What You’ll Find This Week

In April 2025, Shopify CEO Tobi Lütke told every employee that using AI was "a baseline expectation," and that anyone asking for more headcount or resources had to justify first why AI couldn't do the job. This month, the same CEO lamented that his employees were tossing AI "slop grenades" at each other.

Lütke hasn't changed his mind about AI. He's describing a cost his own mandate never planned for: work that took seconds to generate and hours for someone else to read.

This week: why AI made writing nearly free and reading more expensive than it's ever been, who ends up paying for it, and why company-wide AI mandates keep producing more of it.

Don’t Miss Our Latest Podcast

This Week’s Article

In April 2025, Tobi Lütke sent Shopify's employees a memo that got quoted in nearly every conversation about enterprise AI for months. "Reflexive AI usage is now a baseline expectation at Shopify," he wrote. Teams asking for more headcount or resources would first have to justify why AI couldn't do the work, and AI competency would become part of performance reviews and hiring.

In mid-September, Lütke sat down with Shane Parrish on The Knowledge Project, and described what some of that usage looks like from inside the company. "We call those 'slop grenades' that people toss at each other. That's definitely a bad thing," he said. "You don't really read it, and now it has to be reviewed by your colleagues, and they are like, 'this doesn't look right'. You're just letting AI do the work for you."

Lütke isn't walking back the AI mandate and nothing in the interview suggests Shopify plans to use less AI. He's naming a cost the mandate never accounted for. And he isn't the only executive naming it.

Ryan Roslansky runs Office and LinkedIn at Microsoft, which puts him over Word, PowerPoint, and Outlook, the apps where Microsoft's Copilot assistant drafts the documents, decks, and emails people send each other. On August 31, he posted on LinkedIn about a document that landed on his desk. "Someone sent me what was clearly an AI generated doc. I read it, realized it was AI, had my AI summarize it, and ended up with nothing that brought new thinking to the idea. This is a doom loop."

One of these executives mandated AI across his whole company. The other runs the software that produced the document he was complaining about. Both are describing the same habit: people sending AI output they haven't read themselves, to colleagues who now have to.

Writing used to be expensive. A two-page memo could cost its author an afternoon, and that afternoon did a job nobody tracked. It forced the author to work out what they actually thought before anyone else had to read it. Length was a rough signal too. A reader could assume a long memo had cost somebody something.

AI took the cost out of writing and made reading more expensive at the same time, because AI output runs long. Lütke's own example on the podcast was email. AI is supposed to help synthesize the points in an email, he said, rather than turn it into "a big missive" that wastes everyone's time. So the work didn't go away. It got bigger, and it moved from the person who sent the document to the person who received it.

Perfect example: a few days ago I asked Claude a yes/no question. Claude responded with 905 words. (Fortunately, I had the wherewithal not to forward that response along to colleagues.)

BetterUp, a workplace coaching company, and Stanford's Social Media Lab surveyed 1,150 full-time US desk workers in September 2025 and gave the problem a name. "Workslop" is AI-generated work that looks polished and complete and doesn't hold up once someone tries to use it. Forty percent of the workers surveyed had received workslop in the previous month. Each instance took about an hour and 51 minutes to deal with, roughly 20 minutes longer than doing the work by hand. Researchers put the cost at $186 per employee per month, or more than $9 million a year (for a company of 10,000 people).

None of those hours show up where companies measure AI, and the person doing the cleanup pays for them twice. The sender gets credit for an afternoon saved. The recipient spends two hours fixing someone else's work, and at a company that grades people on how much they use AI, those are two hours of their own work that didn't get done. On paper, the colleague who threw the grenade looks more productive than the one who cleaned up after it. Leadership sees AI adoption climbing and never sees the hours people spend repairing AI output.

The damage doesn't stop at time. About half of the people who received workslop said they now saw the sender as less creative, capable, and reliable. Forty-two percent saw them as less trustworthy, and nearly a third said they were less likely to want to work with that person again. The sender saves an afternoon and pays for it with their colleagues' trust.

A handful of careless employees can't produce numbers this large. In the same survey, 53% of workers admitted to sending workslop themselves.

BetterUp's follow-up work, published in August, traces where it comes from. Workslop spreads when leadership mandates AI faster than it provides training and support, and employees start turning in work whose main purpose is to check the AI box. The strongest predictor of whether AI investment actually improved performance was employees' trust in their leaders. Adoption rates, the number most AI rollouts are managed against, predicted less.

Follow that through. A company tells employees AI use is expected and makes AI competency part of their reviews. The easiest evidence of AI use for a manager to see is output: the drafted deck, the generated report, the three-paragraph reply to a one-line question. Nothing in that loop asks whether the person on the receiving end read the deck, acted on the report, or had to rewrite the reply. A mandate judged by how much people use AI ends up rewarding the exact behavior Lütke described on the podcast.

Meta has gone further than Shopify. This year it started grading employees on "AI-driven impact" in performance reviews, and it runs dashboards tracking AI adoption across teams. In April, a Meta employee built an internal leaderboard called Claudeonomics that ranked more than 85,000 coworkers by how many AI tokens they used and handed out titles like "Token Legend" and "Cache Wizard." The top user averaged 281 billion tokens over 30 days, and some employees put AI agents to work for hours just to push their numbers up. The leaderboard came down two days after the story broke. An agent left running for hours produces a bigger number and nothing anyone will read, which is exactly what a usage metric rewards.

I've written about AI adoption twice already. In June, If AI Isn't Tearing Your Company Apart, That's a Problem argued that visible AI activity often stands in for the organizational change it's supposed to produce. In August, AI is Becoming Commoditized argued that companies report adoption numbers as if they were returns. Workslop adds a harder problem to both. The activity carries a cost of its own, and that cost gets charged to someone other than the person who created it.

Lütke and Roslansky both spotted the problem from the receiving end, which is the only place it shows up. The person who sends workslop rarely experiences it as a problem. The person who reads it does, for about two hours at a time.

The fix has to start on the receiving end as well. If you run a team under an AI mandate, stop asking how much your people use AI and start asking what happens to what they send. Did the recipient act on the document? Answer the email? Rewrite the deck from scratch? When was the last time someone on your team sent AI-generated work back as unusable, and would they feel safe doing it?

Lütke's memo made employees justify why AI couldn't do a job before they asked for help. The next memo should make them show that someone read what AI did.