AI is killing jobs and creating them at the same time. That’s not a dodge — it’s what the data actually shows. And the split is becoming one of the most consequential fault lines in the entire AI story.

A new report released this week found that companies classified as “high-intensity AI adopters” — those spending roughly $30 per employee per month on AI tools in their first quarter of adoption — saw overall headcount grow by 10.2%. More striking: entry-level headcount at those firms rose by 12%. That’s the cohort everyone has been telling us is doomed.

Meanwhile, Goldman Sachs data shows that AI has erased approximately 16,000 net jobs per month over the past year, with Gen Z and entry-level workers bearing the disproportionate brunt. These two facts are both true, and they paint a picture far more complicated than either the AI utopian or the AI apocalypse crowd wants to acknowledge.

The firm-expansion theory

The most useful framework I’ve seen for understanding why both things can be simultaneously true comes from the research report itself. For software and technology firms, AI can make core outputs — writing code, debugging, building internal tools, producing technical documentation — cheaper and faster to produce. Lower production costs in these workflows raise the return to expanding the whole firm, not just the engineering team.

In other words: if AI makes it cheaper to build software, you build more software. More software means more products, more customers, more support, more sales. The efficiency gains don’t just replace workers — they fund growth that wouldn’t have been economically viable before.

This is the “firm-expansion” theory of AI and jobs, and it has real empirical backing in this week’s data.

The other half of the story

But notice who the hiring is happening to: tech-forward, knowledge-work firms. VC-backed companies growing fast. The kind of organisations that would have been expanding anyway. The researchers are honest about this — the paper does not show that AI universally creates jobs. It shows that AI correlates with job creation in a specific type of organisation, pursuing a specific kind of growth.

The companies that buy subscriptions, run pilots, and never make sustained investment? They see no headcount gains. The report explicitly distinguishes between the two groups. Which means the AI productivity dividend is flowing disproportionately to firms with capital, technical staff, founder networks, and management bandwidth to actually absorb and deploy these tools deeply.

That’s not a small caveat. That’s the whole story.

The widening gap

The implication is uncomfortable: AI may be amplifying existing structural advantages rather than democratising productivity. Firms that were already winning are getting more efficient at winning. Firms stuck in the experimentation phase — running ChatGPT subscriptions, maybe a pilot or two, no real integration — are falling behind on productivity without gaining the offsetting growth that sustained AI adoption brings.

This is the widening gap the report describes, and it’s one that traditional labour policy is completely unprepared for. We have a lot of mechanisms for helping workers displaced by automation find new work. We have almost no mechanisms for helping firms that can’t absorb AI effectively become firms that can.

The workers at the companies winning are mostly fine. The workers at the companies losing are not.

What this actually means

The AI jobs story isn’t a single story. It’s two parallel economies operating on completely different timescales and incentive structures. In one, AI is a tool for firm expansion and human leverage — the most productive junior engineers in history, producing more than their predecessors could dream of, hired at premium salaries because they generate outsized returns. In the other, AI is a displacement machine for routine administrative and knowledge-work tasks, quietly eliminating the on-ramp positions that used to train people into those higher-productivity roles.

The junior developers who are winning are the ones at AI-native firms building with AI. The junior developers who are losing are the ones doing the kind of structured, repetitive knowledge work — data entry, document processing, basic service roles — that AI now handles at lower cost and higher consistency.

Neither side of this is wrong. Both are real. The mistake is treating it as a single phenomenon.

Where you land on this largely depends on where you sit. And that’s a policy and moral problem as much as an economic one.