The Apprenticeship Bottleneck: AI, Article Clerks, and the First Rung of Work in South Africa

By Dr Riaan Steenberg

A junior accountant sits with a set of working papers. The task is not glamorous. There are reconciliations to test, invoices to inspect, exceptions to follow up, notes to clean, schedules to prepare, and a senior waiting for the file to make sense.

To the firm, much of this work looks routine. To the trainee, it is not routine at all. It is the place where judgement starts to form.

The trainee learns what a normal transaction looks like, when a number smells wrong, which client explanation is plausible, and which one needs evidence. They learn that a clean spreadsheet can still hide a weak control.

This is why the AI question is more serious than the usual debate about job losses. The risk is not only that artificial intelligence will remove some entry-level tasks. The deeper risk is that it may remove the work through which entry-level people become professionals.

South Africa cannot treat this as a narrow technology issue. We are already dealing with a weak first rung of the labour market. In the first quarter of 2026, Stats SA reported an official unemployment rate of 32.7%. Among young people aged 15 to 34, unemployment stood at 45.8%. That means almost half of the young people who are economically active and looking for work cannot find it.

In that context, the article clerk, the trainee accountant, the junior analyst, the audit assistant, the legal candidate, and the graduate consultant are not just labour categories. They are structured ways for young people to convert education into employability.

If AI weakens those routes, the social consequence will be larger than a productivity gain on a partner's dashboard.

The Work That Looks Easy From Above

Many senior people forget how much they learned from boring work.

Reconciliations, drafting minutes, preparing schedules, checking invoices, cleaning datasets, summarising documents, tagging exceptions, updating working papers, and capturing evidence can look like low-value work once a person has mastered the domain. It is easy to say that a machine should do it.

In many cases, a machine should do it.

There is no virtue in forcing a graduate to copy numbers from one system to another if the task can be automated reliably. There is no professional formation in hours of formatting. There is no moral argument for protecting inefficient work for its own sake.

But there is a formation problem hidden inside the efficiency argument.

Entry-level work is often badly designed, but it is not meaningless. The first layer of professional work teaches pattern recognition, vocabulary, the difference between theory and an actual client file, and the rhythm of deadlines, evidence, escalation, review, and correction.

The young accountant does not become useful only because they know accounting standards. They become useful because they have seen enough messy cases to understand how standards behave in practice.

AI can now absorb a growing share of this messy first-pass work. It can draft audit planning notes, summarise contracts, compare policy documents, identify anomalies, generate working paper narratives, prepare first-draft tax research, clean data, reconcile fields, and write a coherent explanation around a variance.

That creates an obvious temptation. A firm can keep the senior people, give them better tools, and reduce the number of juniors required to produce the same output.

For a firm under margin pressure, that looks rational. For a profession, it creates a pipeline problem.

Article Clerks Are A Useful Test Case

The South African accounting profession gives us a clear example because the training model is explicit.

SAICA's CA2025 competency framework sets out the competencies expected of an entry-level CA(SA) at the end of the pre-qualification process. The 2025 training regulations require training offices to provide a sufficient range and depth of relevant work so trainee accountants can obtain the required training and practical experience. The training office must also provide experience of increasing complexity and the necessary range and depth in prescribed tasks and competencies.

This matters. The model is not simply "hire cheap labour for three years." At its best, articles are a structured apprenticeship. The trainee moves from basic work to more complex work. They receive feedback. Their evidence is reviewed. They develop professional values, technical competence, business acumen, judgement, relational skill, and digital acumen.

That is the theory.

The practical question is what happens when the basic work changes shape faster than the training model can adapt.

If AI prepares the first draft of the working paper, what exactly did the trainee learn? If AI identifies the sample, flags the exceptions, writes the client query, and drafts the conclusion, what is the trainee's role? Are they evaluating the machine's output with professional scepticism, or are they simply accepting a polished answer because it looks plausible?

