AI Is Breaking the Apprenticeship Model

A firm can become more productive while weakening its ability to produce experienced people.

Consider a professional-services team that uses AI to accelerate research, prepare initial analyses, and draft routine materials. Senior professionals can complete assignments with less junior support. The immediate economics may be attractive. But a second question sits behind the staffing decision: where will the next generation of senior professionals acquire its experience?

Junior work has often served two purposes. It produces something the organization needs, and it puts less-experienced people close enough to real problems to learn. AI can change the economics of the first function without providing a replacement for the second.

That is the apprenticeship problem. The risk is not that every entry-level role disappears, or that an established career model should be preserved unchanged. It is that organizations remove parts of their development system without recognizing that they are doing so.

Normatic's position is that AI transformation must redesign how people learn to do the work that remains, alongside how that work gets done. A future-state workflow is incomplete if it improves today's output while leaving tomorrow's expertise to chance.

The employment evidence raises a broader organizational question

Stanford's August 2026 Canaries in the Coal Mine? study examines ADP payroll data through June 2026. Employment among 22–25-year-olds in AI-exposed occupations was 19% below a comparison in which it had matched less-exposed peers' growth. Older workers showed no comparable gap. Weaker hiring, rather than increased separations, accounts for the divergence.

Declines concentrate in automation-oriented occupations; complementary usage shows different patterns. The authors also find suggestive associations between codified knowledge and weaker young-worker employment, and between tacit knowledge and stronger experienced-worker employment.

These are descriptive findings, not causal estimates. Education controls weaken the patterns; some divergence predates generative AI; national survey comparisons are less pronounced. The paper finds no widespread economy-wide displacement. Its 19% figure is a relative shortfall, not the proportion of junior jobs eliminated by AI. August 2026 paper.

Age is also an imperfect guide to experience: a career changer can be a novice at forty. For organization design, the question is which opportunities to acquire competence are changing, whatever the worker's age.

The employment evidence makes that question timely. It does not establish that organizations are already exhausting their future supply of experts. The expertise-pipeline argument is a forward-looking implication to investigate within actual workflows.

Junior work has bundled production with learning

In a familiar development pattern, a junior employee starts with relatively structured tasks, observes experienced colleagues, and encounters recurring situations. With guidance, repetition becomes comparison: why did the same approach work in one case and fail in another? The employee begins to recognize exceptions, understand organizational context, and make decisions with less supervision.

The productive assignment supplies the setting for that learning. A research task brings an analyst into contact with conflicting evidence. Preparing a model reveals which assumptions matter. Reviewing a document creates an opportunity to discuss why apparently standard language is unsuitable in this particular transaction.

The economically useful work and the development of expertise have been bundled together. AI may begin separating them.

This is not a claim that all junior work teaches well. Some tasks provide little variety, feedback, or connection to decisions. Some employees receive rich exposure while others remain confined to repetitive support. Nor does time served automatically produce expertise. The quality of participation matters.

Stephen Billett's foundational research on workplace learning emphasizes guidance, access to progressively different activities, and the learner's engagement. That framework helps explain why being employed near experienced people is insufficient: the work must offer meaningful participation and support. This is an established learning lens, not evidence about current AI effects. Toward a Workplace Pedagogy.

The organizational consequence is specific. Removing a task can remove both its production cost and the opportunities for observation, practice, and feedback that surrounded it. A task-level business case will capture the former more readily than the latter.

Exhibit 1. A task can produce an output and a learning opportunity

Exhibit 1: Normatic illustrative analysis. Practices differ across organizations and professions. Automating a task does not necessarily remove its learning value if participation and feedback are deliberately retained.

The experience paradox

An organization may use AI to reduce the work for which it needs novices while placing greater value on people who already possess judgment. Yet some of those experienced people acquired their judgment through the very work now being automated.

Our earlier Perspective, Intelligence Is Becoming Abundant. Judgment Becomes the Scarce Resource., considered how organizations use judgment. This article asks how they replenish it.

Judgment does not appear fully formed when someone receives a senior title. It develops through making distinctions, testing explanations, receiving correction, and encountering the consequences of decisions. A useful learning path includes ordinary cases that establish a baseline as well as unusual cases that challenge it.

This creates an important qualification to The Future Organization Runs on Exceptions. If people increasingly handle exceptions, organizations must ensure they have learned enough about normal operation to recognize what makes a case exceptional. Sending novices directly into the hardest unresolved problems is not an adequate apprenticeship design.

Tacit knowledge should not become a comforting label for work assumed to be permanently beyond AI. Some experience can be represented in examples, records, and tools; capabilities will continue to change. The practical concern is narrower: wherever people remain responsible for consequential decisions, they need a credible route to the competence those decisions require.

The paradox is therefore conditional but consequential. If junior participation contracts while demand for experienced judgment persists, the organization needs another way to create that judgment. Without one, short-term efficiency could conceal a future capability deficit.

