AI, Entry-Level Work and Career Formation

The copier that would not follow the manual

In the 1980s, the anthropologist Julian Orr followed service technicians who repaired photocopiers.

The corporation had tried to make the work orderly. Its diagnostic documentation linked error codes to known failures and laid out the route to a repair. On one call, the machine supplied plenty of information. It flashed codes and crashed when tested. The codes and the crashes did not agree.

The technician called a specialist whose work combined troubleshooting, supervision and occasional instruction. He was baffled too. Replacing the machine would have cost money and damaged the customer's confidence. So the two men began telling stories.

They recalled other machines and compared symptoms. They ran tests, watched this machine respond, revised their accounts and tried again. The conversation continued through lunch and into the afternoon. After five hours and a dozen anecdotes, the stories, observations and memories converged on a diagnosis. Three months later, Orr heard a shortened version of the false-error-code story being passed among technicians over a game of cribbage.

The repair became part of the occupation's memory.

Orr's ethnography, and John Seely Brown and Paul Duguid's account of it, exposed a gap inside many jobs. Formal instructions describe the expected work. Competent practice also depends on how people respond when the situation refuses to match the description.

That gap is where much entry-level learning happens. A newcomer sees which clues an experienced person notices, how a possible explanation is tested, when a rule is useful, when the case needs escalation and how a technical decision affects a customer or colleague. The finished repair shows the answer. Participation reveals how an answer was reached.

Generative AI can now produce cognitive artefacts such as instructions, code, summaries, classifications and plausible diagnoses in seconds. It can make part of this learning easier. It can also deliver the finished-looking answer before a beginner has encountered the difficulty from which judgement grows.

The question is therefore larger than whether AI removes junior tasks:

When entry-level work changes, how will newcomers gain the experience, judgement, relationships and credibility from which later careers are made?

The short answer

Entry-level work is one of the systems through which a society reproduces capability. It supplies goods and services now, while giving a newcomer access to cases, tools, standards, colleagues, feedback and progressively greater responsibility. It can also produce a wage, a work identity, a network and evidence that other people recognise.

AI can improve that system. It can demonstrate a method, retrieve a precedent, translate unfamiliar language, provide practice, expose a novice to expert patterns and make feedback available at the moment of need. In some settings, less experienced workers have gained most from assistance.

AI can weaken the system too. If an organisation removes the tasks that justified junior headcount, gives the remaining cases to experienced workers, or lets beginners submit generated work they cannot inspect, it may preserve today's output while reducing tomorrow's supply of people able to handle exceptions and carry responsibility.

Neither outcome follows from exposure alone. The result depends on adoption, labour demand and the design of work: which tasks remain available, whether newcomers can observe practice, what they must attempt themselves, how feedback arrives, who checks understanding and how increasing competence leads to recognition and mobility.

A useful question for any changed junior role is:

What is the newcomer producing, what are they learning, who can observe it, and what responsibility can they earn next?

Entry-level describes a position, not a kind of person

An entry-level worker may be 17 or 57. They may be entering paid work for the first time, changing occupations after twenty years, returning after care or illness, translating overseas experience into a new institutional setting, or beginning a profession after extensive formal study.

Age, tenure and level overlap, but they are not interchangeable. A 23-year-old can be highly experienced in a family enterprise. A senior manager can be a novice in clinical practice. A qualified migrant may know the work and still lack local recognition, vocabulary or access. An apprentice is simultaneously a worker and a learner by design.

This page uses entry-level work for a position near the boundary of an occupation, organisation or practice where a person has limited recognised experience and responsibility. Early career refers to a period in a person's trajectory. Junior refers to a relative level of authority or expected independence. These are Guidebeam working distinctions, because labour datasets, employers and professions classify them differently.

The distinction matters for evidence. One influential United States study uses ages 22 to 25 as its early-career group. That creates a measurable cohort; it does not define every entrant. An Australian report examines people aged 15 to 24, young graduates aged 20 to 24 and occupations with different exposure scores. Those measures answer related questions from different angles.

It also matters for fairness. If organisations assume that every beginner is young, unconstrained and able to accept unpaid experience, they narrow access before any assessment of capability occurs. Career formation then reflects financial support, social networks and institutional recognition as well as what a person can learn.

