The silent substitution

The point is not the sudden disappearance of entire professions. The point is the change in the learning and work layer through which career entrants have traditionally grown into responsibility. AI systems therefore change more than individual tasks. They change who still gains experience inside organisations, who prepares decisions and how many experienced people are needed for a given volume of work.

The evidence is not yet causally conclusive. It is strong enough, however, to treat career entry in especially exposed fields as an early warning signal.

Keywords: generative AI; agentic AI; career entry; knowledge work; labour market; non-hiring; expert work; productivity; organisational redesign

Silent substitution begins where entry-level, support and learning work is calculated out of organisations. This layer was rarely glamorous. It consisted of research, drafting, review, documentation, standard cases and follow-up. But it was precisely through this work that young employees learned how a domain really operates. When this work disappears, output is not the only thing that disappears. Experience disappears as well.

The first signal is therefore not necessarily a wave of layoffs. More often it is non-hiring: fewer advertised entry-level roles, departures that are not fully replaced, less external support work, smaller trainee programmes, fewer junior roles and more work landing with experienced employees who are supported by systems. The personnel pyramid is not first cut at the top. It is no longer reliably rebuilt at the bottom.

This mechanism is politically difficult to grasp. It does not produce a single event. There is no factory closure, no clearly bounded wave of layoffs, no occupational group that disappears overnight. There are many small decisions: an analyst role not refilled; a trainee programme paused; fewer working students; fewer junior developers; less legal research; less first-level support; less clerical processing. In aggregate, this can become a larger labour-market rupture than a visible dismissal round.

Core thesis

In short: when AI removes entry-level, support and learning work from organisations, work volume is not the only thing that is lost. The experience layer from which later responsibility emerges is missing.

That experienced specialists benefit in the short term fits this picture. People who know the domain, customers, risks and quality standards can steer significantly more cases, analyses, contracts, tickets or projects with agent systems. But higher productivity does not automatically lead to more employment under cost pressure. It can also mean that fewer people are needed for the same revenue, case or project volume.

Agentic AI means systems that work toward an objective across multiple steps: they take in context, access data and tools, generate intermediate results, obtain approvals, document decisions, escalate exceptions and close process cases. Human responsibility does not disappear as a result. But it shifts: away from execution and support work, toward objective setting, control, exception handling, liability and system design.

The difference is practical, not semantic. A chatbot accelerates a junior task. An agentic system can absorb parts of a junior role. Several interacting agents can change a process so that a department needs fewer people for the same throughput. This changes not only work, but also the personnel structure.

The evidence is still incomplete. For Germany, it cannot yet be shown robustly that the decline in entry-level roles is monocausally caused by AI. Economic weakness, the interest-rate turn, cost pressure, skills mismatch and sector-specific investment restraint also play a role. The point is therefore more cautious: pressure on junior roles is not proof that AI is the sole cause. It is a relevant warning signal because it appears in fields that generative and agentic systems can address particularly well.

The sources used here do different things. Peer-reviewed studies and working papers carry the theoretical and empirical base. Job-ad analyses and company surveys show early market movement, but they do not replace official labour-market statistics. Market forecasts and vendor reports say little about employment effects; they mainly help classify diffusion and product development. These distinctions matter because otherwise indicators quickly become certainties.

Germany: employment expectations and political framing

The ifo Institute reaches a similar background signal from another direction. In a special analysis on AI use, 27.1 percent of companies expect AI to lead to job reductions over the next five years; only 5.2 percent expect additional jobs, while two thirds expect no change. These are expectations, not measured employment effects. They do show, however, that many companies read AI more as a future reduction impulse than as an expansion impulse (Wohlrabe, 2025).

The German Council of Economic Experts also classifies AI as a technology that, more strongly than earlier automation waves, can affect non-routine activities in highly qualified professions. At the same time, the macroeconomic effect remains open and is embedded in productivity and structural-change narratives. This caution is methodologically correct. It should not, however, lead to overlooking early shifts in career entry (Sachverständigenrat, 2025).

In January 2026, the German Federal Government stated in response to a parliamentary question that there were so far no indications that the use of AI technologies had reduced career-entry opportunities. That is understandable if the question is hard causality. For policy, it is still risky because hiring and role indicators react earlier than unemployment statistics. If action begins only once causal proof is officially clean, it may come too late (Deutscher Bundestag, 2026).

International signals: age and exposure

Evidence from the United States is closer to this mechanism. Brynjolfsson, Chandar and Chen examine employment data and find a relative employment decline of 16 percent for 22- to 25-year-olds in highly AI-exposed occupations, while more experienced employees in the same or less exposed occupations remained stable or continued to grow. This is a strong signal, but not proof for Germany (Brynjolfsson, Chandar and Chen, 2025).

