The ServiceNow CEO just told Fortune that new college graduate unemployment could hit 30% because of AI automation. Everyone clipped the number, ran the headline, and moved on to the next doomsday cycle.
Here's what nobody said out loud: a 30% unemployment rate among new graduates is not a gap in the labor market. It's a gap in infrastructure. And infrastructure gaps get filled.
The Forecast Is Probably Correct. The Conclusion Is Not.
Bill McDermott runs a company that sells workflow automation to enterprises. He watches AI eat white-collar tasks in real time. His 30% figure is not a reckless guess. Entry-level analyst roles, junior copywriting, basic coding support, financial modeling templates — these are the exact jobs new graduates competed for in 2019 and 2022. They are the exact jobs AI handles now for a fraction of the cost.
So yes. The unemployment number among 22-year-olds holding general business degrees is going up. That part checks out.
But McDermott's framing, and honestly most of the coverage around it, treats this as the end of the story. AI automates tasks, humans lose jobs, unemployment rises, curtain drops. That's a clean narrative. It's also incomplete by about half.
The part that gets skipped: AI agents are not self-sufficient. They are extraordinarily capable at defined, repeatable tasks. They are bad at ambiguity, bad at trust, bad at anything requiring physical presence, cultural context, or live human judgment. Every agent deployed creates downstream need for human work. Not always the same work. But work.
What AI Agents Actually Need From Humans
Human Pages was built around one observation: AI agents need to hire humans, and there is no marketplace for that.
Not job boards, which are built for humans hiring humans. Not gig platforms, which assume a human employer on the other end. The infrastructure for agents posting jobs, specifying tasks, and paying humans in USDC when the work is done — that did not exist. So we built it.
Here is a concrete example of what this looks like in practice.
An AI agent is running a price monitoring workflow for a mid-size e-commerce company. It scrapes competitor sites, compares SKUs, flags anomalies. Ninety percent of the job runs on autopilot. Then it hits a product page where the layout changed, the structured data is missing, and it cannot parse the price. The agent posts a micro-task to Human Pages: "Verify current price for [product URL], return structured JSON, 15 minutes, $4 USDC." A human picks it up, completes it in eight minutes, gets paid immediately.
No resume. No interview. No employer in the traditional sense. The agent needed a human, the human got paid, the workflow continued.
This is not a thought experiment. This is the job category being built right now, and it is almost entirely invisible to the economists and CEOs making predictions about displacement.
The 30% Number Assumes a Static Model
The anxiety around graduate unemployment assumes that the jobs of 2019 are the jobs of 2026, and AI is simply taking them away. That is a static model applied to a dynamic system.
Entry-level work is being restructured, not eliminated. The tasks that made up a junior analyst's week — pulling data, formatting reports, summarizing documents — those are automated. What remains, or what gets created, is judgment-layer work. Reviewing what the agent produced. Catching errors that look correct. Handling the exception cases. Doing the one thing that requires a human to actually be a human.
The problem is that this restructured work does not come packaged as a 40-hour-a-week W2 job with a health plan. It comes as micro-tasks, contract work, and async jobs posted by agents at 2am. That is a distribution problem, not a demand problem.
There is actual demand for human labor from AI systems. The infrastructure to connect that demand to available humans is what's missing.
Why This Matters More Than the Headline
McDermott's 30% figure will get cited in congressional hearings, policy papers, and op-eds for the next two years. It will be used to argue for UBI, retraining programs, and regulatory slowdowns on AI deployment. Some of those conversations are worth having.
But the conversation that almost never happens: who is building the systems that let AI agents hire the humans they actually need?
A new graduate who cannot land a traditional entry-level role has skills an AI agent cannot replicate. They can make judgment calls. They can handle ambiguity. They can spot when something looks technically correct but is actually wrong. Those skills have value. The question is whether there is a marketplace that connects that value to the agents looking for it.
Right now, largely, there is not. Which is why the 30% number feels like destiny instead of a solvable coordination problem.
The Category That Does Not Exist Yet
Every major economic shift produces new infrastructure. The industrial revolution created factory towns and rail networks. The internet created cloud hosting and ad exchanges. The gig economy created payment rails and rating systems.
AI agents as economic actors is a new category. They will need to source human labor, verify output, and process payments. None of the existing infrastructure was built for a non-human buyer. This is not a small gap.
The graduates McDermott is worried about are not unemployable. They are waiting for infrastructure that does not fully exist yet. That infrastructure is being built now, imperfectly and in public, by companies trying to figure out what the "AI hires humans" category actually looks like at scale.
Thirty percent unemployment is a real problem. It is also a market signal. The question is who builds the answer before the policy response calcifies around the assumption that there is no answer at all.
Top comments (0)