The answer depends on design.

AI can improve the training contract if it is used to accelerate learning. A trainee can use AI to compare interpretations, test reasoning, generate questions, simulate a review comment, or explore why an exception matters. In that model, AI becomes a tutor, a critic, and a productivity tool.

AI can also hollow out the training contract. If it removes the early tasks without replacing them with deliberate learning, the trainee gets less exposure while the firm still records productivity. The person moves faster through the process, but may understand less.

This is the apprenticeship bottleneck.

The Leapfrogging Temptation

South Africa often likes the idea of leapfrogging. We speak about skipping old infrastructure and moving directly to newer models. Mobile money can leapfrog bank branches. Online learning can leapfrog physical capacity. Renewable energy can leapfrog parts of the central grid. AI now promises to leapfrog some of the slow professional-development work that has traditionally sat at the bottom of firms.

There is a positive version of this.

A young person with good AI tools can do work earlier that used to require years of exposure. They can analyse a dataset, prepare a presentation, draft a memo, build a financial model, review legislation, and test assumptions with far less friction. This can help a graduate from a weaker school or a smaller town compete with people who already know the language of elite firms.

But leapfrogging has a danger. It can skip the institution-building step.

In professional services, the old model had many flaws, but it created a ladder. It was not equal for everyone, and it was often too dependent on who gave you a chance. Still, it gave many young professionals a protected route from novice to contributor. They could be slow at first because the model expected them to be slow. They could learn by doing the simple thing before being trusted with the complex thing.

AI can make the simple thing disappear.

Once that happens, employers may start asking entry-level people to arrive with mid-level judgement. This is already visible in global labour-market commentary. The World Economic Forum's 2026 briefing on AI and entry-level work notes that routine entry-level tasks are being automated, expectations for entry-level roles are changing, and new skills are in demand. PwC's work on early careers also shows that entry-level workers are often curious and excited about AI, but that gaps are emerging between business expectations and worker confidence.

The result is not a clean replacement story. It is a raising of the floor.

The graduate is not told, "We no longer need you." They are told, "We need you to be useful faster."

That sounds fair until we ask who had the opportunity to become useful before they were hired.

The South African Problem

In a high-employment economy, a shrinking first rung is painful. In South Africa, it is dangerous.

Our labour market already fails too many young people at the point of entry. A qualification does not guarantee work. Work experience is demanded before work experience is available. Networks matter. English confidence matters. Transport matters. Digital access matters. The first job is often less about income and more about becoming legible to the next employer.

This is why entry-level professional roles carry more weight than their job descriptions suggest. An article clerk position is not only a salary. It is a signal. It says the person has entered a recognised pathway. It gives them supervised exposure, professional language, client context, deadlines, systems, and a record of competence. It helps them move from potential to evidence.

AI may not remove that pathway immediately. In some firms it may even improve it. But the perceived risk among graduates is rational. They can see the same thing senior people can see: much of the work that used to justify a large junior cohort is now technically automatable.

This creates a perception problem before it creates a statistical job-loss problem. Young people may start to believe that the ladder is being pulled up. Parents may question whether a long professional route still carries the same promise. Universities may find that employability depends on whether graduates can work with intelligent tools from day one.

The pressure will not be evenly distributed.

Well-resourced students will use AI before they enter the workplace. They will practise with tools, improve their writing, simulate interviews, and learn the language of professional work. Poorer students may encounter these tools later, through constrained access, weaker devices, limited data, or institutions that have not integrated AI well.

If the entry-level market starts rewarding AI fluency before the education system has made AI fluency broadly available, inequality will deepen.

The Wrong Response

The wrong response is to defend old junior work as if inefficiency were a public good. There is no future in making article clerks do work that machines can do better, faster, and with fewer errors. That will not protect the profession. It will only make the profession dishonest about the nature of competence.

The other wrong response is to declare that AI will automatically create better jobs and that displaced entry-level work will reappear higher up the value chain.