One job, several forms of formation

The first months in a role can look inefficient. A newcomer asks questions, works slowly, needs review and sometimes makes a recoverable mistake. Measured only against today's output, an experienced worker or configured system may appear to be the obvious substitute.

The comparison misses what the arrangement is producing over time.

This is a career-formation map, not a claim that every junior job provides all seven. Many entry roles have always been precarious, repetitive, unsafe, underpaid or deliberately denied progression. Some people acquire capability through education, community practice, volunteering or independent work. Some routine tasks teach little after they are mastered.

The map instead shows why a task inventory is incomplete. A task can have modest productive value and substantial learning value. Another can consume time without supplying useful variation or feedback. Automation may remove drudgery, remove rehearsal, or do both at different moments.

The relevant unit is the learning pathway through work, not the isolated task.

From the edge of a practice

In 1991, Jean Lave and Etienne Wenger gave a name to a pattern they had seen across apprenticeships and communities: legitimate peripheral participation.

Newcomers begin at a periphery where their participation is limited but real. They have access to practitioners, activities, artefacts and the social life of the work. Over time, their participation changes. Learning includes new knowledge and skill, but also movement towards fuller membership of a practice.

The phrase contains three safeguards.

Legitimate means the newcomer is allowed to be there. They can ask, observe and contribute without being treated as an intruder. Peripheral means their responsibility is bounded while they gain access to the whole. Participation means they are involved in the practice rather than merely receiving information about it.

Brown and Duguid described apprentice butchers whose position failed this test. They were present in the workplace but could not watch the journeymen cut and saw meat. Formal training could name the activity. The arrangement denied access to its performance.

Later research made the organisational choice clearer. Alison Fuller and Lorna Unwin studied apprentices and older workers in the English steel industry. Their expansive–restrictive continuum distinguished workplaces that gave apprentices broad participation, planned on- and off-the-job learning, feedback and a route for progression from those that confined them to a narrow station and immediate output.

The job title apprentice did not guarantee apprenticeship-quality learning. Participation, personal development and institutional arrangements did.

The same is true of jobs without a formal apprenticeship contract. Michael Eraut's research with early-career nurses, engineers and accountants found that much workplace learning was informal, combining learning from other people with learning from experience. His 2004 synthesis emphasised an interaction among challenge, support and confidence. A challenge can extend capability when the person has enough support to attempt it. Remove the challenge and learning stalls; remove the support and the person may retreat from the work.

AI changes both sides. It can supply support at the point of difficulty. It can also remove the difficulty, conceal it behind an answer or increase the volume of work without increasing the time for reflection.

The expert's traces in the novice's screen

A large customer-support deployment shows the constructive possibility.

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied a generative-AI assistant introduced among 5,172 agents at a business-software company. The system watched text conversations and suggested replies, drawing partly on records of successful past interactions. Agents could accept, edit or ignore the suggestions.

In the peer-reviewed study, access increased issues resolved per hour by 15 per cent on average. Less experienced and lower-skilled agents gained more. Agents with two months of tenure and access to the system performed about as well as untreated agents with more than six months of tenure. During unexpected outages, more-exposed agents retained some gains relative to their own pre-AI performance, evidence consistent with learning rather than momentary copying alone.

The system had become a carrier of experienced practice. It could place a useful question, phrase or diagnostic route in front of someone who had not yet accumulated the same history.

That does not make the organisation's learning system automatic. The deployment existed inside a structured job. Agents handled real customer problems, remained responsible for the conversation, received weekly one-to-one feedback from managers and sometimes worked without a suggestion. The researchers could observe output and medium-run performance in one firm. They could not observe wages, total labour demand or the long-run composition of hiring.

There is another dependency. The assistant learned from records created by workers, with top performers' conversations given additional weight. If future workers mostly follow existing suggestions, the organisation may produce fewer novel examples from which the next system and the occupation can learn. The study's authors raise that problem directly.

The case shows AI acting as a bridge between accumulated practice and a newcomer. It does not show that a bridge can replace the place, people and responsibility on either bank.

When performance and learning part company

Immediate performance is tempting evidence of formation. A beginner produces acceptable code, a good report or a resolved case, and the output seems to demonstrate capability.

