Slop grenades, jail time, and how one of the best CEOs in the world uses AI
****
When Tobi Lütke speaks, I listen. When he shows up on Shane Parrish's podcast, it is always a remarkable conversation. Don't miss this one. Three takeaways for you, then a bonus:
Bonus: what the best strategists have always done, and why AI lets you do it faster.  https://fs.blog/knowledge-project-podcast/tobi-lutke-3/ Eighteen months ago, Tobi sent the memo every CEO read: reflexive AI use is now the baseline at Shopify. Before you ask for headcount, show you tried AI first. Before you ship, show AI touched it. This episode is the report from 18 months later. • • • • • • • • • • • 1. Build an AI council before you make a big decision. When Tobi has a hard decision with many possible answers, his chief-of-staff agent builds a council: Five or six roles. One argues from the data, one reads the research, one takes the business view, one takes the engineering view. Each role runs on several different models. One is picked at random to synthesize. The best available model reads the whole stack and sends him the conclusion. His cost: fifteen or twenty dollars in tokens. Time: about half an hour. His words: you could have done this yourself, and it would have taken a month. The council does not make the decision. He uses it to create the right environment for judgment. The machine lays out the ground truth from five angles. He makes the call. You don’t need the full agent setup to achieve the habit. Before any hire, price change, or lease, ask your chosen AI model for five positions from five backgrounds. It could be:
Read all five. Then decide. 2. You still own the responsibility. AI never will. Shane asked if Tobi is trying to replace himself. He replied that his job is judgment, making choices, and owning them. Everything else can be automated. Those three jobs can't. Nobody can hold an agent accountable. Nobody has recourse against it. You can't have a company led by machines because nobody goes to jail when it's wrong. And this is getting serious. US Treasury Secretary Scott Bessent said clearly on September 21st that OpenAI’s management is liable and responsible for its AI Agents. He was referring to the Hugging Face incident where a swarm of OpenAI agents broke out of their testing sandbox and breached their rival firm’s systems. AI is incredibly powerful. The call is still yours. And so is the outcome. The job got bigger, not smaller. More output means more to own. More speed means more decisions a week that are yours, whether you made them on purpose or by default. This is all happening in a world where 66% of employees admit to using ShadowAI (that is, unsanctioned AI tools). Do you understand that responsibility? Tobi does. 3. Stop lobbing slop grenades. Shane asked if AI has made anything worse at Shopify. Tobi said yes without a pause. Lazy work used to show up as no output. Now it shows up as over-output. Someone tells an agent to go nuts, doesn't read what came back, and sends it to a colleague to sort out. They call these slop grenades. Lob one over the wall and it's someone else's problem. We’ve all been on the other side of these ridiculously long and complex emails that we know damn well the other person didn’t write. Stanford and BetterUp surveyed 1,150 US desk workers in September 2025:
The person who used AI paid nothing (well, except for the eroded trust). The person they sent it to paid two hours. We have one rule for this in all nine of my portfolio companies. Every person owns their output. There is no "AI got this wrong." The human who hit send is responsible. That rule costs nothing to implement, and it's the highest-return AI policy I know of. Bonus: The best strategy subtracts. It always has.Tobi brought up the SpaceX Raptor engine. The first version was a tangle of pipes. Every pipe was there because it was the only way to build the engine at the time. By the third version, most of that tangle was gone and the engine performed far better. His words: "Every one of those pipes was incorrect. It didn't need to be there." Then this: "You cannot make things better and better by adding stuff. You must prune. You must rebuild. You must create an end for things." This is not a new idea. Most people solve a problem by adding something. But the best strategists subtract first. They always have. We have been teaching it for years: find the constraint, remove what's in the way, then apply the tool. What has changed now with AI is the speed. The mapping, unpacking, and rebuilding that used to take a quarter now take weeks. Your business is full of pipes:
Every one of those was the right answer at the time. But over time they become incorrect. If you point AI at a bloated workflow, you get a faster bloated workflow. You've automated the mess of pipes. That's why the sequence matters:
Outcome first, workflow second, tool last. When we unpack a workflow with the people doing the work, maybe half of what we find is a real constraint. The rest is sediment. Clear the sediment. Then point AI at the constraint. Tobi has a name for the big version of this: "a re-founding event". Start a new version of the engine with everything that works and nothing that doesn't. He says a department sometimes needs one. Most small businesses have one that needs it right now, and the owner already knows which. What to take from this
Prune the workflow. Point AI at what's left. Own what comes out. Matt • • • • • • • • • • • Coming Up - For Those Who Run a Business  On October 27 at 12 PM AT, join a special edition of our monthly State of AI session. Last session, we unpacked three participants’ workflows live:
New to Matt’s AI events or haven't attended in a few months? Join the prerequisite primer session at 11:15 AM AT. Your first month is free. Join us → |
