The AI Jobs Data Is Telling Two Stories at Once
Anthony Pompliano reversed his position on AI and jobs this week. Marc Andreessen amplified it with three words: "This is the way." Inside of a few hours, the post had a quarter million views and the executive class had a clean new talking point.
Pompliano's argument is straightforward. He used to believe AI would replace entry-level jobs and work its way up the org chart. The data, he says, has changed his mind. He cites software engineer hiring growth, a 5.6% increase in new-grad hiring over the last twelve months, falling unemployment for college-educated 20 to 24 year olds, and the WSJ's reporting that AI created 640,000 new jobs in the U.S. between 2023 and 2025. His conclusion is that AI makes companies more productive, productive companies hire more people, and the result is more startups, more jobs, more economic growth.
The argument is clean and the data points are real. There are also other data points, equally real, that say almost the opposite.
The numbers Pompliano cites are accurate
The 640,000 figure traces to LinkedIn's analysis of job posting data, which the Wall Street Journal covered in detail. Roles like Head of AI, AI Engineer, and Data Annotator are growing fast. The World Economic Forum cites a broader LinkedIn figure of 1.3 million new AI-enabled roles plus another 600,000 in AI-supporting data center work. The directional claim that AI is creating new categories of work at scale is well-supported.
The 5.6% new-grad hiring increase comes from the National Association of Colleges and Employers and is a genuine projection. Companies hiring aggressively to keep up with product demand, which Pompliano sees in his own portfolio, is a real pattern.
If you read only these sources, the conclusion writes itself. Productivity gains drive hiring, and the early headlines about AI eliminating jobs were premature.
The numbers other people cite are also accurate
SignalFire studied hiring at the largest public tech firms and venture-backed startups between 2019 and 2024. They found a 50% decline in new-role starts for workers with less than one year of post-graduate experience. That decline held consistent across sales, marketing, engineering, recruiting, operations, design, finance, and legal. It was not a tech-only phenomenon.
Federal Reserve data, summarized by Josh Bersin in late 2025, shows the unemployment rate for recent college graduates near 10%, the highest level since the 2011 recession recovery. The CNBC reporting and Bloomberg Businessweek's May 2026 cover story both frame 2026 as potentially the worst job market for new grads since the pandemic. Cengage Group's research found that only 30% of 2025 graduates had secured full-time work in their field, down from 41% the year before.
So we have two data sets. One says AI is creating hundreds of thousands of new jobs. The other says entry-level hiring at the largest companies has fallen by half and recent-grad unemployment is near a fifteen-year high. Both are true. The interesting question is why.
The gap between the two stories is the entire AI Operations conversation
Headline AI roles are growing at companies that are taking adoption seriously. They are hiring Heads of AI because they have decided they need someone to own the function. They are hiring AI engineers because they have built things that need to be maintained. They are hiring data annotators because they have realized that the quality of their AI output depends on the quality of their training data, and that work has to be done by humans.
These companies are using AI to expand what they can do. The new roles exist because the new work exists, and the new work exists because someone took the time to figure out what AI should actually be doing inside their operation. We've written before about why this kind of operational thinking is what separates AI that works from AI that just runs.
Entry-level hiring is collapsing at companies that adopted AI as a substitution for labor. Leadership looked at their org chart, identified the tasks junior employees did, and used AI to do those tasks instead. The work that used to be a stepping stone into the company is now a prompt. The career ladder loses its first rung, and the company saves money in the short term while quietly destroying its own talent pipeline.
Same technology, two outcomes, driven entirely by whether leadership treats AI adoption as transformation or substitution.
This is a pattern we've been writing about for a while. We argued earlier this year that AI pilots fail because the workflow underneath them never gets rebuilt. The labor data is what that pattern looks like at the macro level. The companies showing up in the 640,000-new-jobs number are the ones that rebuilt the work. The companies showing up in the SignalFire 50% decline are the ones that pointed AI at the existing org chart and let it cut.
Pompliano's optimistic framing misses this distinction. So does the doom-loop narrative on the other side. The macroeconomic question of whether AI creates jobs or destroys jobs is unanswerable in the aggregate, because both things are happening at once depending on which kind of company you look at.
The Jevons paradox argument doesn't say what people think it says
The most-liked reply to Pompliano's tweet pushes on a real weakness in his argument. Jake notes that even if Jevons paradox applies to AI and labor, "that doesn't mean it's the outcome people want." More productivity leads to more demand, which leads to more work, not less. This is technically true and philosophically unsatisfying.
The implicit promise of AI in the workplace, repeated for the last three years, is that it will give people their time back. It will handle the toil. It will let humans focus on creative, strategic, meaningful work. Pompliano's argument quietly inverts that promise. AI doesn't free you from work in his framing. It makes you more productive at work, which leads your employer to want more of you doing more work. That's not a bug in his argument. That's the argument, and it's worth being honest about.
The companies handling this well are the ones that decided early what they wanted AI to actually do for their people. Some companies chose throughput, and their teams now produce more. Others chose capacity, and their teams now take on work the company never could before. A few chose relief, and their teams genuinely have more time for things that matter. None of these are wrong choices, but they are choices, and pretending the technology will make them on its own is how you end up with a workforce that is more productive, more anxious, and more disillusioned at the same time.
What this means if you are running a company
Pompliano is correct that the macro picture is more complicated than the early doomsayers suggested. The data on AI-created roles is real and the growth is real. The trap is reading his argument as permission to assume things will work out automatically. They will work out for the companies that do the work of figuring out what AI is actually for in their organization. They will not work out for the companies that just point AI at their existing workflows and expect the cost savings to compound forever.
Figuring out what AI is for is harder than it sounds. It's the Discovery Discipline we built our methodology around. Most leadership teams skip it because it feels slower than just deploying tools and counting licenses. The companies that don't skip it are the ones that end up hiring instead of cutting.
If you are leading a company right now, the question worth sitting with is not whether AI will replace your people. It is what you want your people to be doing in three years that they are not doing today, and whether your current AI strategy is moving you toward that or away from it. Most strategies we see are moving companies toward a smaller version of themselves. There are other directions available, but they require a different kind of work.
The companies that figure this out will keep showing up in the optimistic data. The companies that don't will keep showing up in the pessimistic data. Both data sets will keep getting bigger, and the macro debate will keep generating tweets and counter-tweets, because the answer depends entirely on which kind of company you are looking at.
For any individual company, the work is making sure you end up on the right side of that split. That work is not a debate to follow on Twitter. It's a set of decisions about how your people work, what your workflows look like, and what you are actually trying to build. The companies that take those decisions seriously are hiring. The companies that don't are quietly deleting their own future bench.
Pompliano changed his mind because new data came in, which is the right instinct. The further version of it is to keep asking which data, from which kinds of companies, doing which kinds of things with the technology. The patterns are visible if you look for them, and the choice of which pattern your company ends up in is still yours to make.