AI Didn't Fix Discovery. It Made the Problem Harder to Ignore.

There's a conversation happening in product circles right now about how AI has changed the game. Most of it focuses on delivery: engineers shipping faster, code getting written in minutes, features that used to take sprints now taking days.

And that part is real. Delivery has genuinely been transformed.

But here's what I keep running into: teams that have turbocharged their delivery engine while their discovery process is still running on fumes. They're building faster than ever. They're also finding out faster than ever that they built the wrong thing.

Marty Cagan wrote recently about "build to learn vs build to earn" the distinction between prototyping to discover a solution worth building versus actually building the thing customers run their business on. It's a clean framework, and I think he's right that the real bottleneck has always been discovery. AI just made that truth impossible to hide.

I've been sitting with this firsthand through my work with Billie, a workflow tool I've been advising on. Early in the process, we were describing features in requirements documents. Reasonable approach. Standard practice. And for most of the team, it felt like enough.

Except it wasn't.

The users we were building for already had deeply established workflow patterns. They weren't coming to us as blank slates. And when you put abstract requirements in front of someone who has been doing their job the same way for years, you get polite nods. You get "yeah, that sounds useful." You don't get the truth.

The truth only came out when we put a prototype in front of them.

Not a finished product. Not a polished demo. A rough, functional, fast prototype that AI tooling helped us spin up in a fraction of the time it used to take. Something they could actually touch and click through. And then something interesting happened.

Some features landed exactly as we expected. Users got them immediately. They fit into existing mental models, filled a gap, made sense. Good signal. Worth building.

But the more valuable signal came from the features that didn't fit. The ones where users paused, looked slightly uncomfortable, and said something like "I'm not sure how I'd use this." That friction, that moment of cognitive mismatch wasn't a red flag. It was a signal.

Those were the transformational features. Not because users were wrong, but because those features were asking them to work differently, not just work easier. Incremental features slide into existing workflows. Transformational ones require a workflow to change. And you can't discover that distinction from a requirements doc. You can only discover it from a reaction.

This is where AI has genuinely changed discovery, and it's more subtle than the "ship faster" narrative suggests. The old constraint in prototyping wasn't ideas. It was time and cost. Building enough prototypes to generate real learning meant investing significant engineering or design resources before you had any confidence you were heading in the right direction. So teams made fewer prototypes, ran fewer tests, and relied more heavily on assumptions dressed up as requirements.

AI has collapsed that constraint. We can now prototype quickly enough to put multiple versions, multiple approaches, multiple conceptual bets in front of users in the time it used to take to spec a single one. Discovery can actually be discovery, iterative, experimental, genuinely exploratory, rather than a checkbox before the "real work" begins.

But, you still have to know what you're looking for.

The tool doesn't tell you what the friction means. It doesn't tell you whether the user's confusion signals a bad idea or a breakthrough one. It doesn't tell you whether to kill the feature or redesign the onboarding around it. That judgment still requires product sense. It still requires someone who can sit in that awkward moment when a user doesn't know what to do with your prototype and ask the right follow-up question.

What AI has done is compress the loop between hypothesis and signal. It's shortened the distance between "here's what we think users need" and "here's what users actually do when we put something in front of them." That's enormous. But compression isn't the same as insight.

The teams that will win aren't the ones who use AI to prototype faster. They're the ones who use AI to learn faster and who have the product judgment to know what they've learned.

Discovery is still the hard part. AI just gave us fewer excuses to skip it.