Your AI Doesn't Know What Your Team Wouldn't Ship

I helped start Yahoo! Answers. We brought it back years later. Then I made the call to end it.

It looked like my call. It wasn’t, really. Ending something millions of people still used came out of a read that had built up over time, across a lot of people, about what the thing was worth and what it quietly cost us to keep. I could make the call because I was carrying all of that context at once. Not one of us could have fully explained it.

That’s the part we keep handing to AI as if it were a to-do list.

Your company runs on judgment like that everywhere, and you can’t point to where it lives.

I had the unit wrong

For a few weeks now I’ve been arguing that when AI commoditizes outputs, the thing that compounds is your judgment. I still think that’s right.

I just had the unit wrong. The judgment that compounds most isn’t yours. It’s your team’s.

Every company runs on a body of judgment nobody wrote down. What you build and what you kill. Which bets are worth making here and would be reckless anywhere else. What you’d never ship no matter how good the demo looked. What “good” means in this building, in a way that’s different from the company across the street. It isn’t in a document.

It’s spread across a dozen people who would each describe it differently, and it holds anyway, right up until you try to hand it to a machine.

Blind to the bar

Watch what happens when you drop a capable model into a real team. It’s genuinely good at the work and completely blind to the bar.

It builds the feature three of you already decided, last year, wasn’t worth the maintenance.

It gives you ten options and no opinion about which one your company would put its name on.

It clears every bar except the one that matters, the one only your team can see. Every single output looks correct. And a team that ships all of them, because each one looked fine on its own, is just a faster version of a team with no taste.

What speed reveals

It took me twenty years of building products at scale to see this clearly.

The doing was always the visible part of the job. The building, the shipping, the scaling. The deciding was the part underneath it, the part experience actually bought you.

Knowing what to kill was part of that judgment, not a separate instinct. When the doing was slow, the gap between the two stayed hidden. Everyone was busy, so everyone looked useful. Speed pulls that cover off.

When the work takes a minute, what’s left standing is the judgment, and you can suddenly see who has it and who was just moving quickly.

Here’s why that judgment is worth more than the model, not less. It doesn’t come from a manual. It’s path dependent. It’s the residue of a hundred calls you got wrong, of watching how a decision actually landed

versus how you thought it would, of learning which instincts to trust and which to override.

Which is exactly why a general-purpose model shows up knowing none of it. So does your competitor’s model. The building blocks are the same for everyone now. The only thing that isn’t is the judgment your team earned and theirs didn’t, and that is the one part of the whole stack nobody can copy.

The mistake leaders make

So the mistake I watch leaders make is treating this as a productivity rollout.

Hand everyone a tool, count the hours saved, call it a win. Give a team a pile of AI and you don’t get a smarter team. You get the same team shipping twice as much of the wrong thing.

The actual work is quieter and harder. Name what your organization’s judgment actually is, and figure out where it lives, before you automate around it. Otherwise you scale the output and lose the reason it was ever any good.

I don’t have this fully worked out, and I want to be honest about where it gets hard.

Most of this judgment was never said out loud, because it never had to be. It lived in the room, in who winced when a bad idea came up, in what everyone somehow already knew not to try. The people who hold it are usually too close to it to see it as a thing at all.

And the moment you try to write it down, half the team realizes they would have written something different.

Naming what your organization actually believes is good, out loud, is harder and more contested than any tool rollout, and it’s the part nobody schedules.

I’m still testing my way through it. I don’t have clean answers yet.

Your moat, and your blind spot

But I’m sure about the shape of it. The moat was never the model. It’s the judgment the model is missing, the accumulated sense of what to build and what to kill, what good means in this building and nowhere

else. That’s the part that compounds. That’s the part nobody can copy.

It’s also the part you never wrote down. Every company runs on judgment nobody wrote down. That’s your moat, and your blind spot.

And that was always the real work.

Not building the thing, but knowing, as a team, what deserved to exist at all. AI just made everything around it cheap enough that there’s nowhere left to hide from the one thing it can’t do for you.