I’ve spent the last few days reading several papers on AI and employment, expecting the usual cheerful choice between:
“AI will eliminate half of all jobs.” and “Don’t worry, technology always creates new ones.”
The actual data is considerably more interesting. And, at least so far, considerably less apocalyptic.
There is little evidence of an economy-wide AI jobs collapse in the United States. Stanford’s recent review concludes that aggregate employment effects remain small, even in occupations highly exposed to AI. But underneath those aggregate numbers, something else appears to be happening.
Companies are beginning to separate. A relatively small group is adopting AI seriously, investing heavily, changing how work gets done and growing. The rest aren’t necessarily firing everyone. They’re simply beginning to look less competitive.
That distinction may matter much more than today’s arguments about whether ChatGPT is “taking jobs.”
First, Only a Minority Is Really Doing This
Depending on how you define adoption, you’ll get wildly different numbers. Ramp sees AI spending among more than half of the companies on its platform, but its customer base is disproportionately technological and isn’t representative of the whole U.S. economy.
The U.S. Census Bureau gives the more conservative number: roughly 20% of firms currently report using AI. Other surveys produce much higher figures because they measure employees using AI rather than firms systematically integrating it.
This distinction is important. An accountant using ChatGPT to rewrite an email is AI usage. It is not an AI-native company.
Even among companies classified as adopters, Stanford notes that comprehensive integration remains unusual. Most deployments are still limited to a couple of business functions or relatively infrequent use. So we are still surprisingly early.
Which makes what happens among the serious adopters particularly interesting.
The Companies Spending Heavily on AI Are Hiring More
Ramp and Revelio Labs linked actual AI spending data with workforce records covering 21,559 U.S. firms. This is much more useful than asking executives whether they consider themselves “AI companies.”
Money is wonderfully resistant to PowerPoint. The headline result is counterintuitive. Companies adopting AI didn’t shrink. The strongest adopters grew.
Firms making the largest AI investments increased employment by roughly 10% during the two years following adoption. Low-intensity adopters showed no statistically significant employment change. And look at the chart.
The two groups track each other reasonably closely before adoption. Then they separate.
Two years later, the high-adoption companies are around 21% above their pre-adoption headcount baseline, versus roughly 7% for matched companies that had not yet adopted. That’s a substantial divergence.
The naïve story was:
AI makes each employee more productive,
therefore companies need fewer employees.
The emerging story may instead be:
AI makes good companies more productive,
therefore they win more business,
therefore they need more employees.
Productivity and employment aren’t opposites if productivity expands the company.
Even More Surprisingly: They Hire Juniors
This is where the picture becomes complicated. The conventional argument says entry-level knowledge workers should be first against the wall. And there is evidence supporting part of that concern.
Stanford’s review finds that recent graduates are facing an unusually difficult labor market, and several studies have detected declining demand for younger workers in AI-exposed occupations. But the researchers are careful about causality: interest rates, post-pandemic over-hiring and remote work all complicate the story, and some deterioration began before ChatGPT.
Yet among Ramp’s high-intensity AI adopters, entry-level headcount actually increased by about 12% after adoption. This sounds contradictory. I don’t think it necessarily is.
The question may not be: Will companies still hire inexperienced people?
It may become: Which inexperienced people are worth hiring when AI dramatically amplifies what one capable person can do?
My expectation is that strong AI-native companies will continue hiring juniors. But they’ll become much more selective.
If an inexperienced employee can operate agents, understand systems, verify outputs, learn quickly and coordinate increasingly autonomous workflows, their leverage can be enormous.
If their primary value is producing the first draft of work that an agent can already produce in thirty seconds, the economics become considerably less pleasant.
So the ladder probably doesn’t disappear. The first rung gets narrower.
What About Everyone Else?
Here’s where I think the more important labor-market effect begins.
Stanford finds no broad AI-driven collapse in employment today. Unemployment has not risen faster in the occupations most exposed to AI than in the least exposed ones. Employment in coding-intensive occupations has slowed but remains positive.
Meanwhile, adoption remains highly uneven. That means we’re watching two processes simultaneously:
AI-native firms gaining leverage. and the rest of the economy adjusting much more slowly.
This matters because competitive pressure doesn’t require AI to directly replace a worker.
Imagine two companies. Company A introduces agents across sales, research, software development, operations and administration. Company B doesn’t. Company A doesn’t necessarily fire 30% of its employees. Instead, it ships faster. Experiments more. Serves more customers. Opens markets faster. Builds internal tools for things Company B still buys. Makes decisions with more information. And perhaps hires another 20% because the organization is expanding.
What happens to Company B? Eventually it has to respond. And that is where the employment pressure arrives.
