The Silent Skill Drain: How Entry-Level AI Automation Threatens Future Growth

TL;DR for operators

Entry-level automation is usually discussed as a headcount issue. That is too crude. The sharper operational question is whether automation changes which juniors get access to which experts. A firm can keep the same number of junior roles and still damage its future skill pipeline if more of those roles move away from high-quality mentors.

The paper’s uncomfortable result is that automation can raise output immediately while reducing long-run growth. The reason is not mystical. If senior experts need fewer juniors to complete routine work, the displaced juniors do not vanish in the model. They get reallocated to less-skilled mentors. They still have jobs. They simply inherit weaker tacit knowledge. Very efficient, except for the small matter of tomorrow’s expertise.

For operators, the key metric is not “how many junior tasks did AI replace?” It is “did AI expand or shrink high-quality learning exposure?” The answer depends on the type of technology. Automation of apprentice-bearing work can weaken knowledge transmission. Task-creating technologies can strengthen it by giving top experts more junior-facing work. Scale-expanding automation can also help, but only when it truly lets experts supervise more work and more learners rather than merely doing the same work with fewer juniors.

This is a theoretical paper, not a field experiment. Its value is diagnostic: it gives firms a vocabulary for separating productive automation from skill-pipeline cannibalisation. The practical test is whether AI adoption increases current output while quietly reducing the future supply of people who know how the work is actually done.

The problem is not empty graduate seats; it is worse first teachers

Junior work has always had an awkward dual role. It produces something today, and it teaches someone how production actually works. The spreadsheet, the due diligence file, the patient intake, the legal memo, the model validation note: these are not glamorous tasks. They are also the entry points through which novices learn how experts notice problems, choose exceptions, and exercise judgment.

Enrique Ide’s paper, Automation, AI, and the Intergenerational Transmission of Knowledge, builds a model around that simple but underpriced fact.¹ Its central claim is not the familiar “AI may reduce entry-level employment.” That claim is now sufficiently rehearsed to require its own waiting room. The more interesting claim is subtler: even if all novices remain employed, entry-level automation can reduce long-run growth by changing who learns from whom.

That distinction matters. A company can point to stable junior headcount and still be hollowing out its apprenticeship structure. If the best experts now use AI to complete routine junior tasks with fewer assistants, juniors may be pushed toward less productive teams, smaller projects, weaker mentors, or lower-learning work. The firm still has juniors. It just gives more of them worse first teachers.

That is the paper’s mechanism. AI does not merely substitute for labour. It changes the matching market between novices and experts. Once that market is part of skill formation, the productivity question becomes intertemporal: what looks like output growth this year may be expertise erosion three promotions from now.

The model treats work as both production and transmission

The paper uses a task-based overlapping-generations model. Individuals live for two periods. In the first, they are novices. In the second, they become experts. Experts differ in skill. Novices begin identical and acquire tacit knowledge by working with an expert. In the baseline model, a novice matched with an expert inherits that expert’s skill level in the next period.

That last assumption is intentionally stark. It compresses a messy workplace reality into a clean transmission rule. The point is not that every analyst becomes her managing director. If only. The point is that early-career placement affects the kind of tacit knowledge a worker carries forward.

Production requires expert skill plus routine tasks. Some tasks can be performed by machines; the rest require novice labour. Machines are more productive than novices on the tasks they can perform, but they cannot perform everything. Experts orchestrate the work. Novices execute routine components while observing the expert’s judgment.

The key contractual friction is that tacit knowledge cannot be fully specified, verified, seized, or pledged as collateral. Novices cannot simply borrow against future earnings and pay top experts for access to mentorship. Instead, they “pay” by supplying current labour, often accepting lower wages in exchange for better future learning. The model represents this through a zero lower bound on wages: wages can fall to zero, but not below.

That constraint creates a familiar outcome. Lower-skilled experts may be inactive. Middle experts can hire novices by paying positive wages. The most attractive experts become oversubscribed: novices would like to work with them even at zero wages because the future learning value is so high, but wages cannot go negative to clear the market.

This is not just a labour-market curiosity. It is the engine of growth in the model.

Growth comes from top experts replicating themselves

In equilibrium, only a subset of experts are active producers. Each active expert hires multiple novices. The next generation’s skill distribution is therefore formed by taking the current distribution, cutting off the lower inactive part, and reweighting the upper tail according to how many novices each active expert mentors.

In plain English: growth comes from the best experts replicating their practices through junior workers.

The model’s long-run growth depends on two objects. The first is expert span of control: how many novices an active expert supervises. In the baseline, this is the number of non-automated tasks, $N-I$. The second is the thickness of the upper tail of the skill distribution, captured by $\theta$. When the initial skill distribution has a fat upper tail, the asymptotic growth factor is:

This expression is doing more than decorative algebra. It says growth rises when top experts supervise more novices, and when the economy has a sufficiently rich upper tail of expertise to replicate. If $N-I$ falls toward one, the replication engine weakens. If the upper tail is too thin, sustained asymptotic growth vanishes.

