The apprenticeship void: AI augmentation and entry-level displacement in knowledge work

Something has changed about how people become good at knowledge work. For most of the twentieth century, the path was reasonably predictable: you joined an organisation at the bottom, did the tedious work nobody senior wanted, and slowly – through doing it badly, then less badly, then competently – built the kind of judgement that cannot be taught in a classroom. The junior analyst who wrestled with imperfect data sets did not just learn data analysis; she learnt how to think about data. The trainee solicitor who drafted standard contracts did not just learn contract law; he learned how lawyers read, worry and hedge. That developmental logic is now under considerable strain.

Brynjolfsson et al. (2025b) analysed payroll data covering millions of workers across tens of thousands of firms and found that employment for early-career workers aged 22–25 in the most artificial intelligence (AI)-exposed occupations fell by 16% relative to less-exposed occupations following the widespread adoption of generative AI – while employment for workers over 30 in the same fields remained stable or grew. Davis (2026) provides the structural explanation: AI replicates codified knowledge – the textbook, procedural, rule-based knowledge that constitutes the expert part of entry-level work – but cannot replicate the tacit knowledge acquired through lived experience. In displacing the codifiable tasks, AI removes precisely the work through which tacit knowledge would have been built.

This creates a tension the existing literature has not yet fully addressed. The dominant frameworks treat AI as either an automation or an augmentation technology – asking what tasks it replaces or complements (Raisch and Krakowski, 2021). What these frameworks tend not to ask is what AI does to learning. When a junior worker’s tasks are absorbed by an AI system, the output may still be produced, but the developmental process that would have accompanied producing it is lost. The worker gets the answer without having done the thinking.

The apprenticeship void requires careful differentiation from related prior concepts. Deskilling theory (Braverman, 1974) describes skill degradation within existing roles through task fragmentation – a within-role problem. The apprenticeship void operates earlier, describing the structural removal of the roles and tasks through which foundational capability is first built – a pre-skill problem. Competency traps (Levitt and March, 1988) describe over-reliance on existing competencies that blocks exploration; the void describes the absence of the conditions for competence to develop at all. Automation dependence describes performance decline when tools are withdrawn from workers who have already adapted to them (Parasuraman and Manzey, 2010); the apprenticeship void describes the antecedent mechanism – why those workers never built the underlying capability in the first place. Each existing concept captures a symptom; the apprenticeship void names the structural cause.

This paper introduces the apprenticeship void to describe that loss at a structural level, defined as the systematic absence of learning-rich task engagement that results when AI augmentation displaces the foundational work through which early-career knowledge workers have historically developed expertise. The paper argues that this void is theorised to produce three downstream organisational consequences – synthetic readiness, promotion fragility and pipeline thinning – and that these consequences will become increasingly visible as AI-augmented cohorts progress into roles that require capabilities they were never given the chance to build. Section 2 develops the theoretical foundations. Section 3 presents the framework, construct definitions and formal research propositions. Sections 4 and 5 address contributions and discussion. Sections 6 and 7 cover limitations and implications, and Section 8 concludes.

Lave and Wenger (1991) argue that knowledge is not something a person carries waiting to be applied; it is constructed through participation in practice. The novice learns by doing the work herself – initially in peripheral, supervised ways and gradually taking on more complex and consequential tasks as competence grows. This process of legitimate peripheral participation is not incidental to expertise; it is its primary generative mechanism. Eraut (2000) reinforces this in professional settings, showing that tacit dimensions of professional knowledge are acquired almost entirely through workplace participation rather than formal instruction.

Collins et al. (1989) extend this through cognitive apprenticeship theory, demonstrating that expertise develops through modelling, coaching and scaffolding within real task environments. The critical point is that this depends on the apprentice actually performing the tasks – not reviewing their outputs. The foundational tasks of any knowledge profession are not merely a source of outputs; they are the medium through which expertise is formed.

Faraj et al. (2018) document how algorithmic systems progressively absorb the codifiable components of knowledge work, narrowing the range of tasks available to junior workers. When AI handles research synthesis, first-draft generation and data formatting, the junior worker becomes a reviewer rather than a producer. The learning that occurs in production does not occur in review. Dell’Acqua et al. (2026) confirm this experimentally: while AI raises output quality, its benefits are most pronounced for less experienced workers – consistent with the argument that AI-mediated performance gains do not reflect equivalent capability development.

