AI Made Your Team Faster. Somebody Else Is Paying For It.

The most expensive document in your business this week probably took four minutes to make.

Someone typed a prompt, got back twelve tidy paragraphs with headings and a summary table, skimmed it, and hit send. It looked finished. It read like work. Then it landed on somebody else’s desk, and that person spent the next hour and a half working out what it actually said, which parts were true, and what on earth they were meant to do with it.

There is a name for this now. Workslop. I think it is the most useful word to come out of the AI conversation in the past year, because it describes something owners can feel in their business and haven’t been able to point at.

What The Word Means

Researchers at BetterUp Labs and Stanford’s Social Media Lab coined it in Harvard Business Review in September 2025. Their definition, roughly: content produced with AI that passes for good work and does nothing to move the task forward.

They surveyed 1,150 full-time desk workers in the US. Forty per cent had received workslop in the previous month. Each incident took close to two hours to sort out. Using the salaries people reported, the researchers put the cost at US$186 a month for each employee on the receiving end, and more than US$9 million a year for an organisation of 10,000.

Run the same maths on a team of twenty and you land around US$18,000 a year. Spent on reading.

I already use “slop” for lazy AI output in general, the stuff that fills your social feed. The “work” on the front is what matters here. This is slop with your logo on it, sent by someone you pay, to someone else you pay.

Where The Saved Time Goes

It is hard to see because the person who sent it really did save time. Ask them and they’ll tell you AI has made them faster, and they are right. Their forty minutes became four.

The other thirty-six went to whoever opened it. With interest.

Workday surveyed 3,200 people at larger organisations and published the results in February 2026. Seventy-seven per cent said AI had made them more productive over the previous year. Then the researchers asked about rework. Roughly 37 per cent of the time saved was going straight back into correcting, clarifying and rewriting what the AI produced. Close to four hours handed back out of every ten gained. Only 14 per cent of people were coming out consistently ahead once that was counted.

A Zapier survey of more than 1,100 AI users at US enterprises found the average person spending 4.5 hours a week cleaning up AI output. Half a day. Every week.

Those are big-company numbers from mostly overseas samples, and I would hold the exact figures loosely. The mechanism is what travels. And I think it bites harder in an SME, because a business of fifteen people has no layer of analysts to absorb the cleanup. The person who catches the vague proposal, the wrong figure in the quote, the client email promising something you don’t do, is usually the most senior and most expensive person in the building. Often that’s you.

So if your team is using AI every day and you have never been busier, this is one place to look. It is also why “time saved” is such a slippery way to measure what AI is returning. You have to count the time at both ends of the handover.

The Bill You Never See

The hours are the smaller cost.

In the BetterUp and Stanford research, 42 per cent of people said they saw the sender as less trustworthy afterwards. About half rated them as less creative, less capable and less reliable. One document did that.

Picture it in a team of twelve where everybody knows everybody. Your bookkeeper gets a waffly three-page “analysis” from the operations manager and quietly decides to double-check everything that comes from him from now on. She says nothing. She just stops trusting his numbers. You now carry a tax on every future exchange between those two people, and it will never show up on a report.

There is a slower cost underneath that one. Somebody who ships the model’s first answer every day stops building their own sense of what good looks like. I wrote about that risk for junior staff, and it applies just as much to someone with twenty years behind them.

Check Your Own Sent Folder

Zety surveyed employees in early 2026, and 55 per cent said they had received workslop from a manager or supervisor.

So before you go hunting for the culprit, have a look at what you have sent lately. The strategy summary you generated on a Sunday night and forwarded to the leadership team with “thoughts?”. The policy you asked a chatbot to draft and never read past page one. The 400-word reply to a question that needed a yes or a no.

When the owner does it, two things happen. Everyone learns that this is what acceptable looks like around here. And nobody pushes back, because it came from you.

I also think a lot of workslop starts with a well-meant instruction. “We need to be using AI more.” No detail on which work, to what standard, checked by whom. People do what they have been asked. They use it more. On everything.

Why I Wouldn’t Ban Anything

The reflex is to clamp down. I get it, and I think it is the wrong move. You would give up gains that are real and substantial, and the usage would simply move to personal accounts where you can’t see it.