The paper also shows that, after normalisation, the expert skill distribution converges to a Pareto form. That is not the business takeaway. The business takeaway is simpler: a firm’s future expertise depends on how many juniors are learning from its best operators, not merely on how many juniors it employs.

Automation gives twice before it takes

The paper’s main result comes from an automation shock. Suppose a new technology allows an additional measure $\Delta$ of tasks to be automated. The model restricts attention to the case where the post-shock economy still grows and all novices remain employed. This is important. The result does not depend on junior unemployment.

On impact, output rises. There are two immediate gains.

First, machines are more productive than novices on the newly automated tasks. Existing expert-led production teams become more efficient. That is the obvious gain, and it is the one most dashboards will happily count.

Second, because high-skill experts now need fewer novices per production process, some novices are released into the market. They are reallocated to lower-skill experts who were previously inactive. More experts produce. Aggregate output rises again through this extensive margin.

So far, automation looks excellent. It is cheaper, faster, broader. The consultants may now enter the room with a slide titled “unlocking productivity.” One can almost hear the invoice.

Then the dynamic term arrives.

Because top experts now hire fewer novices, fewer members of the next generation inherit top-tier tacit knowledge. More novices learn from lower-skill experts. The best practices of the frontier diffuse more slowly. Future experts are weaker than they would have been. The immediate level gain is real, but the growth rate falls.

The paper summarises the output ratio after the automation shock as:

The first term is the direct productivity gain: machines outperform novice labour on newly automated tasks. The second term is the immediate extensive-margin gain: displaced novices allow more lower-skill experts to become active producers. Both are positive.

The third term is the slow poison. At the moment of adoption, it equals one because expert skills are already predetermined. After that, it declines. The more time passes, the more the lower-quality mentor composition compounds. In the model, output eventually falls below the no-automation path.

This is the paper’s most useful discipline for executives: an automation project can be privately rational, operationally impressive, and dynamically damaging at the same time.

The figures illustrate the mechanism; they are not empirical forecasts

The paper’s figures are best read as mechanism illustrations, not as empirical estimates of how any particular AI product will affect a real economy.

This matters because theory papers are easy to misuse. The paper does not say “AI automation is bad.” It says the welfare effect of automation depends on whether the technology reallocates novices away from high-quality mentors or toward them.

That is a much less viral claim, which is one reason it is more useful.

The welfare loss comes from the wrong missing price

The welfare result is not just “future output may be lower.” The paper shows that automation can reduce aggregate welfare relative to non-adoption when the discounted future loss from weaker knowledge transmission outweighs the immediate output gain.

The friction is the zero lower bound on wages. If novices could borrow against future earnings and pay experts for mentorship, some high-learning positions would be preserved. Top experts would internalise more of the future value created by training novices.

But tacit knowledge is hard to verify and embodied in the worker. The novice cannot pledge it as collateral, and the expert cannot easily appropriate its future returns once the novice becomes an expert. Current labour becomes the payment mechanism for access to mentorship.

Automation devalues that payment mechanism. If AI lets a senior expert perform junior execution tasks without the junior, the novice has less current labour to offer in exchange for learning. The mentor may adopt automation because it raises current production, even though the social value of the displaced learning slot is larger.

The market failure is therefore not that firms dislike training. It is that the future value of tacit knowledge transmission is not fully priced into the adoption decision. Conveniently, the missing price is attached to people who have not yet become powerful enough to have budget authority. Capitalism does enjoy tidy timing.

Task creation is the clean counterexample

The paper does not treat technology as one blob. It distinguishes automation from technologies that create new labour-intensive tasks.

A task-creating technology introduces new work that novices perform productively under expert supervision. Under the paper’s full-adoption condition, such a technology raises output immediately and increases long-run growth. Why? Because the most skilled experts now hire more novices. More juniors learn from the best mentors. Best practices diffuse faster.

This contrast is crucial for business interpretation. The question is not whether AI touches junior work. The question is which kind of junior work it creates, removes, or transforms.

A tool that eliminates low-learning clerical drudgery may be harmless or even helpful. A tool that removes the very tasks through which novices observe expert judgment is different. A tool that creates new apprentice-bearing work around validation, exception handling, data preparation, client adaptation, safety review, or domain-specific interpretation may strengthen the skill pipeline.

The difference is not moral. It is architectural.

Scale effects are the escape hatch, not the default excuse

The baseline model assumes experts have fixed capacity: each expert oversees a fixed production process. That is intentionally restrictive. Section 6 relaxes it by allowing experts to scale their operations when automation reduces supervision time.

This creates a countervailing force. Automation may reduce the number of novices needed per project, but it may also let a top expert oversee more projects. If the second effect is strong enough, the total number of novices working with top experts can rise. In that case, automation accelerates rather than slows knowledge transmission.

This is the strongest argument for optimistic AI deployment. It is also the easiest argument to overclaim.

For the scale effect to rescue the skill pipeline, AI must genuinely expand expert supervisory capacity. It is not enough for an expert to produce more solo output. The expert must be able to oversee more work in a way that still exposes novices to meaningful judgment. If AI lets senior people disappear into private productivity cocoons, that is not scale. That is a nicely formatted apprenticeship vacuum.