Bandura’s (1986) social cognitive theory holds that self-efficacy is built primarily through mastery experiences – occasions on which a person completes a challenging task through their own effort. In AI-augmented environments, the experience of producing a competent output may no longer constitute such a mastery experience. When the output is largely AI-generated, the worker’s sense of achievement is not tethered to a corresponding development of capability.

Brynjolfsson et al. (2025a) provide empirical grounding: AI assistance raised the productivity of less experienced customer-support agents by 34% by disseminating expert practices, but the experiential substrate from which those practices emerged could not be transferred. Novice workers produced better outputs; they did not become better workers. Noy and Zhang (2023) corroborate this: generative AI compresses productivity distributions, with the largest gains among the least experienced – raising the same question of whether convergence reflects genuine capability development or AI-mediated performance smoothing. Jarrahi (2018) identifies this as structural: AI systems are designed to improve outputs, not to develop the humans producing them.

If performance feedback becomes systematically unreliable as a signal of capability – because performance is increasingly AI-mediated – the mechanism through which self-efficacy develops is disrupted. Workers may develop high confidence in tasks they have in fact never truly performed. This inflated self-assessment is a predictable structural outcome of AI-mediated work, not a matter of personal failing.

Becker (1964) frames skill development as an investment process: workers and organisations invest time and effort in learning-by-doing, producing a stock of capability that compounds over a career. Davis (2026) demonstrates that AI substitutes for entry-level workers in occupations where the experience premium is low – where the gap in value between inexperienced and experienced workers is relatively small – while complementing experienced workers in high-premium occupations. AI is most effective at eliminating the tasks that entry-level workers have always used to build towards the experience premium. It removes the means by which workers have historically climbed the ladder without removing the ladder itself.

Korinek and Stiglitz (2024) identify significant distributional consequences: the workers most at risk are those who have not yet accumulated experiential capital. Autor (2024) argues that AI can rebuild middle-class jobs by amplifying expertise – but this assumes workers possess the expertise to be amplified, leaving unanswered the question of how that expertise is acquired when the entry-level tasks that historically built it have been displaced. Brougham and Haar (2020) note that long-run capability effects remain invisible in the short-run productivity data driving AI adoption decisions. Argote and Miron-Spektor (2011) reinforce the organisational dimension: when AI displaces the experiential learning of junior workers, it disrupts not only individual development but also the collective learning processes dependent on a steady supply of experienced practitioners.

The four theoretical perspectives are not interchangeable – each contributes something unavailable from the others. Situated learning and cognitive apprenticeship theory supply the mechanism: participation in real task environments is the generative process of expertise, and its removal is the central event the paper explains. Social cognitive theory supplies the psychological dimension: it explains why the performance-capability decoupling is invisible to workers themselves because inflated self-efficacy from AI-mediated success mimics the signal of genuine mastery. Human capital theory supplies the economic dimension: it explains why costs accumulate invisibly in productivity data and why effects concentrate in early-career cohorts. Organisational learning theory supplies the systemic dimension: it explains how individual deficits propagate into collective capability depletion. Together, the four perspectives reveal that AI augmentation simultaneously breaks the mechanism of learning, corrupts the signal of capability, arrests experiential accumulation and depletes the collective knowledge base – a diagnosis unavailable from any single lens.

Taken together, the four perspectives converge on three broken assumptions. The first is that learning occurs through participation – when AI performs foundational tasks, this mechanism is severed. The second is that performance reflects capability – when AI-mediated outputs decouple from individual effort, this signal is corrupted. The third is that experience accumulates over time – when AI absorbs the tasks that generate learning, time accumulates without the associated capability development.

The apprenticeship void names the structural space created when all three assumptions break simultaneously. Nonaka and Takeuchi (1995) make the stakes clear: the socialisation process through which tacit knowledge passes from one generation of practitioners to the next is precisely what AI-mediated task displacement interrupts. Polanyi’s (1966) foundational insight – that we know more than we can tell and that what we cannot tell is acquired through doing – is precisely what AI-driven task displacement removes from the entry-level experience.

