. Some botsitting is the job, such as verifying high-stakes outputs. The cause often sits upstream, when every employee prompts a general chatbot on their own. We would drop AI where it adds nothing and redesign the workflows where it does.
How can you find workslop in your own team?
Run a five-day audit on three recurring outputs that someone else depends on. Everyone who touches them logs minutes per instance, including the people who receive them. Then trace three instances of each with the owner, and compare total minutes with how long the output took before AI.
Related reading
Sources
Glean Work AI Institute (Hinds, Hoffman, Baladi, Cao, Lee, Leonardi, Ranganathan, Rhymer, Rogelberg, Sutton), The Work AI Index 2026, June 2026. Survey of 6,000 full-time digital workers in the US, UK and Australia, December 2025 to January 2026.
Niederhoffer, Kellerman, Lee, Liebscher, Rapuano and Hancock, AI-Generated “Workslop” Is Destroying Productivity, Harvard Business Review, September 22, 2025.
BetterUp Labs and Stanford Social Media Lab, Workslop: The Hidden Cost of AI-Generated Busywork. Online survey of 1,150 full-time US desk workers, September 2025.
Holweg and Davenport, Don’t Let AI Slop Muck Up Your Company’s Processes, Harvard Business Review, June 16, 2026.
Workday, New Workday Research: Companies Are Leaving AI Gains on the Table, January 14, 2026. Survey of 3,200 full-time employees, fielded by Hanover Research in November 2025.
Stack Overflow, 2025 Developer Survey: AI, 2025.
KPMG and the University of Melbourne, Trust, attitudes and use of artificial intelligence: A global study 2025, press release, April 2025. Survey of more than 48,000 people in 47 countries.
Lee, Sarkar, Tankelevitch, Drosos, Rintel, Banks and Wilson, The Impact of Generative AI on Critical Thinking, CHI 2025, April 2025.
Faros AI, The AI Productivity Paradox Research Report, analysis as of June 2025. Telemetry from more than 10,000 developers across 1,255 teams.
Humlum and Vestergaard, Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI, NBER Working Paper 33777, May 2025, revised March 2026.
Work Management Institute, What Is Botsitting? The Hidden Labor Behind the AI Productivity Paradox, June 20, 2026, updated July 28, 2026.