You open the deck from a colleague and it looks great. Clean headings and a confident summary up top. Twenty minutes later you realise it doesn't actually say anything, and now it's your job to work out what they meant.
That experience has a name now. Researchers at BetterUp Labs and Stanford's Social Media Lab call it workslop: AI-generated work that looks finished but lacks the substance to move the task forward. They surveyed 1,150 US desk workers. Forty percent had received some in the last month. Fifty-three percent admitted sending it at least some of the time.
I'd bet your numbers look similar. Most leadership teams I talk to rolled out a chatbot licence, sent an email telling people to use it, and waited for the productivity to show up. What showed up instead was more stuff. More pages and more polished drafts that someone downstream has to fix.
The cost lands on the person who opens it
Here's the part that should bother a CFO. The researchers estimate 15.4% of all the work people receive is workslop. Each incident costs the recipient about an hour and 56 minutes to deal with. That's the time spent decoding it, checking it, going back to the sender or quietly redoing it yourself.
They put a dollar figure on it too. Each incident costs $186 a month per affected employee. Scale that across a 10,000-person organisation at a 40% hit rate and you're looking at around $9 million a year in lost productivity. They call that a floor. It doesn't count the damage to trust when you start assuming everything your colleague sends you came out of a prompt.
That last bit matters more than the money. Once people suspect a teammate is outsourcing their thinking, they stop reading that person's work carefully. Or they read it twice. Either way the team slows down, and nobody says why.
It's a leadership problem
The easy move is to blame the sender. They're lazy, they cut corners, they should know better. The research points somewhere less comfortable. Workslop thrives when leaders mandate AI use without saying how or why, and then expect more output because the tools are now in place.
Think about what that message sounds like from the desk. Use AI. We've paid for it. You should be faster now. With no guidance on where it helps and where it hurts, people do the rational thing. They use it everywhere, they produce volume, and they hit send.
Give people a new capability without changing how work gets reviewed, and they'll use it to do the old job faster, including the bad parts.
Two things that cut it
The useful finding in the research is that workslop isn't evenly spread. Two factors predicted a lot less of it.
The first is team trust. Where people felt safe admitting they'd used AI, asking for feedback and raising quality concerns, workslop dropped by 61%. That makes sense. If you can say "I drafted this with AI, can you check the numbers in section two?", the recipient knows where to look. If you have to hide it, you pass the problem along and hope nobody notices.
The second is what the researchers call an agentic mindset. People who felt confident and in control of the tools, the ones they describe as pilots rather than passengers, were half as likely to send low-quality AI work. They used it on purpose. They knew what they wanted before they asked.
Neither of those comes from a licence. Both come from how you train people and how you run the team.
What I'd change on Monday
Start by naming it. Tell your team the word, show them an example and admit you've probably sent some yourself. Once there's a shared word for it, people can call it out without it turning into an accusation.
Then make AI use visible. A one-line note at the top of a draft saying what the tool did and what the human checked takes ten seconds. It tells the reader where the risk sits. It also nudges the sender to actually check something before they claim they did.
Next, change the ask. Stop measuring how much people use AI. Usage dashboards reward volume, and volume is exactly the problem. Ask instead what a piece of work let someone decide or do. A two-paragraph answer that unblocks a decision beats a twelve-page AI summary nobody finishes.
Then fix the review. Most teams review AI-assisted work the same way they reviewed human work, which assumes the writer understood what they wrote. That assumption doesn't hold anymore. Reviewers need to ask the sender to explain the key claim in their own words. If they can't, it goes back.
Training that actually reduces it
This is where most AI training goes wrong. The typical session is a tour of features and a prompt library. People leave able to generate more, faster. That's how you train people to send more workslop.
The training that helps works on the two things the research flagged. It builds the confidence to use the tool deliberately on a real task, with the person deciding what good looks like before they start. And it gives the team a shared way to talk about AI use, so admitting you used it is normal and asking for a second look isn't a confession.
In practice that means people bring live work into the room. A board paper they're drafting, or the monthly report they summarise for their boss. They run it with the tool, then they pull the output apart with a colleague. Where did it invent something? Where did it sound right and say nothing? What would they have to check before sending it to their boss?
That last question is the one that sticks. Once someone has watched a confident AI paragraph fall apart under a two-minute check, they don't forget it. They also get faster at spotting the good bits, which is the whole point.
It's also why I'd pick role-based AI training over all-hands demos every time. Our AI training workshop is built around that kind of practice. Teams work on their own material, and they leave with a short set of team rules covering where AI fits in their work and how they'll flag it when they use it.
The manager's part
Managers set the tone here more than any policy does, which is why AI training for managers deserves its own session. If the manager forwards AI summaries without reading them, the team learns that's fine. If the manager asks "what did you check?" every time, the team learns that too.
Watch what you reward. If the person who sends the longest update gets the praise, you'll get longer updates. If the person who flags a problem early gets the praise, even when their note is three lines, you'll get more of those.
And protect the time. A lot of workslop happens because people are overloaded and the tool offers a way to look done. If you've used AI as an excuse to pile more on each person, you've built the conditions the research describes. Taking something off someone's plate will do more for quality than another prompt template.
Watch the handoffs
Workslop does the most damage where work changes hands. A summary goes from analyst to manager, then manager to executive, and each person adds a layer of polish without checking what's underneath. By the time it reaches the person making the call, it reads beautifully and rests on nothing anyone verified.
Map one of those chains in your own team. Pick a regular report and follow it from first draft to final reader. Ask who checked the substance and who just reformatted it. Most teams find at least one step where the honest answer is that nobody looked closely.
Fix that step first. Make one named person responsible for checking the substance before it moves up, and give them the time to do it. Then tell the people downstream that the check exists, so they stop redoing it out of suspicion. That one change can cut a surprising amount of rework, because it puts the effort where the risk actually sits.
Where this leaves you
None of this means AI at work is a bust. The same research found that the confident, deliberate users sent far less workslop. The gap is between handing people a tool and teaching them to use it well, inside a team where it's safe to be honest about how the work got made.
So count the workslop before you count the hours saved. Ask your team how much of what lands in their inbox looks finished and isn't. If the answer is anywhere near the research, you've found the first thing to fix. Start with the people and the review, then worry about the next tool.
