The entry layer goes first

Two findings from 2025 appear to contradict each other and in fact say the same thing. In Denmark, Humlum and Vestergaard found almost no effect of generative AI on earnings or hours through 2024, despite wide adoption. In the United States, Brynjolfsson, Chandar and Chen found a relative employment decline of around 13 percent among workers aged 22 to 25 in the occupations most exposed to AI, while older workers in the same occupations were unaffected. Neither finding has yet been replicated. Read together, they point to the same place.

The aggregate is quiet because the change is concentrated. It is concentrated at the entry layer because that is where the work is cheapest to check and least dependent on judgment: first drafts, first responses, first passes at research, the routine tasks that a junior person does for two years while learning to do the tasks that cannot be automated. An agent can do those tasks now, and verifying its output is cheap because the output is simple. So that is where agents go first, and that is where the hiring quietly stops.

The field experiments add a twist. In the customer-support study by Brynjolfsson, Li and Raymond, the AI assistant helped the least experienced workers most; it supplied what they had not yet learned. That is good news for a novice who is already employed. It is bad news for the next novice, because the organisation now needs fewer of them to produce the same output, and the agent does not need two years to get up to speed.

Here is the problem that no productivity measure captures. The entry layer is not just a cost. It is the pipeline through which an organisation produces the people who, ten years later, have the judgment to supervise, challenge and improve its systems, including its agents. Automate the entry layer and for five years you feel only the savings. In year six you discover that the senior people who verify the agents are retiring, and the cohort that should replace them was never hired. Catalini, Hui and Wu call this the "missing junior loop," and they pair it with a second mechanism they call the codifier's curse: the experts who write down their knowledge to train the systems are the same experts the systems then replace.

No one in the current structure of a company is positioned to see this coming. The business unit that stopped hiring juniors sees a leaner cost base. Finance sees the same. IT sees a successful deployment. HR sees a hiring freeze it did not request and cannot explain. The problem is visible only from a vantage point that holds the human and the artificial workforce in one view and thinks in years rather than quarters. That vantage point is what the Chief Workforce Officer is for.

What would a CWO do? Three things. Redesign the human roles in agent-heavy task families around verification, exception handling and judgment, and pay them for the expertise those require rather than the hours they take. Rebuild entry-level roles as apprenticeships in verification, so that junior people learn the work by checking it, which is how most crafts have always been learned. And track one number that no finance function tracks: how many people in the organisation could still do each automated task by hand. When that number approaches zero, the organisation has not become efficient. It has become dependent, and it will find out on the day the agent fails.

The Danish null and the American decline are the first two data points in a long series. They are not yet a trend, and I would not build policy on them. But they are exactly the pattern the argument predicts, and they arrive at the layer the argument says they should. That is reason enough to give someone the job of watching.