Anthropic approaches the issue through observed usage patterns and finds a more cautious, but similar signal. For 22- to 25-year-olds in exposed occupations, there is no clear increase in unemployment, but there is a weaker job-finding rate. After the introduction of ChatGPT, it is around 14 percent lower than in 2022; the source itself describes the finding as narrowly statistically significant and data-dependent. That fits silent substitution precisely: the effect appears more in slowed entry than in open unemployment (Anthropic, 2026).

Productivity evidence as a precursor to changed staffing needs

In the short term, such findings support the complementarity thesis: employees become better, faster and more productive. That is real and often desirable. The labour-market consequence depends on what companies do with this productivity. When standard tasks are completed faster and more cheaply, the need to maintain large volumes of junior work declines. When experienced employees coordinate the work of entire teams with system support, the number of experienced employees needed per work volume can later decline as well.

Phase 1: junior substitution

This does not automatically lead to layoffs. More often, demand for entry disappears. Graduate programmes become smaller, project teams are planned without junior staffing, internal training seems too expensive, and applicants are expected to already bring what they previously learned inside the company. This is the paradox: precisely because AI could make beginners more productive, fewer beginners are hired if their original learning tasks are replaced.

Phase 2: senior leverage

After that, experienced specialists benefit. They know the domain, customers, risk, exceptions and quality standards. With AI agents, they can steer more cases, more code, more analyses, more contracts, more support processes or more operating processes. In the short term, this makes the senior look more indispensable, not more replaceable.

That picture is correct, but only for the first step. The senior is not replaced immediately; the senior's work is more strongly leveraged. Exactly that later changes staffing needs. If one experienced employee with agents can produce or control the output of several previous roles, demand for experienced employees does not have to grow at the same rate. Under competitive and cost pressure, it can decline per work volume.

Phase 3: fewer senior roles per work volume

The third phase is the most uncertain part of the model and should be read as a hypothesis: after junior substitution, demand for experienced roles can also decline. Not every senior role disappears. But roles lose scarcity when their value consists mainly of recurring pattern recognition, process knowledge, domain memory, standard judgement and control of support work. These are exactly the elements that agentic systems can increasingly prepare, document or partly execute.

What is protected is not the title "senior". What is protected is responsibility, liability, strategic prioritisation, customer trust, negotiation, exceptional judgement, cross-functional systems understanding and the ability to define objectives and risks under uncertainty. That work remains scarce. But it is not identical with today's quantity of experienced specialists. Scarcity moves from execution to responsibility.

The pipeline risk

The most sensitive long-term consequence comes from the combination of these phases. If juniors are no longer hired, the pipeline for future seniors is missing. If, at the same time, seniors are strongly leveraged by agents, this shortage is barely visible at first. Companies can work for several years with existing experienced people and AI systems. Later, a generation of people is missing that learned responsibility in the domain through real cases.

Such a structure would be unstable: few experienced accountable people, many graduates who struggle to enter, and growing dependence on technical process support. That is why the decline in entry-level roles is more than a problem for young people. It is an indication that the structure of knowledge work is changing.

A second objection: productivity gains can create new demand. If services become cheaper, new products, more projects and additional work can emerge. This speaks against simple zero-sum models. The silent-substitution thesis therefore does not claim that productivity always reduces employment. It claims that productivity changes personnel architectures under certain cost and competitive conditions.

Third, AI can empower career entrants. In some studies, less experienced employees benefit strongly from AI support. But this does not automatically mean that companies hire more entrants. Enablement at the task level can coincide with declining demand for entry-level roles at the organisational level.

Another objection concerns platform data. Job advertisements on Stepstone, LinkedIn or other platforms show important signals, but they do not capture every hire, internal mobility, informal recruitment or public employment services. A robust early-warning system must therefore combine several data sources.

The stock error: employees instead of hiring flows

Unemployment rates and employment stocks react late. Silent substitution starts earlier: in hiring plans, non-replacement, reduced trainee programmes, changed project roles and rising experience expectations. Whoever measures only the stock sees the redesign only after an entry generation has already received fewer learning cases.

The occupational-category error: occupations remain, roles disappear

Many studies classify AI risk by occupation. That is useful, but coarse. Companies organise value creation not only by professions, but by roles in processes. A law firm still needs lawyers, a software company still needs developers, an industrial company still needs controllers and operations managers. But it may need fewer people for research, pre-review, documentation, standard cases, support work, reporting and internal coordination. The occupation remains visible while its entry and middle roles shrink.

