The Productivity That Gets Paid For at the Next Desk

Think about the last complicated email you received. One of those long threads, with three people copied and a tangled decision inside. You had two paths. The first was to read it calmly, understand what they were really asking for and reply “agreed, let’s go ahead.” Two minutes of reading, four words of reply. The second was to paste it into an AI, ask for a summary, ask for a reply that sounded up to the task, skim it and send it.

The second path feels more productive. It looks more professional. And it often ends in an exchange of ten messages, three clarifications and a meeting to confirm that everyone agreed from the start.

What sped up wasn’t solving. It was producing.

What workslop is and what it costs

This distinction seems minor and it isn’t, because that’s where much of the money companies think they’re saving with artificial intelligence is going. In September 2025, a team from BetterUp Labs and the Stanford Social Media Lab published a term for this in Harvard Business Review: workslop. They defined it as AI-generated content that looks like good work but lacks the substance to move the task forward. In their survey of 1,150 US workers, 41% had received something like this in the previous month, and each episode took them an average of one hour and 56 minutes to deal with. Calculated on reported salaries, that’s about 186 dollars a month per employee. In an organization of ten thousand people, around nine million dollars a year.

But the number that interests me most in that study isn’t the cost. It’s the social effect. More than half of those who received workslop reported being annoyed, 42% considered the sender less trustworthy and about a third said they’d rather not work with that person again. In other words, the tool we use to look competent is achieving exactly the opposite, and whoever uses it doesn’t find out.

Poorly thought-out work moves

There’s the whole mechanism. Poorly thought-out work doesn’t disappear, it moves. Whoever generates the document saves twenty minutes and whoever receives it loses two hours interpreting, correcting or redoing it. In the company’s accounts, the saving shows up and the cost doesn’t, because it happens at another desk, in another department, in another quarter. It’s a transfer disguised as efficiency.

We’re terrible judges of our own speed

The underlying problem is that we’re terrible judges of our own speed. In July 2025, the independent lab METR ran a controlled experiment with experienced developers working in their own repositories. Before starting, they estimated AI would make them 24% faster. When they finished, they felt they had been 20% faster. The measurement showed they took 19% longer. It’s worth saying that METR itself warned in February 2026 that this result no longer necessarily reflects current tools and that its follow-up study had selection problems. It shouldn’t be used to claim that AI makes us slower. But what does stand, and it’s the uncomfortable part, is that perception and the stopwatch pointed in opposite directions, and nobody in the room noticed.

What the code shows

In code the same thing happens, and it gets recorded. GitClear analyzed 211 million changed lines between 2020 and 2024. Lines associated with refactoring, the work of tidying and reusing what already exists, fell from about 25% in 2021 to under 10% in 2024. Copy-pasted code went up, and 2024 was the first year on record in which copy-paste outnumbered reorganizing. Duplicated blocks multiplied eightfold in a single year. More code was written than ever and less was tidied than ever. Someone is going to pay for that difference, and it won’t be whoever pressed tab.

It’s not a problem of lazy people

It’s tempting to read all this as a problem of lazy people misusing the tool. It isn’t, and that reading is what prevents fixing it. A later study by the same researchers, published in January 2026, found that workslop appears where there are unclear AI mandates and overloaded teams. Translated: when the instruction from above is “use AI” and the metric is adoption, the organization gets exactly what it asked for. People using AI. Nobody asked for understanding, so nobody delivered it.

What your dashboard doesn’t measure

And here’s where the problem stops being technological. If your dashboard shows adoption rate, number of queries or estimated hours saved, you’re measuring the cheap half of the equation. The expensive half is rework, and almost no company measures it, because measuring it would mean asking whoever receives the work how long it took them to make it usable. It’s a simple question and almost nobody asks it.

Producing stopped being the bottleneck. Understanding never stopped being one. The difference is that before, this imbalance went unnoticed, because producing was also expensive and slow, and that slowness forced us to think a little before starting.

So the question that follows, and that I’m leaving open for the next column, is who in your organization today has the explicit job of thinking before someone presses generate.

How much of your last week went into fixing something that arrived polished and poorly thought out, and who did you report that cost to?