Not: AI replaced Sarah in accounting.
But: Company A took 15% of Company B’s market, so Company B needs to restructure.
That is a completely different mechanism.
AI May Destroy Jobs Through Competition, Not Automation
I think this distinction gets lost in most discussions. We keep imagining AI replacement at the task level:
human task → AI task → human disappears
But economies operate at the company level too.
The actual chain may look more like:
AI → higher organizational productivity → lower costs / better products → market-share gains → competitors forced to restructure
The worker may lose their job in a company that barely used AI at all. Their job was still indirectly displaced by AI.
Just not by an agent sitting in their chair. This would also explain why aggregate labor statistics can remain fairly boring during the early stages.
Diffusion takes time. The strongest adopters pull ahead first. Competitors respond later. Then suppliers respond. Then wages and hiring patterns adjust. Then entire organizational structures change. Technology moves quickly. Economies have considerably more paperwork.
The Jobs I Expect to Grow First
If this interpretation is roughly correct, I would expect unusually strong demand for people in what I’d loosely call near-technical systems roles. Not necessarily computer scientists.
People who can understand a business process deeply enough to rebuild it around AI. Someone has to connect models to databases. Design agent permissions. Define verification gates. Restructure workflows. Build feedback loops. Determine what should remain human. Translate messy organizational knowledge into machine-readable state. Investigate failures. Design evaluations. Coordinate agents. Decide when automation is actually making things worse. The valuable skill isn’t merely “knowing AI.” Soon everyone will know AI.
It’s being able to look at a functioning human organization and ask: How should this system operate if intelligence becomes cheap?
That’s a much rarer capability.
Eventually Every Serious Company Becomes AI-Native
This is where I think the current discussion becomes too focused on adoption percentages. Whether today’s number is 20%, 40% or 50% isn’t ultimately the important question. If the productivity and growth differences persist, competitors won’t have much choice.
AI adoption stops being an innovation project. It becomes basic competitive hygiene.
Nobody today proudly announces that their company is: internet-first.
Or: spreadsheet-enabled. Or: cloud-compatible.
These technologies disappeared into the normal architecture of organizations. AI probably does the same. But becoming AI-native is more profound than buying Microsoft Copilot licenses.
It changes how the company itself is organized. And that brings us to what I think happens after the current phase.
The Next Competition Isn’t Human Coordination
For the last century, companies became better largely by improving systems for coordinating humans.
Management. Org charts. ERP. Email. Slack. CRMs. Project management. OKRs. Meetings. Dashboards. The corporation is, among other things, an enormous technology for making groups of humans behave somewhat coherently. AI agents change the object being coordinated.
The competitive question increasingly becomes: Who can build the better system of humans, agents, data, tools and feedback loops? And eventually even that description may become outdated. Because the strongest systems won’t merely execute workflows.
They’ll learn from them. An agent makes a mistake. The system captures it. An evaluation identifies the failure. Memory changes. A workflow changes. A permission changes. Another agent verifies the modification. The next execution improves.
At that point the competitive advantage isn’t: Our employees use better AI.
It’s: Our company learns faster than your company.
That’s a much more formidable moat.
This Is the Part I Think We’re Underestimating
The popular debate asks whether AI replaces workers. The current evidence suggests that question is premature and possibly badly framed. Heavy AI adopters are currently associated with more employment growth, not less.
Aggregate U.S. employment data does not yet show an AI jobs apocalypse. But adoption is uneven, and that’s precisely why I wouldn’t find the current stability particularly reassuring.
The first-order effect may be productivity.
The second-order effect may be competition.
And the third-order effect may be organizational evolution.
My working hypothesis for the next few years is therefore:
The best AI-native companies become dramatically more efficient and continue growing and hiring.
Their competitors lose relative productivity and market share and are forced to optimize.
Hiring shifts toward people capable of designing, supervising and improving human-agent systems.
Eventually AI-native operation becomes mandatory rather than distinctive.
And then the competition changes again. The winning company won’t be the one whose people communicate most efficiently. It won’t even necessarily be the company with the smartest model.
Models increasingly come from the same handful of providers. The winner will be the company that builds the best learning system around those models.
A system capable of observing its own work. Finding failures. Changing itself. Testing the change. Remembering the result. And doing the next iteration better.
For a hundred years, we optimized companies to coordinate people. We may now be entering the period where we optimize companies to learn. That sounds like a subtle difference. I suspect it isn’t.
Splitting into two economies rather than collapsing employment is a much more interesting read of the data, and it matches what I see from the small-business side. The gap is not between people who use AI and people who do not, it is between organisations that changed how work flows and ones that bolted a tool onto the old process.