Assistance complicates the inference. The observed result belongs to a configuration of person, tool, instructions, time and review. To infer independent capability, we need evidence of what the person can understand, adapt, debug or perform when the configuration changes.

Judy Hanwen Shen and Alex Tamkin tested that difference with 52 software developers, most of them junior, who used Python regularly but did not know the Trio asynchronous-programming library. Participants were randomly assigned to learn and use the library with web search alone or with web search plus a coding assistant.

The 2026 preprint reports no statistically significant average completion-time gain. On a subsequent unaided quiz covering concepts, code reading and debugging, the AI-assisted group scored 4.15 points lower on a 27-point scale. Screen recordings revealed different patterns within the assisted group. People who delegated the code tended to finish more quickly and learn less. Those who asked conceptual questions or tried to explain generated code preserved more understanding.

This is a short, developer-specific experiment, not a forecast for occupations. It does show that a completed task and acquired skill can move in different directions.

A second recent experiment prevents the warning from becoming a law. Zara Contractor and Germán Reyes randomly assigned 210 undergraduates to learn an unfamiliar topic and write an analytical essay with or without generative AI in proctored sessions. Their July 2026 preprint reports a 0.27-standard-deviation gain on an immediate unaided knowledge test that persisted one week later. Delayed writing gains were stronger among students who used AI for explanations; short-run gains among those using it to generate prose disappeared when assistance was removed.

The studies involve different people, domains, tools, tasks and outcomes. Together they replace a false choice with a design question. AI can answer, demonstrate, prompt, critique or coach. The learner can delegate, inspect, practise, compare or explain. Those interactions do not produce the same formation.

This is why the capability page separates a successful performance from the knowledge, skill, judgement and support that produced it. An AI-assisted work sample may be relevant evidence for an AI-assisted job. It cannot, by itself, establish unaided understanding or readiness to supervise the system.

The contested bottom rung

Learning inside a job becomes impossible when the job is never offered.

Erik Brynjolfsson, Bharat Chandar and Ruyu Chen examined monthly records from ADP, the largest payroll processor in the United States. After restrictions, their sample contained between 3.5 and 5 million workers a month from firms present throughout January 2021 to September 2025. They compared employment trends by age and occupational exposure to generative AI.

Their November 2025 working paper found a 16 per cent relative employment decline for workers aged 22 to 25 in the most exposed occupations after controlling for shocks common to each firm at each time. The divergence was concentrated in occupations where observed AI use was classified as more automative. Base-pay trends differed much less than employment.

The result deserves attention because it uses actual payroll employment rather than task capability alone. It also has firm boundaries. ADP's clients do not exactly reproduce the United States economy. Recorded job titles were unavailable for some workers. Exposure measures join occupation task descriptions to model or usage classifications. The estimate is relative, not a 16 per cent fall in all young employment. Most importantly, the authors say that factors other than generative AI may contribute. Their evidence is consistent with an AI effect; it does not isolate every cause.

The possible mechanism connects directly to career formation. If codified and checkable junior tasks can be generated while experienced workers retain relationship, process and exception work, an organisation may reduce the intake through which people once accumulated that experience. The current expert remains productive. The future expert has fewer places to begin.

That mechanism is a hypothesis requiring direct organisational evidence. A firm could respond differently: hire more entrants because assistance lowers training costs, expand output, redesign progression or create new tasks. The technology page explains why capability, adoption, task redistribution, job redesign and employment cannot be collapsed into one step.

Australia is not showing the same current pattern

Evidence from Australia supplies an important counterweight.

Jobs and Skills Australia's 2025 Gen AI Capacity Study found no clear evidence at that time that generative AI had reduced entry-level roles. Its work on occupational pathways described a mobile labour market in which about 15 per cent of workers change roles from one year to the next, while stressing upskilling, redeployment and role change.

The Australian Department of Employment and Workplace Relations then examined occupation-level data through February 2026. Its 2026 report found that employment among people aged 20 to 24 had grown slightly faster than employment among those aged 25 and above since late 2022. The unemployment rate for graduates aged 20 to 24 was 5.4 per cent on a four-quarter rolling basis, low by the report's pre-pandemic comparison.