The productivity error: complementarity can become substitution

Productivity improvements initially sound like good news. Economically, they can be good news if new demand, new products and new activities emerge. For individual companies under cost pressure, however, productivity often means: the same output with fewer people or more output without proportional staff growth. Complementarity at the task level can therefore become replacement or relocation at the organisational level.

The infrastructure error: diffusion happens through standard software

A significant part of diffusion will not start as a major transformation programme, but through existing enterprise software: CRM, ERP, office suites, ticketing systems, HR systems, project management, development environments and industry applications. Gartner expects that by the end of 2026, 40 percent of enterprise applications will include task-specific agents, up from less than 5 percent in 2025. Microsoft describes a development from assistants to human-agent teams and then to organisations in which people set direction and agents execute business processes. Both are market or vendor perspectives, not neutral labour-market studies. For the diffusion question, they are still relevant (Gartner, 2025; Microsoft, 2025).

The imagination error: after juniors come seniors

Perhaps the biggest blind spot lies in the assumption that experienced specialists are permanently protected. In the short term, they do benefit because they can use and control AI better. In the long term, exactly this productivity gain can be the mechanism that reduces their number. The question is not whether one experienced person is fully replaced. The question is how many experienced people an organisation still needs per process, revenue unit, mandate, case or product.

Research: from occupational risk to process and role risk

Research should not only ask which occupations are AI-exposed, but which roles inside an occupation previously had learning, support and control functions. This requires task- and process-near datasets. The relevant unit is not only "lawyer", "controller" or "software developer", but contract review, research, pre-analysis, standard decision, case closure, test generation, customer triage or dispatch proposal.

Longitudinal studies inside companies that introduce agentic systems would be especially useful. They would need to measure how hiring volume, non-replacement, senior/junior ratio, throughput times, output per head, error rates, wage structure, learning paths and internal mobility change. The decisive point is not only employment, but staffing need per unit of process performance.

Statistics: early indicators for the experience ladder

An early-warning system would need to combine several indicators:

• Share and absolute number of advertised entry-level roles by occupational field and AI exposure.
• Share of job advertisements with de facto experience requirements despite entry-level labels.
• Time to first qualification-adequate job after graduation.
• Applications per successful entry, differentiated by field of study.
• Non-replacement rates after exits from junior and clerical roles.
• Share of AI-supported or AI-led process steps in companies.
• Senior/junior ratio in knowledge-intensive occupations.
• Learning-case volume per entrant: how many real cases does a junior still handle directly?

Without such indicators, the debate remains fixed on unemployment rates. That is convenient, but too late and too coarse.

Education: away from support work, toward agentic responsibility

Universities and vocational education must teach more than AI tools. They must create learning paths in which students and career entrants learn to review outcomes, define process objectives, evaluate evidence, identify chains of error, take responsibility and decide with incomplete information. If simple research, standard text, standard code and standard analysis are automated, education systems can no longer treat these activities as central proof of capability.

Companies, too, must not fully outsource the learning layer. They still need people who understand the domain, customers, systems and risk from their own experience. Anyone who delegates all learning cases to AI for cost reasons saves in the short term, but creates a long-term shortage of accountable people.

Labour-market policy: not only skills, but access

Labour-market policy often treats further training as the answer to technological change. In the case of silent substitution, that is not enough. If companies offer fewer entry-level positions, additional qualification alone does not solve the access problem. A graduate can be well qualified and still find no entry if the economic function of the entry-level role has been automated.

Politically relevant instruments would therefore secure access to real learning and responsibility environments: funding programmes for AI-supported trainee models, public data on entry rates, tax incentives for qualification-adequate entry roles, new models of dual knowledge work or sector-specific standards for talent pipelines. The state does not have to preserve inefficient roles. But it should make visible and address the disappearance of career paths.

Enterprise strategy: internalise the pipeline risk

The better strategy is not to remove every junior. It is to redefine junior work: not as cheap support work, but as controlled training in agent steering, validation, exception handling and process responsibility. Companies that preserve this pipeline may later have an advantage when there are many AI users, but few people with real accountability routines.

That picture is this: generative AI and agentic systems first reorganise the support and learning layer of knowledge work. As a result, demand for career entrants can decline. Then agent systems increase the productivity of experienced specialists so strongly that the number of senior roles needed per work volume can also decline. Organisations become flatter, more process-driven and more strongly shaped by a few accountable humans plus many digital agents. Scarcity moves from execution to responsibility.

The decline of entry-level roles is therefore not a side issue. It is an early signal of a possible reorganisation of knowledge work. Policy understandably waits for unambiguous causal evidence, while companies already act economically. This asymmetry is risky. If the unemployment statistics show the effect clearly, career paths may already have been damaged.