Preserve learning, not busywork

There is little value in requiring a junior employee to spend hours reformatting a document merely because previous generations did so. Repetition without feedback can consume time without improving understanding. Automation can remove frustration and release capacity for more useful participation.

The right design question is which learning experiences deserve preservation when the task changes. Did the employee learn to question a source, recognize an anomaly, or understand a client's priorities? Or was the activity largely mechanical? Those differences determine whether a replacement learning mechanism is needed.

Take an illustrative reconciliation workflow. If software resolves straightforward matches, a developing professional need not repeat those matches manually to prove diligence. But the person may still need to understand the transaction flow, explain representative matches, investigate selected discrepancies, and follow an issue through to resolution. The learning objective survives even when its delivery changes.

The same logic applies to producing an analysis. AI can assemble an initial view while the learner is responsible for identifying assumptions, testing a competing explanation, and discussing the conclusion with a reviewer. Less manual preparation can create more time for reasoning, provided the workflow actually allocates that time.

Simply giving the saved hours back as additional production targets will not necessarily produce this result. The organization has to decide how much capacity to devote to development and what evidence of progress it expects.

AI can shorten the path to competence—or bypass it

There is evidence for both opportunity and risk, although the studies examine different settings.

In Generative AI at Work, published in The Quarterly Journal of Economics in 2025, Brynjolfsson, Danielle Li, and Lindsey Raymond studied 5,172 customer-support agents. AI assistance increased issues resolved per hour by 15% on average, with larger benefits for less-experienced and lower-skilled workers. The authors also found evidence consistent with learning, including performance during periods when the tool was unavailable, although those outage estimates were noisy. This is evidence from one workplace using an earlier system, not proof that every AI assistant accelerates professional development. Published study.

A different result appears in Shen and Tamkin's 2026 preprint How AI Impacts Skill Formation. Randomized experiments on learning an unfamiliar programming library found weaker subsequent understanding and debugging with AI assistance, without a significant average efficiency gain. More cognitively engaged interaction patterns were associated with better learning, but those patterns were not themselves randomized. The study tests acquisition of a bounded technical skill; it does not establish long-term deskilling across professions. February 2026 version.

The findings should not be averaged into a universal verdict. They concern different work, systems, participants, and outcomes. Together, they make a useful distinction visible: improved performance with a tool and development of the person using it are separate outcomes to evaluate.

Normatic's interpretation is that AI-assisted apprenticeship can be designed around active engagement. A system might present a case, ask for an initial diagnosis, offer a counterexample, and provide feedback grounded in reviewed material. It might help a learner compare several explanations or explore how a decision changes under different conditions.

These are design possibilities, not guaranteed effects. A fluent explanation can be wrong. A simulated case can simplify the relationships and consequences that make real work difficult. Coaching material needs an appropriate evidence base, and learners need opportunities to challenge it with capable people.

The aim is not independence from every tool. It is enough understanding to use tools responsibly: to recognize uncertainty, check relevant evidence, explain a decision, and know when assistance or escalation is needed.

Exhibit 2. Automating production leaves a separate learning decision

Experience needs a deliberate design

The starting point is the capability the organization needs people to acquire. A generic ambition to develop judgment is difficult to operationalize. A more useful goal might be recognizing when an apparently routine request contains conflicting commitments, diagnosing a production failure, or identifying a material weakness in an argument.

From there, the organization can design a progression of participation. Early exposure might combine reviewed cases and observation of senior decisions. Later work might require the learner to propose a course of action, defend it, and compare it with the eventual outcome. Decision authority can then expand within defined boundaries as evidence of readiness improves.

Simulations can provide access to unusual situations without exposing live work to avoidable error. Rotations can reveal how one team's output affects another. Post-decision reviews can make expert reasoning visible, including what was uncertain at the time. Mentoring can connect these experiences to the individual's development needs.

No single mechanism replaces apprenticeship. A library of cases without discussion can become another content repository. Shadowing without an opportunity to reason can remain passive. Responsibility without support can become abandonment. The mechanisms need to connect to real work and to each other.

Exhibit 3. Design the experience and the evidence of readiness

Exhibit 3: Normatic illustrative framework. The progression, evidence threshold, and permitted actions should fit the profession and consequences. Completion of training alone does not establish readiness.

An illustrative software team might automate routine implementation while retaining a developmental assignment around each selected change. The junior developer explains the intended behavior, identifies a failure case, reviews the generated code, and proposes a test. A senior colleague reviews the reasoning and later discusses what happened in operation. The purpose is to make learning observable while benefiting from automation.

The design should also protect access. If only the fastest employees receive difficult cases and senior attention, others may never acquire the experience required to advance. Case allocation, mentoring capacity, and opportunities to make supervised decisions are organizational choices. They should not depend entirely on who happens to be present when a problem occurs.

