What actually matters now

The most quoted number in the AI-jobs discourse is Dario Amodei’s prediction that AI could wipe out half of entry-level white-collar jobs within one to five years. It’s a good headline. It’s also the wrong thing to be scared of, and the fear it produces sends people to exactly the wrong defenses.

Here is the number that actually tells you what to do. In PwC’s 2026 Global AI Jobs Barometer, the most AI-exposed junior roles are seven times more likely to demand skills we used to consider senior — judgement, leadership, the ability to decide under ambiguity. Those roles grew 35 percent since 2019. The other entry-level roles, the ones that were really just structured tasks in a trench coat, shrank 10 percent.

Read those two facts together and the picture isn’t “AI takes the jobs.” It’s “AI deletes the bottom rung and moves the floor up.” The apprenticeship layer — the years where you did the boring, structured work badly and slowly until you got good — is being compressed into a prompt. That’s a more specific problem than mass unemployment, and it has more specific answers.

What stops mattering

Start with the losers, because being honest about them is the whole point.

The thing that stops mattering is the ability to produce the artifact. The deck. The first-pass analysis. The competitive summary. The clean-but-generic draft. For twenty years, being the person who could reliably turn a vague ask into a competent document was a career. That skill is now a feature in a product that costs twenty dollars a month.

The data backs the discomfort. Commodity creative work — template design, basic copy, routine reporting — is seeing rates collapse, while premium, strategic creative roles grow. Recent graduates in computer science are posting higher-than-average unemployment. “Learn to code” was advice for a world that ended.

What actually matters, in order

I’ll be concrete, because vague advice about “human skills” is how consultants avoid saying anything.

Judgement under ambiguity. Not answers — the model has answers. The scarce thing is knowing which question to ask, which of five plausible answers is right for this situation, and when the confident output is confidently wrong. This is why PwC finds junior roles demanding senior skills: the tasks that survive are the ones where being wrong is expensive and the machine can’t be trusted alone.

Taste. When production is free, the bottleneck becomes selection. A thousand generated options are worthless without someone who can tell the good one from the merely competent, and defend the call. Taste used to be a nice-to-have on top of craft. It’s becoming the craft.

The ability to own an outcome. AI produces outputs. It does not own results. Someone still has to stand in the room and say “this is the position, I decided it, here’s why, and I’ll answer for it if it’s wrong.” That accountability is the least automatable thing in any organization, and it’s where the compensation is migrating.

Fluency with the tools, as table stakes. The wage premium for demonstrable AI skill is real — PwC and multiple labor datasets put it well into double digits. But notice the framing: fluency is the price of entry now, not the differentiator. Being “good at AI” in 2028 will be like being “good at email” — assumed, invisible, and worthless as a selling point on its own.

What actually matters: education

The credential story is inverting, and slowly, which makes it easy to miss.

A degree was a proxy for “this person can do structured knowledge work.” That’s precisely the capability AI commoditized. So the signal value of the generic degree is falling, while the value of things a degree was never good at proving — judgement, taste, the ability to ship something real — is rising.

This doesn’t mean education stops mattering. It means the evidence shifts from credentials to artifacts. In the next decade, “here is a thing I built and the thinking behind it” will beat “here is where I studied” in more rooms than it used to. The portfolio eats the résumé. (I’m aware of the irony of writing that on a portfolio site. I built it because I believe it.)

What actually matters: geography

The lazy take is that AI makes location irrelevant — work from anywhere, talent is global, the map dissolves. The data says the opposite is happening, at least for now.

AI-exposed firms are pulling away from their peers: dramatically higher productivity growth, and they’re expanding hiring faster, not slower. That concentration has a geography. The advantage compounds where the density already is — the places with the capital, the frontier firms, and the clusters of people who learned to use the tools first. Remote work loosened the map; AI is quietly re-tightening it around a smaller number of winners.

The honest read for an individual: proximity to where the frontier actually gets built — physical or networked — still matters more than the “work from anywhere” story admits.

So here is the advice I’d give someone starting out, stripped of comfort. Don’t compete with the model at producing artifacts — you will lose, and the wage data already shows people losing. Compete at the things it can’t do: decide, discriminate, own the result. Build evidence instead of collecting credentials. And get close, in whatever way you can, to where the work is actually moving.

The bottom rung is gone. That’s real, and pretending otherwise is cruel. But the ladder is taller than it’s ever been for the people who figure out how to start climbing from the middle.