The framework traces a causal chain from AI augmentation through the apprenticeship void to downstream organisational consequences, moderated by factors that partially restore the developmental conditions AI adoption disrupts. The chain begins with AI augmentation and entry-level task displacement: as AI is adopted in knowledge-intensive settings, it absorbs the tasks junior workers have historically performed – drafting, summarising, retrieving, synthesising, formatting and iterating (Kellogg et al., 2020). The immediate effect is a productivity gain; the less visible effect is a reduction in learning-rich work available to early-career employees.

This reduction is theorised to produce the apprenticeship void – the structural absence of foundational task engagement resulting from AI-driven task displacement in knowledge work. The term “void” is used deliberately, following its established usage in management theory (cf. Khanna and Palepu, 1997, on institutional voids; Inkpen and Tsang, 2005, on knowledge voids), to denote not a diminishment but a structural absence of a functional mechanism. The void is not a gap that can be bridged by working harder; it is the removal of the conditions under which the relevant kind of development occurs.

Three downstream constructs are theorised to follow from the void. Table 1 presents their defining characteristics across five dimensions, differentiating them from each other and from adjacent constructs in the existing literature.

Synthetic readiness is the individual-level condition that emerges from sustained operation within the apprenticeship void. A synthetically ready worker performs competently in AI-supported environments – meeting benchmarks and producing acceptable outputs – but lacks the independent experiential foundation to perform at the same level without AI assistance. Critically, synthetic readiness is defined by present-state characteristics, not future outcomes: the synthetically ready worker cannot explain the process behind outputs the AI produced, cannot identify AI errors without external verification, and cannot adapt the AI’s approach to novel situations that fall outside its training distribution. These observable characteristics distinguish synthetic readiness from mere performance variation and from automation bias (Parasuraman and Manzey, 2010), which describes over-trust in automation rather than the prior failure of capability to develop.

Promotion fragility describes the condition that develops as synthetically ready individuals advance into more senior roles. Workers are expected to exercise independent judgement and perform where AI assistance is less complete or reliable. Those who reached their current level through AI-mediated rather than genuine capability development will find these conditions more demanding than their prior performance suggested. Promotion fragility extends the Peter Principle (Peter and Hull, 1969) – the observation that workers are promoted to their level of incompetence – by identifying a specific AI-mediated mechanism through which incompetence-at-level is structurally produced rather than incidentally encountered.

Pipeline thinning is the long-run organisational consequence. As successive cohorts progress through synthetic readiness and encounter the limits of promotion fragility, the pipeline through which organisations develop their next generation of deeply experienced practitioners becomes progressively thinner. The organisation appears to function – outputs are still produced, and metrics are still met – but its reserves of deep experiential capability diminish.

Three moderating factors – drawn from existing organisational learning theory and applied here to the apprenticeship void context specifically – are theorised to shape the severity of these consequences. Task redesign (the deliberate restructuring of AI-assisted workflows to preserve learning-rich activities) is proposed to attenuate the void directly: if foundational tasks are preserved in the workflow, the participatory learning mechanism remains partially intact even where AI handles the remainder. Mentorship systems operate through socialisation – direct practitioner-to-practitioner knowledge transfer – partially substituting for the intergenerational transmission that AI-driven displacement interrupts (Nonaka and Takeuchi, 1995). Organisational learning culture determines whether task redesign and mentorship are actually implemented, resourced and sustained over time (Senge, 1990; Argote and Miron-Spektor, 2011).

Figure 1 illustrates the full causal chain from AI augmentation through the reduction in experiential learning opportunities to the apprenticeship void, with moderating factors at the central node and the three downstream constructs below.

The following formal propositions are advanced to guide future empirical research on the apprenticeship void:

This paper makes four distinct theoretical contributions.

Firstly, it introduces the apprenticeship void as an original construct, precisely differentiated from deskilling (which describes within-role skill degradation), competency traps (which describe over-reliance on existing competencies) and automation dependence (which describes post-adaptation performance decline). The void operates earlier than all three – at the level of conditions for capability formation rather than its degradation or its absence after adaptation. The observation that AI displaces entry-level work exists in the labour economics literature (Brynjolfsson et al., 2025b; Davis, 2026), but a management-theoretic account of what this means for capability development has been absent. The apprenticeship void fills that gap.