Workforce planning needs two views of capacity

A production plan asks how much work the organization can deliver with the people and systems it has. A capability plan asks which forms of expertise it will need, where they will come from, and how long development may take. AI changes both calculations, but not necessarily at the same rate.

A role may contribute less billable or measurable output after automation while the organization still needs a route into that profession. That does not establish a case for retaining the role unchanged. The response might be a redesigned junior role, a structured development placement, a different mix of work and study, or a partnership that provides supervised experience.

The investment should be explicit. Its cost includes expert coaching time, suitable case access, assessment, and the operational support that makes supervised participation possible. Its value depends on future capability needs, retention, mobility, and whether learning transfers into useful performance.

Measures should therefore go beyond training attendance or the volume of AI-assisted output. Useful evidence could include performance on unfamiliar cases, quality of explanations, recognition of limits, time to demonstrated readiness, and the amount of support still required. Where appropriate, selected assessments can limit AI assistance to distinguish the person's understanding from the tool's contribution.

These measures should guide development rather than encourage premature independence. A learner who escalates a consequential uncertainty appropriately may be demonstrating better judgment than one who completes the task confidently and incorrectly.

The ownership is shared. Business and workflow leaders understand the capabilities required by the work. Experienced practitioners provide guidance and assess performance. HR and learning teams help establish pathways and access. Transformation leaders need to connect these responsibilities to the workflow redesign and its business case.

A more senior firm still needs a route to seniority

Many professional-services models have combined a broad junior base with smaller managerial and senior layers. Junior production helps deliver current engagements while assignments, review, and client exposure contribute to development. The balance varies considerably by profession and firm.

If AI absorbs a substantial share of the production work, some firms may operate smaller teams with more senior expertise and greater technological support. That can be a rational delivery model. It leaves the question of replenishment unresolved.

A firm that depends increasingly on experienced professionals should be able to explain whether it develops them, recruits them, or combines the two. External hiring can meet genuine needs, but it does not eliminate the wider requirement for someone to create experience. If many organizations reduce entry routes at once, they could intensify competition for an expertise pool that is not replenishing at the same pace. This is a scenario to monitor, not a demonstrated economy-wide outcome.

Everyone wants experience. Someone still has to create it.

There are countervailing possibilities. Lower delivery costs could expand demand and create new forms of junior participation. AI could let developing professionals attempt richer work earlier. New training arrangements could produce competence more efficiently. The future supply of expertise depends on these responses as well as on hiring decisions.

For an individual firm, the actionable issue is whether its own operating model contains a credible development pathway. A partner's productivity gain should not automatically become a smaller training budget. Nor should every managerial hour released by automation become a wider span of control. As argued in AI Will Change the Economics of Management, coaching and coordination compete for attention, and their demands differ.

This applies to Normatic as much as to its clients. An offer built around senior transformation capability should help strengthen the client's ability to understand, own, and sustain change. Joint delivery and capability transfer need opportunities to exercise judgment, not only documentation of what the senior team has done.

Redesign work and learning together

For any material workflow change, leaders should identify the tasks being removed or altered, the learning those tasks previously enabled, and the capabilities people will still need. They can then decide which experiences to retain, replace, or improve, and how readiness will be demonstrated before responsibility expands.

This connects apprenticeship to The Enterprise AI Moat Is Not the Model. An organization's ability to develop sound judgment alongside its systems may become a valuable capability in its own right. Its value would come from better decisions and continuing renewal, not from maintaining familiar job titles or claiming that human expertise is beyond technological change.

The opportunity is to make apprenticeship more intentional and potentially more effective. AI can remove low-value effort, widen access to examples, and support practice. Organizations must ensure that those possibilities become part of the design rather than assume that learning will take care of itself.

If AI removes the work through which novices historically became experts, organizations will need to redesign apprenticeship as deliberately as they redesign workflows.

Sources

Employment evidence

• Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, August 2026. Stanford publication page.

AI-assisted performance and learning

• Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, Generative AI at Work, The Quarterly Journal of Economics, 2025. Workplace field evidence; the published version reports 5,172 agents and a 15% average productivity improvement. Earlier working-paper estimates differ.
• Judy Hanwen Shen and Alex Tamkin, How AI Impacts Skill Formation, preprint, revised February 1, 2026. Experimental evidence on a specific programming skill; exploratory interaction patterns do not establish which learning design causes better outcomes.

Foundational workplace-learning framework

• Stephen Billett, Toward a Workplace Pedagogy: Guidance, Participation, and Engagement, Adult Education Quarterly, 2002. Used for its framework of guided participation and access to workplace learning, not as evidence of current AI capability.

Normatic interpretation

The apprenticeship argument, profession examples, three exhibits, proposed development arrangements, and experience-trap scenario are Normatic analysis. The sources do not establish a causal, economy-wide collapse of junior employment or a measured future shortage of experts. The Perspective extends Normatic's existing positions on judgment, exceptions, management, and organizational capability. Sources checked September 29, 2026.