The same report found slower recent growth in more AI-exposed occupations and sensitivity across measures and model specifications. It did not treat the relatively solid youth indicators as proof that no junior task or hiring channel had changed.

The contrast with the United States study is substantive. Countries differ in occupational structure, institutions, adoption, population growth and data. The age bands and exposure measures differ too. One set of results should not be used to erase the other.

As of July 2026, the defensible conclusion is narrower than either alarm or reassurance: there are credible signs of pressure on some exposed young-worker employment in the United States, while broad Australian indicators do not show a comparable deterioration in entry-level or young-graduate employment. Both records are early. Neither reveals whether the people hired today are receiving the formation they need.

A career-formation chain

A robust entry pathway gives a newcomer movement through several kinds of participation:

access → legitimate contribution → representative cases → feedback and explanation → expanding judgement → recognised responsibility → mobility

AI can alter every link. It may open access through translation or lower training costs. It may reduce the number of available positions. It may widen the cases a novice can attempt, or divert all difficult cases to a senior worker. It may provide immediate explanation, or generate an answer that hides the reasoning. It may record a person's growing capability, or make authorship and contribution harder to interpret.

Entry-level work · Formation

What happens to the pathway when a junior task changes?

Current output and future capability are different outcomes. AI can strengthen one link in a pathway while weakening another.

Access→Participation→Representative cases→Feedback→Judgement→Recognised responsibility→Mobility

Career-formation pathways for three entry-level cases

Formation stage	Customer-support agent	Junior software developer	Trainee accountant
Access	Translation and answer support may lower entry barriers; position numbers still matter.	Scaffolds may help a newcomer attempt code; hiring access still comes first.	Document assistance may lower routine load; supervised trainee places remain the gate.
Participation	Handle live cases with protected escalation, not only observe generated replies.	Work in the shared codebase with review, not an isolated answer box.	Join real engagements within permission and confidentiality boundaries.
Representative cases	Retain varied requests and exceptions rather than only repetitive tickets.	Attempt changes that expose architecture, users and failure modes.	See transactions, controls, judgement calls and client context.
Feedback	A senior explains why a response works; generated text alone can hide reasoning.	Review comments and tests connect changes to consequences.	Supervisor review connects workpapers to standards and evidence.
Judgement	Practise when to answer, clarify, escalate or refuse.	Practise trade-offs, debugging and when not to accept a suggestion.	Practise materiality, scepticism and when evidence is insufficient.
Responsibility	Gain bounded ownership of cases with an accountable escalation path.	Own a small change through review and release.	Sign off only within delegated authority while the supervisor remains accountable.
Mobility	Observed judgement opens more complex queues or roles.	Demonstrated work supports broader systems responsibility.	Recognised experience supports progression toward wider professional duties.

Task removed

Output may continue while practice disappears.

Answer generated

Completion is not evidence of learning.

AI explanation

Useful feedback still needs checking and context.

Protected practice

A novice can attempt work within bounded risk.

Supervised escalation

Authority stays visible while judgement grows.

No progression

A role can produce output without opening a next step.

A seven-stage pathway runs from access through participation, representative cases, feedback, judgement, recognised responsibility and mobility. A table compares a support agent, junior developer and trainee accountant, keeping immediate output, learning evidence, authority and the next opportunity separate.

This is Guidebeam's working synthesis, not a universal career ladder. Careers branch, pause and cross institutions. People enter practices through many routes. The chain identifies dependencies that become easy to lose when work is redesigned one task at a time.

It also changes what should be measured. Throughput, cost and error rates describe current production. A formation system also needs evidence of case exposure, unaided understanding, quality of feedback, progression in responsibility, retention, access and later mobility. A dashboard of completed junior tasks can look efficient while the pathway behind it is narrowing.

The OECD's study of training in enterprises groups informal learning into learning by doing, learning from others and keeping up to date. It also finds that management support and some institutionalisation help create a learning environment. Work does not become developmental simply because a beginner performs it.

Who pays for the missing experience?

When an employer stops providing a first opportunity, the need for experience does not disappear. It moves.

