The Framework That Changed How I Think About What AI Should Automate

I sat in a talk a few months ago and watched a slide go up that made me put my pen down. Not because it was clever. Because it named something I'd felt but never had language for.

The talk wasn't about AI with AI. It was a framework someone else built, and I want to be honest about that up front: this is a framework from a talk I attended, not something I invented. But it changed how I think about every automation decision I've made since, so I want to walk through it the way it was walked through to me.

The Judgment Pyramid, as it was presented, has three layers.

At the bottom: Hidden Repetition. The drafts nobody sees. The drills. The tenth version of an email before the one you send. Low-stakes reps that never make it into a portfolio.

In the middle: Pattern Exposure. Small decisions with recoverable mistakes. You try something, it doesn't quite work, you notice why, you adjust. Nobody's watching. The cost of being wrong is an afternoon, not a client.

At the top: the thing everyone actually notices. Judgment. Strategy. Intuition. The read on a room, the instinct for which idea will land, the sense of when something is almost right versus actually wrong.

Here's the pivot the talk made, and the one that stopped me: most people only value the top of the pyramid. They see someone with sharp judgment and assume that judgment arrived on its own, or that it's a personality trait. It isn't. It's compounded from the bottom two layers. The hidden reps and the small recoverable mistakes are not the boring prerequisite to good judgment. They are how judgment gets built.

Which means the "easy" work you're tempted to hand off entirely, the drafts, the low-stakes attempts, the stuff that feels like busywork, might be exactly the reps your judgment is standing on. Automate away all the hidden repetition and pattern exposure, and you don't get a shortcut to good judgment. You get a hollowed-out top layer with nothing underneath it.

Where AI actually helps, and the question to ask first

The same talk broke this into five things AI is genuinely good at, each paired with a question to ask before shipping it: does speeding this up remove a rep someone needed, or does it just remove friction that wasn't teaching anyone anything?

The categories to be careful automating first: execution work, creative work, analytical work, expert judgment calls, and managerial decisions. Not because AI can't touch these. Because these are exactly the places where the hidden reps matter most, where the person doing the work is still building the pattern recognition that becomes their judgment later.

The line that's stuck with me since: design for speed without degrading judgment.

That's not an argument against automation. I build AI tools for a living, so it would be a strange place for me to land. It's an argument for noticing which reps you're removing when you speed something up, and being honest about whether that rep was actually building something in you.

This is the whole reason AI with AI Labs teaches hands-on building instead of just prompt tips. A prompt you copy-paste gives you the output without the rep. A build you actually do, even with AI doing a lot of the heavy lifting, still puts you inside the pattern exposure layer. You still have to notice what worked, adjust, try again. That's the part that compounds.

If you're deciding whether to automate something in your own work, ask the question the talk left me with: is this the kind of rep that's building judgment I'll need later, or is it just friction with nothing underneath it? The answer changes what you should hand off, and what you should still be doing yourself, slowly, badly, for a while.