Synthetic users will get you to the middle of the market. That’s exactly where you don’t want to be.

Everyone is excited about synthetic users right now.
Feed an AI your customer profile. Ask it anything. Get answers in minutes.
No recruiting. No scheduling. No messy transcripts.

I use these tools on occasion. They’re fast, they scale, and for certain questions, they’re exactly right.

But they have a structural limitation—one that can be the key differentiator when building from 0 to 1.
They are designed to find the pattern.
The common need.
The majority behavior.
The answer that feels safe to build against.

And that’s the problem.
Because the pattern is the least interesting thing your users do.
The products that actually change markets are rarely built from what’s common.
They’re built from the edge— the offhand comment, the contradiction, the behavior that doesn’t quite make sense.
The thing that almost gets dismissed as noise.

Edge-Hunting in Practice
I’m currently working on an AI platform for a highly technical audience.
The obvious assumption—so reasonable no one questioned it—was this:
these users want the best, fastest, most authoritative answer.
Every competitor is building toward that “single source of truth.”
Ask users directly, and many will say they want the same.
But that’s not how they actually work. They don’t trust single answers.
For them, confidence isn’t something you get from one source— it’s something you build through triangulation.
The same conclusion, tested across different inputs, methods, and moments in time. Only then does it feel true enough to act on.

That’s not a preference. It’s an epistemology.
And it completely changed what we’re building. The opportunity isn’t to provide a better answer. It’s to help users orchestrate inputs until confidence converges.
A second insight is reshaping the interface itself.

Another key insight: Most AI tools assume thinking is linear from question → answer → application. But real work—especially deep technical work—is not.
It loops. It revisits. New information reshuffles what came before.
No one said this directly.
It showed up in tangents when people doubled back, compared contradictions, reworked their thinking in real time.

So we’re not designing a chat. We’re designing a workspace—where conversation and a living document evolve together.

Neither of these insights came from the aggregate.
They came from the edge.
And they’re exactly the kind of signals synthetic users are least likely to produce—because they don’t represent the average.