Each arbitration is rational. The aggregate result dries up the tech pipeline without anyone intending it. Anatomy of a skill debt that AI accelerates.
Based on data-driven work from the Stanford Digital Economy Lab ADPemployment of developers aged 22 to 25 in the United States was, in the fall of 2025, around 20% lower than its peak at the end of 2022. Among the large American tech companies monitored by SignalFire, recent graduates only represent 7% of hires in 2025, with hiring volumes of new grads down by more than 50% compared to 2019.
Behind these figures there is something deeper than a tension of recruitment. A skills debt is accumulating: the savings made today by ceasing to train the profiles who should lead the organization in five to ten years. This debt does not appear in any balance sheet, it arises from individually rational arbitrations, and it will only become visible when it is too late to treat it at a reasonable cost.
Two superimposed phenomena, a single diagnosis
This decline mixes two dynamics that must be separated. The first is cyclical: correction of pandemic over-hiring, rise in interest rates, waves of layoffs in tech since 2023. This layer is cyclical and will be absorbed at least partially.
The second is structural. On a growing portion of codifiable tasks (generation of standard code, tests, documentation, scripts, local refactoring), a senior equipped with agents produces much faster what a junior-senior pair achieved before. For a manager evaluated on his quarterly deliverables, the trade-off has shifted. This layer will remain after the cycle.
And this is where the problem goes beyond the question of employment. The tasks that AI absorbs the fastest are precisely those that historically served as learning grounds for beginner profiles. What then disappears is not a position, it is part of the mechanism by which a beginner becomes, in a few years, a profile capable of independent technical judgment.
The French data do not allow, at this stage, to establish the same diagnosis with the same precision. L’Apec However, confirms a contraction in executive recruitment and a slowdown in IT recruitment in 2024-2025. Combined with cost pressures and the widespread use of code agents, these signals make the risk serious enough to be treated as a matter of governance.
An incentive mechanism without a fix
In some organizations, reducing junior hiring is an explicit choice, presented as strategic. Elsewhere, it is an emerging result: each manager optimizes locally, and the pipeline empties by accumulation.
The common point in both cases is the absence of quantification of the long-term cost. No one takes into account the value of the pool of skills that the organization no longer constitutes.
The HR manager manages staff and costs over 18 or 24 months. The skills pyramid at ten years old is not in anyone’s scorecard. There CIO manages its teams, the COMEX arbiters the programs, the general management defends its margin. At each level, arbitration holds up when examined alone. It is by aggregating it that a drying up of the pipeline appears that no one wanted and that no one corrects.
The most pernicious thing is that this lack of control provides immediate relief. For an HR manager under pressure on costs, not hiring juniors improves the indicators. For a manager pressed for deliverables, not training frees up senior time. For a DG which defends its margin, reducing incoming staff reassures analysts. Everyone benefits immediately from what collectively produces the debt.
The cost of inaction
Giving up training juniors does not eliminate the cost of skill, it shifts it over time. Tomorrow it will take the form of seniors more expensive in a tightening market, an increased dependence on service providers, a progressive loss of control of the IS, and technical decisions taken by profiles who know how to orchestrate agents without ever having borne the consequences of a real incident in production.
The history of computing calls for caution. With each wave of abstraction, compilers, object-oriented programming, frameworks, cloudthe total demand for developers ended up increasing, because the fall in marginal cost opened up new uses. Employment volume will likely recover.
An objection deserves to be taken seriously. The role of the senior will evolve. If seniority is redefined around the orchestration of agents, the supervision of systems and the framing of constraints, the path to skill development could shorten. It’s plausible. The fact remains that contextual judgment, that which is acquired in friction with reality (incidents in production, architectural arbitrations under constraint, negotiation with the profession), remains a tacit skill which requires time and exposure. This path will change shape without disappearing.
The time it takes to rebuild a skills pipeline is measured in years, although we don’t know exactly how many. What we can measure is that each year without action increases the cost of reconstitution and reduces the number of options available.
Let us be precise about the status of this reasoning. The available data shows a drop in junior hiring, not yet a proven drying up of the pool of seniors, which by construction will only be measured in several years. What I am describing is a reasoned hypothesis, not an established fact. But this is precisely the nature of a debt: we contract it well before paying the interest, and waiting for proof of its existence amounts to waiting until the moment when it will be too late to act.
In a previous article, I described the debt of credibility that successive transformation programs leave in organizations. The debt of competence obeys the same logic: invisible, produced by individually rational arbitrations, and discovered too late.
What an organization can do now
The point is not to save the junior out of moral duty. The implicit model of small-task learning is weakening, and it will no longer be enough. It is now necessary to deliberately rebuild a transmission system that the previous model produced for free.
Five actions are operational. All of them have a real political cost, and that is what measures their seriousness.
- Produce the critical skills pyramid at five years old. The age pyramid doesn’t say anything useful here. What matters is theinventory real operational experience: who knows how to stabilize a production on fire, who knows how to arbitrate a technical debt without help, who knows how to say no to a sponsor on factual bases. A simple table, presented once a year to the COMEX, with a projection based on constant numbers and policies.
- Redefine the KING of a junior. We measure it today on his direct productivity at 12 or 24 months, and on this basis, arbitration is lost in advance against an agent. A fair calculation includes the option of seniority at five or seven years, and the market cost of a senior trained elsewhere. An organization that knows how to do this calculation stops thinking in terms of junior versus agent and starts reasoning in junior plus agent versus external market at seven years.
- Index part of the remuneration variable of managers on the measurable development of juniors. Measurable means observable increase in autonomy based on defined criteria, not hours of training followed. Without this indexation, training costs senior time and lowers the manager’s short-term productivity, which amounts to discouraging him from devoting real effort to it.
- Integrate a junior in a supervised pair into each team that uses code generation agents extensively. The junior works with the senior and the agent. The senior validates, the junior learns judgment in a real context. The objective is measurable autonomy over twelve months over a defined scope. This is the only operational way to make acceleration by AI and the transmission of judgment coexist.
- Entrust the COMEX with an annual indicator for the renewal of critical skills, with a named owner, a dedicated budget, and a target trajectory by skill family. Such an indicator could follow a few simple measures: number of juniors supervised per senior, time to access defined autonomy, real exposure to critical incidents or arbitrations, share of teams using agents with an explicit transmission system. Without an owner, without a budget and without a target, the subject remains a hallway conversation. It gains the status of subject of governance the day someone reports on it.
The question that remains open
None of these actions will spontaneously go up on the COMEX. They assume that a manager agrees to defend an investment whose benefits will go to his successor, while the costs will appear on his P&L from this year.
The question that will remain asked for the next few years can be summed up in one sentence. Who in this organization has the responsibility to protect an asset whose performance is long-term, when all performance indicators are short-term? In the absence of a response, the debt continues to accumulate. And it will pay off at the precise moment when the organizations which have ignored it will discover that the seniors they need are rare, expensive and already captured by those who have continued to train.
