Tobias Lütke, the CEO of Shopify, spent the last two years telling his staff that using AI at work was not a suggestion. Before you ask for more headcount, he said, prove that AI genuinely cannot do the job. That is about as strong an AI mandate as a real company has made public. So it was worth noticing when, on a podcast released this September, he described what actually happened inside his own company. He used a phrase his staff had coined for it: slop grenades.
What Lütke actually found
A slop grenade, in Lütke's telling, is a piece of AI-generated work, an email or a report, handed to a colleague who never really read it before sending it on. "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," he said. The person who generated the text got the benefit: it took them five minutes instead of thirty. The cost did not disappear. It moved to whoever opened it next.
This is not a story about AI writing being bad. Modern AI is often very good at producing something that reads fine on a first pass. The problem shows up after: nobody had decided who was responsible for catching the parts that were not actually right, so nobody did, until it landed on a desk it should not have.
Not just a Shopify problem
A 2025 study by BetterUp and Stanford's Social Media Lab, surveying more than a thousand full-time desk workers, gave this pattern a name: workslop, AI-generated content that looks like real work but does not hold up once someone tries to use it. Forty percent of workers said they had received some in the last month. On average, people estimated that around 15% of everything landing in their inbox was workslop, costing them close to three and a half hours a month just to fix. None of that shows up as a line item anywhere. It quietly taxes whoever is downstream.
Here is the actual mechanism, in plain terms: AI made producing a first draft nearly free. It did nothing to make checking a first draft nearly free. Those two used to cost roughly the same, so nobody had to think about which one they were paying for. Now they are wildly different, and if a business does not decide on purpose who does the checking and when, the checking cost does not vanish. It gets handed to whoever receives the work next, without anyone agreeing to that deal.
What this looks like at your size
You do not need two thousand employees for this to bite. Picture a small architecture studio: one architect, one junior assistant. The assistant starts using AI to draft the weekly project-update email to clients, cutting a task that used to take forty minutes down to five. For a while, that looks like a clear win. Then the architect starts noticing small errors slipping into client mail: a materials cost quoted from an earlier draft, a deadline that no longer matches the schedule. Now the architect has to read every one of those emails in full before it goes out, which takes almost as long as writing them from scratch used to. The studio did not get faster. It moved the same thirty minutes from the assistant's day to the architect's, and hid it there.
Honest limits
None of this is an argument against using AI for drafts, reports, or client updates. It is an argument against skipping the one decision that actually matters: for anything AI drafts that is going to reach someone outside your own head, who checks it, and how carefully, before it goes? If the honest answer is "nobody, it just goes," that is a slop grenade waiting for a pin. If the task is genuinely low stakes, a note only you will ever read, a draft you are rewriting anyway, none of this applies; the cost only appears the moment work reaches another person uninspected. We wrote about drawing that same line for AI agents that act on your behalf in Do I Really Have to Check Everything My AI Does?; the same two questions (is it cheap to undo, does it touch someone else's money, safety, or reputation before a person sees it) work just as well for AI-drafted text a person is about to send.
One thing to do this week
Before you tell anyone on your team, including yourself, to just use AI for a task, say out loud who reads the result before it leaves the building, and how long that is supposed to take. If you cannot answer that in one sentence, you have not actually saved any time. You have moved it, and hidden where it went.