I’ve been watching smart people use AI for two years now. Not casual users—founders, consultants, operators who genuinely understand the tools.

And most of them have nothing to show for it.

Not because they lack skill. The problem is subtler: they’re trapped in a cycle where everything they build almost works.

Here’s the pattern. Someone identifies a problem—email overload, client onboarding, research synthesis. They spend a weekend building an AI solution. It works in the test environment. It’s clever. They show it to a friend.

Then they never use it again.

Why? Because the last 10%—the part that makes a solution actually usable—never gets closed. The app lives in a development environment. The workflow requires three manual steps that kill the time savings. The automation breaks when the input format changes slightly.

Ninety percent of the way there. Zero percent of the value captured.

I see this constantly. People building organization systems they won’t access. Research tools that require more effort to query than to just do the research manually. Client deliverable templates that are impressive in demos and friction-heavy in practice.

The technology worked. The deployment didn’t.

I believe the deeper problem is actually target selection rather than execution.

Most people tinkering with AI are solving problems they could have—not problems they do have. They’re building tools for hypothetical workflows rather than identifying the single constraint that’s actually limiting their income or time.

Ask yourself: What is the ONE bottleneck in your work that, if removed, would directly translate to more revenue or recovered hours?

Not “what would be cool to automate.” Not “what’s an interesting use case.” The thing that’s actually choking your economics right now.

Most people can’t answer this clearly. So they build sideways—impressive capability with no economic target.

A Few Questions Worth Sitting With

I’m not going to give you a framework. But I’ve noticed that people who escape the trap tend to ask themselves different questions:

“If this works perfectly, what specifically changes about my week?” If you can’t answer in concrete terms—dollars, hours, clients—you’re probably building a demo, not a system.

“What’s the minimum version that I would actually use daily?” The 90% Trap is often a scope problem. People build the sophisticated version instead of the ugly-but-deployed version.

“Am I solving a problem I have, or a problem I think I should have?” Most AI content showcases sexy use cases. Sexy isn’t the same as valuable. Your bottleneck might be boring. Automate the boring thing.

“What would make this impossible to ignore?” The solutions that stick aren’t the cleverest—they’re the ones that insert themselves into your existing workflow so seamlessly you can’t avoid using them.

There’s another version of this that has nothing to do with work.

“What am I paying for right now that I could build—and customize—for free?”

A mom built a custom streaming app for her kids. No algorithm pushing questionable content. No ads. Just the shows she approves, organized the way she wants. She canceled YouTube Premium. The app took a weekend.

It’s not a business workflow. It’s just solving an actual problem she had, with a tool that now exists.

The 90% Trap happens less often when the use case is personal and specific. You know immediately whether it works because you’re the user. There’s no hypothetical ROI to calculate—either you use it or you don’t.

Sometimes the unlock isn’t “how do I apply AI to my business.” It’s “what am I annoyed about paying for that I could just... make?”

Here’s where this connects to something bigger.

There’s a thesis floating around that AI will benefit the “lower half of the K” through cost compression. Small operators, solo founders, freelancers will access capabilities that previously required headcount. The playing field levels.

I want to believe this. But the 90% Trap suggests a problem.

If individuals can’t close the gap between AI capability and deployed value, the productivity gains don’t accrue to them. They accrue to organizations with the infrastructure, engineering resources, and operational discipline to close that last 10% at scale.

Which means the actual AI productivity story might not be hiring 1,000x engineers or democratized capability for all. It might be that companies simply avoided entry-level hires and margin expansion.

That’s a different outcome. It’s capital owners capturing gains through headcount reduction while individuals struggle to convert their AI experiments into economic value. It’s deflationary without being democratizing.

Here’s how we’ll know which story is true.

If AI is genuinely democratizing productivity, we should see it show up in labor productivity statistics—output per worker rising across the economy, including small businesses and solo operators.

If it’s primarily enabling hiring avoidance, we’ll see it in labor force participation instead. Fewer entry-level jobs. Longer job searches. Productivity gains concentrated in organizations, not individuals.

Right now, the latter signal is stronger. Companies are publicly celebrating “doing more with less.”

The infrastructure bull case assumes the productivity gains are real and broadly distributed. The 90% Trap suggests they might be real but narrowly captured.

I don’t have a thesis on where this lands.

What I’m confident about: the micro and macro are connected. The same dynamic that keeps smart individuals stuck in demo mode is the dynamic that could keep AI’s economic benefits concentrated at the top of the K.

If you’re building AI tools and nothing’s changing about your economics, you’re not alone. But you might be a data point in a larger pattern.

But the tinkering phase needs to end eventually—for individuals and for the economy. The productivity gains have to show up somewhere other than investor decks and avoided job postings.

We’ll see.

VL