The paper highlights two cases where displacement is likely to dominate. First, when automation is “so-so”: machines are adopted but only modestly more productive than novices, so they save little expert time. Second, when expert attention remains bottlenecked by the hardest-to-supervise human tasks. If the limiting factor is judgment, coordination, or accountability, automating a slice of routine work may not free enough senior attention to mentor more juniors.

That boundary is directly relevant to generative AI. Many current tools speed up drafting, summarising, coding, and search. They do not automatically increase the number of juniors a senior can supervise well. The tool may make the expert faster without making the expert more available.

The business metric is learning exposure, not junior headcount

The paper’s most actionable idea is that firms should audit the allocation of novices across experts before and after AI deployment. Headcount is too blunt. “Hours saved” is too self-congratulatory. The missing metric is high-quality learning exposure.

A practical AI adoption review should ask four questions.

First, which junior tasks are learning-bearing? These are tasks where novices see how experts frame ambiguity, detect anomalies, prioritise risks, negotiate trade-offs, and integrate evidence. They may look routine from the outside. They are often not routine from the inside.

Second, which tasks are execution-only? Automating genuinely low-learning work is less dangerous and may even improve training if it frees time for better supervision. The paper’s appendix explicitly allows for tasks with no learning value; automating those does not weaken knowledge transmission and can be beneficial.

Third, does the tool increase or decrease the number of juniors working with top experts? The relevant denominator is not total junior staff. It is junior exposure to the people and teams whose practices should be replicated.

Fourth, does AI scale mentorship or merely scale individual output? A tool that helps a senior review more junior work, explain decisions, generate feedback, and supervise more projects may strengthen transmission. A tool that lets the senior bypass juniors may weaken it.

This suggests a simple operating framework:

The managerial conclusion is not “keep inefficient tasks because tradition.” That is nostalgia wearing a CFO disguise. The conclusion is that firms should redesign entry-level work so that automation removes waste without removing access to tacit judgment.

What the paper directly shows, and what Cognaptus infers

The paper directly shows a theoretical possibility: in a model where novices learn tacit knowledge from experts, knowledge-transfer contracts are incomplete, and expert capacity shapes novice allocation, entry-level automation can raise output on adoption while lowering long-run growth and welfare. It also directly shows that this effect does not require entry-level unemployment. Mentor composition alone can do the damage.

The paper also directly shows that the effect is not universal. Task-creating technologies can improve output, growth, and welfare by increasing the number of novices who learn from high-skill experts. Scale effects can reverse the automation result if automation allows top experts to supervise more total novices.

Cognaptus infers the operational lesson: AI governance should include a skill-transmission audit. Before replacing junior execution, firms should identify whether the task is part of an apprenticeship loop. If it is, the firm should either preserve meaningful junior involvement or create substitute learning channels that keep novices close to expert judgment.

What remains uncertain is empirical magnitude. The paper reviews early evidence consistent with junior displacement in AI-exposed work, but it is careful: those studies do not yet directly track whether displaced juniors move to worse mentors or how their future skill trajectories change. That is precisely the hard measurement problem. It is also precisely the one firms should not wait ten years to discover accidentally.

Where the argument stops

The model is clean because it is narrow. Real workplaces are messier in at least six ways.

Novices are not identical. Some learn faster, some need more structure, and some use AI as a private tutor rather than a crutch. Expert quality is not perfectly observable. Institutional prestige can be a noisy proxy for actual mentorship. Tacit knowledge transmission is not exact; a novice does not clone the mentor’s skill like a low-budget science-fiction premise. Some junior tasks have little learning value. Some AI systems may create new learning surfaces. And contracts may evolve: firms may build training bonds, internal academies, residencies, simulation environments, or AI-mediated coaching systems.

The paper recognises several of these boundaries. It extends the baseline to consider scale effects, occupational choice by experts, non-learning tasks, and imperfect or span-dependent transmission. These extensions do not erase the main mechanism. They specify when it bites.

The practical boundary is therefore clear. The paper should not be used as a blanket argument against automation. It should be used against lazy automation accounting. If a firm measures only current productivity and ignores the future production of expertise, it is not being data-driven. It is simply choosing a short measurement window and hoping nobody notices the missing asset.

Automate the present without liquidating the future

The real danger in entry-level AI automation is not that every junior task is sacred. Many are not. Some deserve a dignified burial and perhaps a small invoice for the time they stole from civilisation.

The danger is that firms automate the work through which novices learn how experts think, then act surprised when the next generation has fewer experts. In Ide’s model, that surprise is not a cultural complaint. It is an equilibrium outcome.

The operational mandate is simple: automate low-learning work, preserve high-learning access, and design new tasks that keep novices attached to expert judgment. The firms that get this right will not merely save labour. They will compound capability. The firms that get it wrong may enjoy a few clean quarters before discovering that they have optimised away their own succession plan.

Cognaptus: Automate the Present, Incubate the Future.