The first labour-market crisis of AI does not have to begin as mass layoffs. It can begin as silent non-hiring. That is exactly why it can be overlooked for a long time.

The thesis is testable. If it is wrong, entry-level roles in AI-exposed occupations should rise significantly again after cyclical adjustment, junior/senior ratios should remain stable, non-replacement should play no systematic role and agent integration should not reduce staffing need per work volume. If it is right, the coming years will show continued compression of entry-level work, rising experience requirements, higher output per head and later declining volume demand even for mid-level and experienced specialists.

For an ageing economy with weak productivity growth, AI is a major opportunity. But this opportunity has a shadow side: productivity can damage career ladders if they are not institutionally rebuilt. That is where the political discussion should begin.

Anthropic (2026): Labor market impacts of AI: A new measure and early evidence. Online: anthropic.com/research/labor-market-impacts (accessed May 6, 2026).

Autor, D. H. (2015): Why Are There Still So Many Jobs? The History and Future of Workplace Automation. Journal of Economic Perspectives, 29(3), 3-30. DOI: 10.1257/jep.29.3.3.

Brynjolfsson, E., Chandar, B. and Chen, R. (2025): Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. Online: digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine (accessed May 6, 2026).

Brynjolfsson, E., Li, D. and Raymond, L. R. (2025): Generative AI at Work. Quarterly Journal of Economics, 140(2), 889-942. DOI: 10.1093/qje/qjae044; first published as NBER Working Paper No. 31161.

CDU, CSU and SPD (2025): Verantwortung für Deutschland. Koalitionsvertrag 2025. Online: cdu.de/downloads/koalitionsvertrag-2025 (accessed May 6, 2026).

Deutscher Bundestag (2026): Regierung: Berufseinstieg durch KI nicht negativ beeinflusst. hib short notice 57/2026, January 27, 2026. Online: bundestag.de/presse/hib/kurzmeldungen-1139672 (accessed May 6, 2026).

Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2024): GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306-1308. DOI: 10.1126/science.adj0998.

Friedrich, M. and Kagerl, C. (2026): Künstliche Intelligenz in deutschen Betrieben: Jeder vierte Betrieb nutzt mittlerweile generative KI. IAB-Kurzbericht 8/2026. Online: iab.de/presseinfo/jeder-vierte-betrieb-in-deutschland-nutzt-generative-ki (accessed May 6, 2026).

Gartner (2025): Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025. Press release, August 26, 2025. Online: gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025 (accessed May 6, 2026).

Hennrich, J. and Schaller, D. (2026): Berufseinsteigerinnen und Berufseinsteiger im Fokus: Unternehmen setzen auf Soft Skills und Künstliche Intelligenz. Randstad-ifo-HR survey, Q4 2025. Online: ifo.de/fakten/2026-01-19/berufseinsteigerinnen-im-fokus-unternehmen-setzen-auf-soft-skills-und-ki (accessed May 6, 2026).

Microsoft (2025): 2025 Work Trend Index: The Year the Frontier Firm Is Born. Online: microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born (accessed May 6, 2026).

Noy, S. and Zhang, W. (2023): Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science, 381(6654), 187-192. DOI: 10.1126/science.adh2586.

Peng, S., Kalliamvakou, E., Cihon, P. and Demirer, M. (2023): The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv:2302.06590. Online: arxiv.org/abs/2302.06590 (accessed May 6, 2026).

Sachverständigenrat zur Begutachtung der gesamtwirtschaftlichen Entwicklung (2025): Nationaler Produktivitätsbericht 2025. Online: sachverstaendigenrat-wirtschaft.de/fileadmin/dateiablage/gutachten/fg2025/2025_Nationaler_Produktivitaetsbericht.pdf (accessed May 6, 2026).

Stepstone Group (2026): Analyse-Update für das Gesamtjahr: Einstiegsjobs auch 2025 deutlich unter dem Fünfjahresschnitt. Press release, January 29, 2026. Online: thestepstonegroup.com/deutsch/newsroom/pressemitteilungen/analyse-update-fuer-das-gesamtjahr-einstiegsjobs-auch-2025-deutlich-unter-dem-fuenfjahresschnitt (accessed May 6, 2026).

Wohlrabe, K. (2025): Mehr KI, weniger Jobs? Was Unternehmen in Deutschland erwarten. ifo Schnelldienst digital 8/2025. Online: ifo.de/publikationen/2025/aufsatz-zeitschrift/mehr-ki-weniger-jobs-was-unternehmen-deutschland-erwarten (accessed May 6, 2026).