Secondly, it identifies the decoupling of performance from capability as theoretically significant. Brynjolfsson et al.'s (2025a) finding that AI raises productivity most among novices and Noy and Zhang’s (2023) parallel result in writing tasks both appear to support the view that AI democratises expertise. This paper argues the opposite reading: that AI-mediated gains among novices represent synthetic readiness rather than genuine capability development, and that conflating the two leads to systematic misunderstanding of what AI adoption does to the workforce over time.

Thirdly, it introduces three downstream constructs – synthetic readiness, promotion fragility and pipeline thinning – as theoretically grounded and empirically tractable concepts defined by present-state observable characteristics rather than future predictions. The constructs extend human capital theory (Becker, 1964) by specifying mechanisms through which AI adoption may erode experience accumulation, and extend organisational learning theory (Argote and Miron-Spektor, 2011) by identifying a new pathway through which organisational capability can be depleted without any immediate signal of deterioration.

Fourthly, it reconceptualises AI as a developmental disruptor, distinct from its roles as automator or augmentor. Autor (2024) argues that AI can amplify expertise and rebuild middle-class employment. This paper complements that optimism with a necessary caveat: augmenting performance is not equivalent to developing the expertise that makes performance durable. The developmental dimension – what AI does to the conditions under which capability is formed – is equally important and considerably less examined than the performance dimension.

The framework rests on a distinction that is easy to state but difficult to operationalise: the difference between a worker who can produce a good output and one who has genuinely become capable. The experienced solicitor and the AI-assisted novice look similar in the output, but they are very different in what they can do next – when the AI is wrong, when the situation falls outside its training distribution, or when independent judgement is urgently required.

The apprenticeship void matters because organisations are in danger of systematically underestimating this gap. The standard metrics by which AI adoption is evaluated – productivity, output quality, and cost efficiency – will tend to look positive in the short run even as developmental damage accumulates. The damage does not show up in the performance data of current entry-level workers; it shows up in the performance data of the senior workers they will eventually become.

One might argue that AI-mediated work is itself a new form of learning – that prompt engineering, output evaluation and human-AI collaboration are genuinely complex skills that partially substitute for traditional experiential tasks (Acemoglu and Johnson, 2023). This paper does not dismiss that possibility. The critical question is whether these new competencies produce the tacit knowledge that senior knowledge roles require, or whether they produce a different and more brittle set of skills. Davis (2026) suggests the latter: AI can replicate codified knowledge but not the experiential substrate that makes that knowledge deployable in novel situations. Dell’Acqua et al. (2026) introduced the jagged technological frontier concept: AI assistance improves performance within its frontier, but workers using AI outside that frontier perform worse than those without it. The apprenticeship void describes the long-run consequence of a workforce that has never developed the judgement to know where the frontier lies.

The empirical findings of Brynjolfsson et al. (2025b) are important context. The concentration of employment decline in the 22-to-25 age group represents the progressive exclusion of an entire cohort from the entry-level roles through which experiential capital has historically been built. If a generation of knowledge workers enters organisations primarily in AI-mediated roles from the beginning, the apprenticeship void is not a temporary disruption – it is the new normal.

The moderating factors – task redesign, mentorship systems and organisational learning culture – point towards organisational responses that could limit the severity of these consequences. But their limitations should be acknowledged. Mentorship transmits tacit knowledge between individuals; it cannot fully replicate the tacit knowledge built through doing the work itself. In competitive environments there will be constant pressure to narrow the scope of task redesign. The moderators matter, but they work against a structural force and should not be treated as a complete solution. The causal relationships proposed in this paper remain theoretical propositions requiring empirical validation; the framework offers a structured starting point for that work rather than a settled account.

The most significant limitation is that the framework has not been empirically tested. The constructs and four propositions require operationalisation into validated measures. Developing instruments that capture synthetic readiness as a present-state construct – AI-contingent performance, process explanation capability and error-detection capacity – is a priority for future work.

