Experience is two skills now, priced in opposite directions

You’ve been told the same thing all year: your experience is the part AI can’t take. The models write the code now, but they don’t know which problem is worth solving, which feature moves the business, which complaint is a symptom rather than the disease. You’ve spent a decade learning that. So your judgment is the moat, and every year you’ve banked makes you safer.

Marty Cagan, who founded Silicon Valley Product Group and is the most-read writer in product management, spent this summer arguing both halves of that sentence. They don’t agree.

The background, quickly, for anyone who hasn’t been following it. As AI collapses the cost of building software, teams are splitting into different operating models, meaning different ways a company actually turns an idea into shipped product. One is the discovery-heavy model, where the slow, careful part is figuring out what to build. Another is what Cagan calls the AI-native model, where AI does most of the building and the whole cycle is nearly instant. In two SVPG essays this summer, “AI Product Management 2 Years In” and “A Fresh Definition of The Product Role,” Cagan argues those models don’t just prefer different PM skills. They select for opposite ones. And the skills that make you excellent in the first can make you fail in the second.

That’s a strange thing to hear from the same person whose discovery-bottleneck argument is the strongest case going for experience as a moat. Read one Cagan and your accumulated PM discipline compounds. Read the other and it’s the thing sinking you.

Both are right, because “experience” is two things wearing one word, and a build cost falling toward zero prices them in opposite directions.

Pull the two claims apart and they stop fighting. One is about domain judgment: knowing what’s worth building in a given market. The other is about process discipline: how much you specify before you build, how much consensus you gather first, how much rigor you front-load. For the whole history of the profession these arrived fused, because you learned them in the same job, on the same clock. “Ten years of experience” was one credential.

They have opposite exposure to a build cost falling toward zero. Domain judgment doesn’t care what building costs. Knowing which problem matters is worth the same whether the build takes six months or six minutes. Process discipline is priced entirely off that cost. Careful upfront specification is smart insurance when a wrong build burns a quarter. It’s dead weight when a wrong build costs an afternoon.

So the exact same habit, “I always align everyone thoroughly before we commit to build,” was an asset under the old cost structure and is a liability under the new one. Nothing about the PM changed. The price of being wrong changed, and it repriced half of their experience without touching the other half. Aakash G., who writes one of the most-read PM newsletters, puts numbers on it: pre-build alignment work that ran 40 to 60% of a PM’s time in traditional teams compresses to under 15% in AI-native ones, because implementation is nearly instant. The senior PM still burning 40% of the week on that alignment has become the bottleneck, even though it still feels like diligence. Aakash’s word for what happens to them is blunt. They “become too slow,” not through any drop in ability, but through disciplined execution of the wrong model.

If that were the whole story, the advice would be easy. Keep the judgment, drop the pacing. But the two don’t come apart that cleanly, and this is the part that should worry anyone with a long résumé.

Look at what discovery judgment actually is when you do it. It’s workshops and validation cycles, the deliberate front-loading of rigor. That is process discipline. The thing Cagan calls the durable moat and the thing Aakash calls the liability aren’t neighbors. At the level of what you do on a Tuesday, they’re the same activity. What separates them is when you stop. And knowing when enough validation is enough is exactly the operating-model-specific instinct that doesn’t carry across.

So the clean split has a seam that fails under load. Knowing what to build and knowing how much to validate before building it aren’t two skills you can hold in separate hands. The second is how the first gets exercised. You can’t run the judgment without running some quantity of the process, and the quantity that made you great in a discovery-heavy org is the quantity that makes you too slow in an AI-native one. The moat and the liability are one muscle, calibrated to a build cost that no longer exists. That’s a much worse problem than “learn a new skill.” The instinct that fires your best judgment also fires your worst pacing, and you can’t keep one without retraining the other.

The case for experience-as-moat has a real number behind it. Anthropic, the AI lab behind Claude, studied 400,000 sessions and found expert users taking 12 actions per instruction against 5 for novices, evidence that experience genuinely multiplies what you get out of an AI. I don’t think that’s wrong. But it measures that multiplier inside one operating model. It says nothing about what happens when the model itself changes underneath a person, which is the case Cagan and Aakash are describing.

Which leaves a question sharp enough to matter to anyone hiring, or being hired. The moat argument says hire for accumulated PM experience. The inversion argument says experience from the wrong operating model is precisely what sinks the hire, and Cagan’s own point is that a normal interview can’t see the difference. A discovery-model PM and an AI-native PM score about the same on generic questions like “tell me about a time you influenced without authority.” So the durable moat and the disqualifying liability show up identically at the moment you decide.

What question tells “ten years that compound” apart from “ten years that have to be unlearned,” when the only instrument you’ve got measures neither? I don’t have it yet. But I’m fairly sure it isn’t on any interview loop running today, and the people most exposed to getting it wrong are the ones with the longest track records.