The missing junior loop

There is a slow failure that no quarterly report will ever show, and it is the one I would most want a Chief Workforce Officer to be watching. It concerns how an organisation makes its own experts.

Expertise in most professional work is produced in one way. A junior person does the routine version of the task, badly and then adequately, for two or three years, under the eye of someone senior. Through repetition they develop the thing that cannot be taught directly: the sense of when a result is wrong, the feel for the case that does not fit the pattern, the judgment. Then they become the senior person, and the loop runs again. Every profession that depends on judgment, from auditing to medicine to engineering to law, has some version of this apprenticeship, usually unofficial, always essential.

Agents enter at exactly the point where the loop begins. The routine version of the task is what they do best and what is cheapest to check. So the organisation gives it to them, stops hiring the juniors who used to do it, and congratulates itself on the saving. For five years nothing visible happens, because the senior people are still there, and they are now the ones checking the agents, which they can do because they learned the work by hand. In year six or seven they begin to retire or move on. The organisation looks for their replacements and finds that it stopped making them.

Catalini, Hui and Wu, whose 2026 paper on the economics of advanced AI I have cited throughout this series, give this the name it deserves: the missing junior loop. They pair it with a second mechanism they call the codifier's curse. The experts who write down their knowledge in order to train and instruct the agents are precisely the experts whose knowledge the agents then make redundant, so the act of teaching the system erases the position from which teaching was possible. The two mechanisms together describe an organisation that consumes its own capacity to verify, and does not notice until the capacity is gone.

The early labour-market evidence is consistent with the first half of this story. The Stanford finding of a relative employment decline among 22-to-25-year-olds in AI-exposed occupations, with older workers unaffected, is what the missing junior loop looks like in its first years. It is not yet replicated, and I would not overstate it. But it is the right shape, in the right place, at the right time.

What can an organisation do? Not refuse the agents; the routine work will be automated, and should be. The answer is to redesign the apprenticeship rather than abolish it. Junior roles become apprenticeships in verification: the junior checks the agent's work, under the eye of someone senior, and learns the task by learning what a wrong answer looks like, which is how craft has always been transmitted. It is slower than doing the work by hand, and it requires the organisation to employ people whose immediate output is lower than the agent's, in exchange for having experts a decade from now.

No business unit will make that trade on its own, because the cost is theirs and the benefit is the organisation's, later. No finance function will propose it. It requires someone whose remit is the whole productive capacity of the organisation over a horizon longer than a budget cycle. That is the Chief Workforce Officer, and this, more than any dashboard or lifecycle, is the reason the role has to exist: someone has to be responsible for the organisation still knowing how to do its own work.