The entry-level roles that AI is eliminating are the ones that produced the senior judgment institutions now need to govern it.
This is not a prediction about a future labour market. It is a description of something already underway in legal services, financial services, consulting, and risk functions. Junior analysts are fewer. The work they used to do, drafting and reviewing and synthesising and pattern-matching across large volumes of material, is increasingly handled by models that are faster, cheaper, and available at every hour of the day. The productivity case for this is clear and largely correct. The governance case is where it gets complicated, and the complication is not obvious until you ask where senior judgment actually comes from.
The hidden function of junior work
Entry-level work in a knowledge-intensive function looks like output generation. Documents get drafted. Contracts get reviewed. Data gets checked. Judged purely by the output, it is exactly the kind of repetitive, high-volume task that automation handles well, and this is why it is the first to go.
But the output was never the only thing the work produced. It was also the mechanism by which a professional built judgment. You learn to review a contract by reviewing a great many of them, making errors and being corrected until an internal model of what matters slowly assembles itself. Risk intuition forms the same way. Not by reading frameworks, but by watching decisions that followed the framework and still failed, and decisions that broke from it and still worked. The accumulation is slow. It is mostly invisible. And it is almost never written down, because it lives in the person, not in the document they produced.
This is the part that the productivity case does not price. When an institution automates the volume work, it captures the visible output and quietly removes the conditions under which judgment used to form. The contracts still get reviewed. The person who would have learned to review them by doing it a thousand times does not get made.
Where the judgment was supposed to go
The judgment built through entry-level work is not a nice-to-have. It is the exact capability that the next generation of governance depends on, and the timing of its disappearance is the heart of the problem.
Governing an automated system well requires a specific kind of judgment: knowing which outputs to trust, which frameworks to apply, and when to override a model recommendation that is locally plausible and globally wrong. That judgment cannot be specified in a rulebook, because its whole value is in the cases the rulebook did not anticipate. It is built the slow way. Through years of close contact with the underlying work, the kind that junior roles used to provide. The senior people who currently hold it were, almost without exception, formed by exactly the entry-level grind that is now being automated away.
So the institution faces a sequence it has not quite registered. It is automating the work that builds judgment, in order to deploy systems that require judgment to govern. The judgment it needs today still exists, held by people who came up before the change. The judgment it will need in ten years is being un-built right now, one eliminated junior cohort at a time, in functions that are measuring only the efficiency they gained and not the pipeline they thinned.
Why this is hard to see from inside the decision
The institutions moving fastest to deploy AI in their knowledge functions are, in many cases, thinning their own future supply of governance judgment without recognising it. The reason they cannot see it is structural, not a failure of intelligence.
The benefit of automation is immediate, measurable, and attributable. Fewer junior hires, faster turnaround, lower cost per unit of work: these show up in the current period and can be pointed to. The cost is deferred, diffuse, and almost impossible to attribute. A shortage of seasoned judgment ten years from now will not arrive with a label explaining that it was caused by automation decisions made a decade earlier. It will look like a generic talent problem, or a run of bad senior decisions, or a governance function that cannot find people who can do the work. The cause and the effect are separated by enough time that nothing connects them in the institution's own accounting.
This is the same asymmetry that produces most slow-building institutional failures. The gain is legible now. The loss is real but unprovable until it arrives, and by then it is someone else's problem to explain. A decision structure that weighs only what it can measure in the current period will choose the automation every time, and will be unable to see what it gave up, because the thing it gave up does not show up anywhere a current report would look.
What taking the problem seriously would mean
None of this is an argument against deploying AI in knowledge functions. The productivity gains are real, and an institution that refused them on principle would lose to one that did not. The argument is about doing it with the pipeline in view rather than out of it.
Taking the problem seriously means treating the development of judgment as a thing that has to be designed for once it stops happening automatically. If the junior work that used to build seasoned professionals is no longer there to build them, the institution that wants seasoned professionals in a decade has to ask where they will now come from, and build that path deliberately. That might mean preserving certain formative work even where a model could do it faster, on the explicit grounds that the point of the work was never only the output. It might mean rethinking how judgment is transmitted when the apprenticeship that used to transmit it has been automated away. What it cannot mean is assuming the supply will be there, because the mechanism that produced it is the precise thing being removed.
The institutions building AI capability today are also, implicitly, deciding how much of the human judgment that will eventually govern that capability they are willing to keep producing. Most are making that second decision without realising they are making it at all. The question the fastest-moving institutions have not yet asked themselves is where the human judgment to govern these systems will come from once the work that built it is gone.
I write about governance, risk, and the decisions institutions find hardest to make. If this is relevant to a problem you are working through, reach me at aan@asifahmednoor.com.
← Back to writing