I set out to make one part of one of our products cheaper to build. I thought I had a fully formed plan. I had clear goals (same function, lower cost, easier to adjust) and even a back-of-the-envelope understanding of what the mechanism would look like. I fired up an AI session to turn it into a version I could take to vendors.
But then I did something I've only recently learned to do, and it changed what was really just an idea into a plan. Instead of just asking the AI agent to write up my concept, I asked it to interrogate me about it. Flesh out the idea. Find the angles I hadn't considered. Challenge the framing before I committed a second of anyone else's time to it.
It dragged two things into the light that I had genuinely stopped seeing.
The first: the part I wanted to cheapen is a consumable. Customers replace it themselves, they like that they can (the part’s a lot cheaper than the overall product), and those replacements are a real revenue stream for us. I knew all of that. I’d known it for years. But my idea ignored it completely, because I was fixated on just the three goals I’d outlined: function, cost, and adjustment.
The second: the part comes from a different vendor than the product it goes into. Changing how it works would change how two suppliers’ components work, not just one. And it would change how the thing gets installed on a brand-new unit before it ships, installation that our small team does in-house. That's not a footnote. That's the kind of detail that turns on-paper cost savings into real red ink that only shows up after the design is fully locked and there’s really no turning back.
Neither of these is clever. That's the point. They were invisible to me precisely because I'd been close to them for two and a half years. An outside set of eyes would have asked about both in the first hour. The AI tool, with no history with the part and no investment in my mental model of it, was a serviceable (and cheaper) stand-in for that outsider.
Which is the whole idea, pointed in the opposite direction.
Last week, I wrote about a sensor project that went weeks down the wrong road because I trusted an AI tool’s answer before I'd made it prove it understood my product. The lesson there was to interrogate the tool before you depend on it. This is the mirror image: also make it interrogate you. In one direction you're checking whether it understood your world. In the other you're using it to check whether you still see your own. Your accumulated context is both your expertise and your blind spot, and the second one grows quietly.
Same instinct both times: the useful catch comes from a fresh perspective, grounded in your reality but with nothing invested in your current approach. And the honest truth is this is the move that would have caught the sensor detour on day one. "What are you assuming about the machine?" was always available to ask. I just hadn't learned to ask for the interrogation.
Now the honest part: It's slow. A lot of what it asks isn't going to get you anywhere directly, and you’ll rehash the obvious. It’s just the model clarifying its way up the learning curve of a business it's never seen. You'll answer half a dozen dull questions to get to the two that matter. Context is still fleeting; even with AI tools’ memory improving, you find yourself re-briefing across sessions, which is a pain. And when it's done with the interview, you don't have a validated answer. You have a better-formed set of questions and a richer plan.
Both of my projects that ran through this are still works in progress, so I don’t have the end of this story just yet. But what I do know is that while this approach isn’t a fast-track to a result, it helped me see the forest while I was fixated on a tree. Arguably the tallest, prettiest, most central, forest-defining tree, but a single tree all the same.
Worth it. Easily.
So, here's the play you can run this week, on any decision you've been circling. Don't just come out and ask the AI tool for a plan; that cedes too much of the thinking. Bring it your idea and tell it to interrogate you: flesh this out, find the angles I'm missing, challenge what I'm assuming, one question at a time. Prime it with your context and accept that you'll re-brief. Then do the part it can't: prune what it comes up with. Much of what surfaces won't matter. Catching the two or three things that do is still your job, and it always will be. Expect to tweak the output. You're not getting a shortcut to the finish, but you're getting a much better place to start.
I run this on everything now, from vendor decisions to a side project that has nothing to do with the business. I use Matt Pocock’s grill-me skill, but others are available or you can write your own. Just remember the slowdown is the whole point. It's the friction that pries you off the one tree.
— Zain
P.S. What's a decision you're circling right now where you already suspect you're staring at just one tree? Hit reply. I read every one.