The Jobs Data Refuses a Simple Story

I spent May grading my predictions. Then PwC published a jobs report that wouldn't fit neatly into the scorecard.

Its 2026 Global AI Jobs Barometer, released June 15, finds that the companies most exposed to AI had stronger headcount growth than the least exposed: 52% against 36%, measured from 2018 to 2025. That period includes several years before ChatGPT arrived, and exposure doesn't establish that AI caused the growth. Even so, I've spent two years warning about displacement. I owe numbers like these more than a defensive footnote.

A company can automate a substantial amount of work and still hire. It can lower prices, find new customers, or offer something it previously couldn't afford to produce. If enough of that happens, employment can grow. I want that outcome. There is no prize for correctly predicting somebody else's unemployment.

Then I got to the entry-level findings.

In PwC's US analysis, the junior roles most exposed to AI were seven times more likely than the least exposed to demand skills usually associated with senior people. Openings for the roles it calls “seniorised” grew 35% since 2019. Other entry-level openings fell 10%.

So there may be a job for you. You just need the judgment and experience that the job used to help you acquire. Good luck with that at twenty-two.

Who gets to be part of the good news?

That question survives the reassuring headline. Suppose a company hires ten experienced people to run a much larger operation with AI, while the twenty junior positions it once used never open. The business might do well. Its customers might get a better service. Someone hoping to enter the profession would have good reason to feel less cheerful about the arrangement.

I'm describing a possibility, not what PwC counted. The point is that a company's growth figure doesn't tell us how newcomers fare. For that we need hiring by experience level, what the advertised jobs require, and who actually gets them.

This is where I think of Asimov's Foundation. Psychohistory concerns the behavior of vast populations. It isn't much help to the individual who needs to know what happens to their own life next Tuesday. An economist can be right about growth while a graduate is right about being shut out. We ought to be capable of hearing both without telling the graduate to take comfort in the aggregate.

The same care applies when comparing studies. The Stanford work I wrote about last August examined employment through payroll records. PwC's barometer combines job advertisements with company and occupational data. Different populations, different time periods, different questions. I can't honestly recruit one study to cancel the other just because this month's result is more convenient.

Cheaper work can create a customer

There is an opportunity here that deserves more attention than it gets in the layoff debate. Imagine a small business that has never commissioned market research because the price was prohibitive. If an AI-assisted service makes it affordable, that business becomes a customer for work that wasn't being bought before. Someone has to understand the business well enough to ask a useful question and deliver something the owner can act on.

That is the sort of opportunity I want to look for in my own venture experiment. Selling a useful service to somebody who couldn't previously afford it is a much more interesting prospect than generating another thousand articles nobody requested.

But demand has to be there. The existence of a capable agent doesn't oblige anyone to become its customer. You still have to find a problem that matters, reach the person who has it, and make the price work. Those questions can get lost when a demonstration makes the production part look effortless.

I would like more research that follows those changes all the way through a business. Did cheaper production lead to lower prices? Did customers buy more? Did the employer hire, reduce hours, or simply keep the savings? Those choices are part of how a capability becomes somebody's livelihood.

OpenAI's Economic Research Exchange, announced June 8, is one new effort to support external empirical work on these effects. I'll be interested in what comes out of it, particularly findings that are awkward for the company paying for the research.

Actionable Steps

If you approve hiring, pull out your junior job descriptions and read them as if you had just graduated. Which requirements could a capable beginner reasonably meet? Which ones would someone have to teach? Put a person's name and paid time against the teaching. Otherwise “we hire for potential” is something nice on the careers page.

If you're applying, show a decision you made with these tools. Explain what you checked and why you changed the first answer. A prospective employer needs to see whether you can be trusted with a problem, and a stack of polished outputs doesn't tell them much about that.

As for my own forecast: I still expect serious disruption. June's evidence makes me less willing to describe it as a uniform retreat of employment. The details will decide who gets through it well, and the details deserve more space in this series.

LLAP. 🖖