BetterUp Labs and Stanford’s Social Media Lab put a number on something most operations leaders already suspect: AI-generated work that looks done but isn’t costs about $186 per employee per month to clean up. At a 10,000-person company, that is over $9 million a year, and it never shows up as a line item.

Shopify’s CEO gave that cost a name on September 15, on The Knowledge Project podcast with Shane Parrish. Tobi Lütke said his own employees throw “slop grenades” at each other. His words: “So we call those slop grenades that people toss at each other.” AI output someone generated, approved without a close read, and passed along. The person on the other end has to review it for real.

The CEO who mandated AI now has a name for the backlash

Eighteen months earlier, on April 7, 2025, this same Lütke posted a memo declaring that “reflexive AI usage is now a baseline expectation” at Shopify. The memo tied a headcount rule to it: before a team could ask for more people, it had to show AI could not get the job done. AI use went into performance reviews too.

Mandating AI use and assigning who reviews its output are two different decisions. Shopify made the first one a year and a half ago. It is inventing the second one live, with a fresh name for the symptom.

What a slop grenade actually looks like

Lütke gave two concrete examples on the podcast. First: someone asks an agent for a code change, approves the pull request without reading it closely, and whoever is next in line has to actually review it. Second: someone uses a model to turn a short point into a long email, and the recipient uses another model to shrink it back down. Nobody read anything in between.

Notice what both examples have in common. The AI step is fast and cheap. The review step is neither, and in both examples it happens twice: once badly, by the sender skimming instead of reading, and once for real, by whoever receives it. Speed at the front of the pipe does not disappear. It moves to the back and gets more expensive on the way.

The term itself is not Lütke’s. Harry Brundage coined it, and Lütke gave him full credit on X after the episode aired, saying he hoped it would catch on. We already covered how Shopify put Anthropic’s agents to work in real production, with real permissions on real systems. The same speed that made that possible is what now produces more output than a human can review in time.

Two audiences, one root cause

Back in August we wrote about the button LinkedIn gave every reader to flag AI slop. There, the reviewer is a stranger scrolling a feed. At Shopify, the reviewer is a coworker with a full inbox. Different audiences, same mechanism underneath: nobody decided who checks the output before it moves, so whoever receives it checks it, at the worst possible moment.

We named this pattern before, on the taste side of the same problem: when nobody exercises judgment over what AI hands them, the debt does not disappear, it moves downstream to whoever inherits the work. Workslop is the same debt, measured in hours instead of quality. In both cases the fix nobody wants to hear is the same one: somebody has to own the review, by name, before the work ships, not after a reader or a coworker flags it.

Assign the reviewer before the first pull request ships

Shopify’s 2025 memo mandated AI use. No public part of it assigns who reviews what that AI produces before it advances. A year and a half later, the company has a name for the consequence and is still solving it case by case, running on human working memory instead of a rule.

That is what we settle before a company turns AI on, not after. In the AI Maestro discovery program, we map every process in your operation before a single tool gets installed: which steps would use AI, what comes out of each one, and who has to sign off before that output moves forward. Review gets assigned by process, with an owner and an explicit standard, instead of a reflex everyone invents on their own the day the first grenade lands on their desk.

Let’s map who reviews each process before you install AI