The framework is most directly applicable to knowledge-intensive occupations – consulting, law, accounting, research and financial analysis – where tacit knowledge development depends heavily on foundational task performance. It is less applicable to highly routinised work or organisations with strong pre-existing developmental structures. Whether it applies with equal force across national and cultural contexts, where apprenticeship traditions and career structures vary, is an open empirical question.

Longitudinal research designs tracking early-career cohorts from initial AI-augmented roles through to mid-career are needed to test the sequential logic the framework proposes and to estimate the lag between task displacement and downstream capability effects. Future research should also examine how variations in AI type, deployment and governance moderate the relationships proposed here.

A fourth limitation concerns a foundational assumption of the framework: that tacit knowledge cannot be meaningfully transmitted through AI-mediated observation alone. This assumption is well-supported within the situated learning tradition (Lave and Wenger, 1991; Eraut, 2000), but it is contested in parts of the knowledge management literature. Some scholars argue that AI systems may capture and transmit certain forms of contextual expertise more effectively than traditional apprenticeship allows (Nonaka and Takeuchi, 1995). If AI can genuinely encode and transfer tacit knowing, the apprenticeship void may be less severe than the framework predicts. The framework takes the more cautious view, consistent with Polanyi’s (1966) position that tacit knowledge is constitutively embodied and cannot be fully externalised, but this remains an empirical question that future research should directly address.

The most pressing practical implication is that the standard metrics used to evaluate AI adoption are insufficient. Productivity, output quality and cost efficiency do not capture what is happening to the developmental pipeline. Organisations serious about long-run capability need to supplement these with indicators of experiential learning depth – measures of the degree to which early-career workers are engaging with learning-rich tasks and building independent judgement alongside AI tools.

This means designing AI-assisted workflows that deliberately preserve foundational task engagement. A junior analyst using AI to summarise a data set should still be required, in some proportion of cases, to produce that summary independently – not as resistance to AI but as deliberate task sequencing in which the comparison between AI-generated and self-generated output drives learning. Mentorship must become more explicit and assessed: when co-working opportunities are absorbed by AI, tacit knowledge transfer cannot remain incidental.

For researchers, the four propositions in Section 3 offer a structured empirical agenda. Survey instruments measuring synthetic readiness should capture three components: AI-contingent performance, process explanation capability and error-detection capacity. Studies examining promotion trajectories in AI-intensive versus less AI-intensive environments would test Proposition 3. Natural experiments – organisations that have adopted AI augmentation at different rates – provide leverage for testing Propositions 1 and 4.

More broadly, the AI and management literature needs a sharper distinction between performance and capability as dependent variables. Theoretical work taking capability development – including its temporal, tacit and social dimensions – as its primary object of analysis is needed to understand what AI adoption is doing to human expertise over the long term (Nonaka and Takeuchi, 1995; Argote and Miron-Spektor, 2011).

The actors most directly affected by the apprenticeship void are: early-career knowledge workers bearing the individual consequences of synthetic readiness; HR and learning professionals responsible for identifying capability gaps; organisational leaders making AI adoption decisions; management educators and university placement offices; and policymakers responsible for workforce development in knowledge-intensive sectors.

There is something quietly important in the distinction between a worker who can produce good outputs and one who has genuinely become good. AI makes that distinction harder to see, and for a time it makes it seem not to matter. The apprenticeship void is the structural condition created when the work through which people become good is systematically removed, leaving behind only the outputs they are supposed to produce.

The downstream consequences – synthetic readiness in individuals, promotion fragility in mid-career, and pipeline thinning over the long run – will not show up in the productivity dashboards that currently drive AI adoption decisions. They will show up later, in the organisations that find themselves thinly staffed with workers who are competent when AI is working and exposed when it is not.

The appropriate response is not resistance to AI, which would be both futile and misguided. It is a deliberate effort to understand what AI adoption does to the developmental conditions within organisations and to redesign those conditions so that AI augments both output and the human capability that, over time, must be able to operate alongside and beyond it. The apprenticeship void is not inevitable. It is the consequence of a particular way of deploying AI – one that can be changed.

The author acknowledges the use of Claude (Anthropic) as an AI writing and research assistance tool during the development and drafting of this manuscript. All theoretical arguments, construct definitions and intellectual contributions represent the author’s own scholarly work and judgement. The final manuscript has been reviewed, revised and approved by the author.

This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.