What gets more valuable as everyone uses AI.

AI made almost everything cheaper to produce. Writing, analysis, code, whole internal tools. The friction that used to keep all of it in check is mostly gone.

That sounds like a pure win, and in a lot of ways it is. But I’ve been watching what happens when the cost of doing more drops close to zero, and it’s not all good. When more of everything is free, the default becomes more of everything. More words, more analysis, more features, more dashboards nobody asked for.

Which means the skills that are actually getting more valuable aren’t the ones AI does for you. They’re the ones that resist the pull toward more. Here are the four I keep noticing.

1. Saying less.

Writing used to be rate-limited by the effort of writing. That friction did quiet work: it forced people to get to the point, because getting to the point was less typing.

That governor is gone. I’ve watched a simple piece of feedback turn into an essay, because the person ran it through AI and the AI happily produced ten paragraphs of what-ifs. “I was looking around and noticed this, not sure if it’s a problem, but here are six things it could be.” The volume is wild relative to the actual signal.

The catch is that someone still has to read it. A human, not another agent, eventually has to close the loop. Every extra paragraph of low-value content is work quietly transferred to whoever’s on the other end. Multiply that across a team and you’ve built a machine for manufacturing each other’s busywork.

AI can help you be concise. It’s genuinely good at compressing a rambling thought into one clean sentence. But it won’t do that by default. Left alone, it expands. Getting brevity out of it takes deliberate effort, and that effort is worth more than it used to be.

2. Guarding scope.

Here’s the one I’m guilty of myself.

I prototype our product designs directly in the codebase now, vibe coding alongside Claude. And when you’re in that flow, it is very easy to start adding scope you wouldn’t have added before. A little extra detail here, a nice-to-have there. Each addition feels free in the moment, because it basically is.

Sometimes that’s fine. On the engineering side, the team is moving fast enough that a little extra scope gets absorbed without much cost. But it doesn’t always get absorbed, and I’ve watched features balloon well past what the goal actually required, mine included, just because I was flowing and everything was cheap to attempt.

I’m seeing this across every discipline, not just engineering. A marketing team that needed a new email drip out the door in five days instead spent a month building a microsite and a set of engagement tools. Tools that, because they weren’t built by engineers, break in places. So the content that AI could have helped them ship fast never shipped, and what did ship is over-built and fragile.

This is where the old startup disciplines matter more, not less. The Pareto principle, ship and iterate, stay ruthless about what you actually need. Those weren’t constraints imposed by a lack of resources. They were good judgment, and good judgment gets more valuable exactly when the resource constraints that used to enforce it disappear.

3. Having a point of view.

This one comes down to how these models work.

An LLM predicts the most likely next piece of text given everything before it. Trained across an enormous amount of human writing, “most likely” tends to mean “most common.” So left unsteered, the output drifts toward the average. The expected phrasing, the expected layout, the expected solution.

For a lot of problems, average is completely fine. You don’t need world-class design to win. You need to solve a real problem. Craigslist is ugly and built an empire because the product did the job. Solving the problem comes first, and it beats polish almost every time.

But average is a ceiling, and the ceiling matters more than people think. A lot of AI-generated UI is going to look the same, because it’s all being pulled toward the same median. Standing out takes the thing the model can’t give you: a point of view. Your own design system. Your brand actually expressed in the interface. A take on the problem that isn’t the most probable one. That’s original thinking, and it’s becoming scarce precisely because everything else is converging.

The irony is that this post is the argument for itself. The observation that AI regresses everyone to the mean is not the most likely thing an AI would tell you about AI. That’s the point.

4. Knowing when it’s wrong.

Related, but not the same. Original thinking is about having a better answer. This is about knowing the default answer is off, and being able to steer the model back.

When I let Claude build something with no guardrails, the first pass is coherent. It looks right at a glance. Then I get into it and it’s a C+. The UX is generic, the structure is subtly illogical, and the deeper into the details I go, the more it needs correcting. I built Trailfold, and the default design it produced was not good. The UX is decent now, but only after a lot of iterating, moving things around, regrouping, telling the model what it got wrong and why.

None of that direction would have been possible without years of doing the work by hand. The knowledge required to fix a UI at that level is esoteric. You have to know it should be different, and know what different looks like. There’s a reason people are racing to build domain-specific models: the broad ones like Claude and Gemini need a lot of hand-holding once you get past the surface. Most of us don’t have a UX-only model to lean on. We have the generalists, and the generalists need an expert holding the wheel.

For now, that makes deep experience more valuable, not less. I see it clearest with engineers. Senior engineers are getting amplified, shipping more and better than ever. Junior engineers are shipping bad code fast and, worse, not learning anything in the process, because the model is doing the reps that used to build their judgment. The same thing is happening in product. A PM who never had to write the epics and sit with the tradeoffs doesn’t develop the instinct to look at what AI generates and say “no, we don’t need this, and this is missing.” That instinct is the whole job, and it’s getting harder to build at exactly the moment it’s getting more valuable.

The pattern underneath.

Look at the four together and they’re the same skill wearing different clothes.

AI lowered the cost of producing more, and of settling for average. So the things appreciating in value are restraint and judgment. Saying less. Building less. Refusing the median answer. Knowing when the machine is wrong. None of it is what AI does for you. All of it is what keeps AI pointed at something worth doing.

The tools got a lot more powerful. The person steering them is worth more than ever. If you’re trying to figure out where AI actually helps your product and where it quietly creates work, that’s the kind of thing I like to dig into. It’s often a judgment problem, not a tooling one.