Workslop is a thinking problem that turns up wearing a tool. Somebody skipped the step where they decide what they actually think, and the software let them skip it more convincingly than ever. Five years ago the same person sent you a half-baked one-pager and you could tell at a glance it was half-baked. Now it arrives in perfect paragraphs with a table.

For me the whole value of this technology is amplifying intelligence, taking thinking a person has already done and extending how far it reaches. Where no thinking went in, the only thing that gets amplified is the word count.

Six Rules That Fit On A Page

• Whoever sends it owns it. “The AI wrote that” is off the table as an explanation. If your name is on it, you have read every line and you would defend it in a meeting.
• Open with your own view. Anything AI-assisted starts with two sentences in your own words. What I think. What I need from you. If you can’t write those two sentences, the document isn’t ready to send.
• Say what the AI did. No shame attached. “Drafted with AI, figures checked against Xero by me” tells the reader exactly how much weight to put on it.
• Shorter is the default. AI makes length free for the writer and expensive for the reader. Ask for the half-page version, then cut it again.
• The receiver can send it back. Give everyone that permission out loud, including permission to send it back to you. “I can’t tell what you’re asking me to do” is a complete and acceptable reply.
• Train on the actual job. In the Zapier survey, 94 per cent of people who had been trained said AI helped their productivity, against 69 per cent of those who had not. A generic prompting workshop moves that very little. An hour on how your estimators should use AI to build a quote, with a real job on the screen, moves it a lot. That is capability, and it is built task by task.

If you already have an AI policy, add these to it. A typical policy covers what staff may paste into a chatbot and stays silent on the standard of what comes out the other side. Working out which tasks suit AI in your business, and what “checked” means for each one, is a big part of what I do in AI consulting work.

One Question For This Week

Find one thing that landed in your inbox recently that read well and told you nothing. Go and see the person who sent it and ask a single question. What do you reckon we should do?

If they answer in thirty seconds with a clear view, good. The thinking was there and the document buried it. Show them rule two.

If they can’t answer, you have found something more valuable. The document was standing in for a decision nobody had made. Sit down and make it together. That conversation will take about an hour, and it is the hour the four-minute document was written to avoid.

Frequently Asked Questions

What is AI workslop?

AI workslop is AI-generated work that looks polished and complete but does nothing to move the task forward. The term was coined by researchers at BetterUp Labs and Stanford’s Social Media Lab in Harvard Business Review in September 2025. Typical examples are long reports with no recommendation, summaries that restate the question, and emails that read well and leave the reader unsure what they are being asked to do. The sender saves time and the receiver spends it working out what the document means.

How much does workslop cost a business?

The BetterUp and Stanford research found each incident took close to two hours to resolve and estimated the cost at US$186 a month for each affected employee, or more than US$9 million a year for a 10,000-person organisation. On the same maths a team of twenty loses roughly US$18,000 a year. Workday’s 2026 research found around 37 per cent of the time people save with AI goes back into correcting and rewriting its output. In a small business the cleanup usually lands on the owner or the most senior people, so the real cost per hour is higher.

How do I stop workslop in my team?

Set a small number of clear expectations. Whoever sends a piece of work owns every line of it. Anything AI-assisted opens with two sentences in the sender’s own words covering what they think and what they need. People say what the AI did and what they checked. Shorter is the default. Anyone can send unclear work back, including to the owner. And train people on the real tasks in their role, because trained staff get far better results from AI than untrained staff.

Should I ban AI tools to prevent workslop?

No. A ban gives up the genuine productivity gains and usually pushes AI use out of sight onto personal accounts, where you have less control over quality and data. Workslop comes from people skipping the thinking and sending the first output unchecked. Clear standards for what gets sent, a culture where unclear work can be returned, and role-specific training fix the cause without losing the benefit. More on how I approach that is on the philosophy page.

How is workslop different from an AI hallucination?

A hallucination is a factual error produced by the model, such as an invented figure or a citation that does not exist. Workslop is a human decision to pass on AI output without adding judgement or checking it. Workslop can be entirely accurate and still useless, because it has no point of view and no clear ask. Hallucinations are caught by verification. Workslop is prevented by expecting the sender to think before they send.