A person may buy further education, accept an unpaid placement, build a portfolio without income, volunteer, rely on family contacts or compete for the smaller number of organisations still willing to train. People with money, time, recognised credentials and professional networks can carry that transfer more easily.

The quality of the substitute matters. The International Labour Organization's review of internships found that paid, well-designed placements were associated with better later employment than unpaid placements; mentorship, employee-like conditions and sufficient duration also mattered. Its wider review of work-based learning found the strongest case for structured arrangements that join work and learning rather than treating any placement as inherently developmental.

The ILO's 2020 youth report described entry jobs as places where young people learn their preferences and capabilities while acquiring work habits, skills and networks. It also found that previous automation risk in an entry occupation was associated with more difficult later transitions across 22 OECD countries. That historical result predates the current wave of generative AI, but the career mechanism is still relevant: lost access can affect more than the first pay packet.

Current policy evidence does not supply a single replacement. The 2026 ILO–World Bank review of 228 youth-employment studies across 62 countries found positive average effects from active labour-market programmes, with substantial variation by programme and context. Training, employment services, wage subsidies and work experience are different interventions. None recreates an occupational community merely by enrolling a participant.

This is the distributional stake in entry-level redesign. An organisation may capture an immediate productivity gain while households, education systems and public programmes absorb more of the cost of producing future capability.

The senior worker who has not been formed yet

The Xerox technicians' five-hour repair was inefficient if viewed as a single closed ticket. It was productive in at least three ways. The customer received a working machine. The technicians increased their understanding. Their story entered a shared repertoire that helped other people diagnose later failures.

AI can participate in each of those achievements. It can help solve the case, reveal similar cases and preserve an account for the next worker. It can also make the first result so easy to count that the other two disappear from view.

The stakes are wider than a threatened bottom rung. Career formation is how occupations renew judgement, how organisations discover who can carry responsibility, and how people turn capability into recognised opportunity. The route has never been equally open or reliably developmental. That history is a reason to design it better, not to preserve every old junior task.

The durable question is:

After the work is done, who has become more capable of doing the next, harder thing?

If the answer is nobody, the system may be consuming expertise accumulated under an earlier design. If the answer includes newcomers who have participated, practised, been corrected and earned wider trust, AI may be helping to build the next generation rather than borrowing silently from the last.

---

Sources and further reading

• Acemoglu, Daron and Pascual Restrepo. “Automation and New Tasks”. 2019.
• Brown, John Seely and Paul Duguid. “Organizational Learning and Communities-of-Practice”. 1991.
• Brynjolfsson, Erik, Bharat Chandar and Ruyu Chen. Canaries in the Coal Mine?. 2025.
• Brynjolfsson, Erik, Danielle Li and Lindsey Raymond. “Generative AI at Work”. 2025.
• Collins, Allan, John Seely Brown and Ann Holum. “Cognitive Apprenticeship”. 1991.
• Comyn, Paul and Laura Brewer. Does Work-Based Learning Facilitate Transitions to Decent Work?. ILO, 2018.
• Contractor, Zara and Germán Reyes. “Experimental Evidence on the Learning Impact of Generative AI”. 2026.
• Australian Department of Employment and Workplace Relations. AI and Employment in Australia. 2026.
• Eraut, Michael. “Informal Learning in the Workplace”. 2004.
• Fuller, Alison and Lorna Unwin. “Learning as Apprentices in the Contemporary UK Workplace”. 2003.
• International Labour Organization. Generative AI and Labour Markets in ASEAN. 2026.
• International Labour Organization. Global Employment Trends for Youth 2020. 2020.
• ILO and World Bank. The Impact of Active Labour Market Programmes on Youth. 2026.
• Jobs and Skills Australia. Our Gen AI Transition: Labour Market Dynamism. 2025.
• Lave, Jean and Etienne Wenger. Situated Learning: Legitimate Peripheral Participation. 1991.
• OECD. Training in Enterprises. 2021.
• O'Higgins, Niall and Luis Pinedo. Interns and Outcomes. ILO, 2018.
• Orr, Julian E. Talking About Machines. 1996.
• Polanyi, Michael. The Tacit Dimension excerpt. 1966.
• Shen, Judy Hanwen and Alex Tamkin. “How AI Impacts Skill Formation”. 2026.