Better jobs do not appear by magic. They appear when institutions redesign work, training, supervision, assessment, and incentives. If a firm captures the productivity benefit but does not rebuild the learning pathway, then the gain is private and the cost is social.

We should be suspicious of both nostalgia and techno-optimism.

The serious question is whether AI is used to develop people or merely to reduce the need for them.

What A Better Model Could Look Like

The old model assumed that juniors learned because they were exposed to work. The new model must be more intentional.

Training offices and professional firms should separate tasks into at least four categories.

First, tasks AI should do because they are mechanical and add little developmental value: formatting, basic extraction, standard cross-referencing, repetitive data cleaning, and first-pass document organisation.

Second, tasks AI can support but the trainee must understand: reconciliations, variance analysis, sampling, working paper conclusions, control descriptions, tax research, and client-query drafting. The trainee should be able to explain the logic, identify risks in the AI output, and defend the conclusion.

Third, tasks AI should be used to teach. The tool can generate alternative explanations, simulate a manager's review note, ask why evidence is sufficient, or compare a weak answer with a stronger one.

Fourth, tasks that remain deeply human because they require judgement, ethics, relationship management, courage, and accountability. These include challenging a client, escalating a risk, deciding whether evidence is enough, admitting uncertainty, and signing off a professional conclusion.

This model changes the role of the junior. The trainee is no longer only a preparer. They become a reviewer of machine-assisted work, a learner in a guided system, and gradually a professional who connects evidence to judgement.

But this requires supervision. A trainee cannot review AI output responsibly if no one has taught them what a good answer looks like. AI will not solve weak training offices. It will expose them.

What Universities Should Do

Universities should not respond by adding one generic AI module and declaring the problem solved.

The issue is not tool familiarity alone. It is professional formation in an AI-rich environment.

Students need to learn how to use AI to think, not how to outsource thinking. Assessments must ask students to show their process, compare alternatives, test evidence, identify hallucinations, and explain why a conclusion is defensible. Students must also learn where AI is useful, where it is risky, and where it is inappropriate.

For accounting students, this could mean working with messy ledgers, simulated audit files, weak controls, contradictory explanations, and AI-generated working papers that contain subtle errors. For law students, it could mean reviewing AI-drafted arguments against primary sources. For business students, it could mean using AI to build scenarios while defending the assumptions.

The point is not to ban the tool. The point is to make judgement visible.

If universities do this well, they can reduce the inequality of preparation. They can give students who do not have private access to professional networks a structured way to practise the new form of work. They can also help employers trust that graduates have not simply used AI to produce fluent nonsense.

What Employers Should Measure

Employers should stop measuring only output per head when they introduce AI into junior work.

They should also measure learning per head.

A firm that automates 30% of junior tasks may look more productive. But if junior capability development falls by 50%, the firm has created a delayed risk. It will show up when the firm needs managers, reviewers, partners, specialists, and leaders who understand the work from the ground up.

The right questions are practical. Are we hiring fewer juniors because the work disappeared, or because we have failed to redesign their role? Are trainees getting enough variety and complexity? Can they explain AI-assisted outputs without the tool? Are seniors reviewing the trainee's judgement, or only the final file? Are we using AI to coach, or only to compress cost?

Professional bodies should ask similar questions. Training requirements may need to recognise AI-assisted work explicitly. Not as a loophole, but as a governed part of competence. A trainee should be able to use AI, but the evidence of competence must still belong to the trainee.

This is especially important in accounting because public trust depends on more than technical fluency. A profession that produces people who can operate tools but cannot exercise judgement has not modernised. It has weakened itself.

The Real Choice

The debate about AI and entry-level jobs is often framed as a choice between protection and progress. That is not the real choice.

The real choice is between extraction and formation.

Extraction uses AI to take cost out of the system while assuming the talent pipeline will somehow repair itself. Formation uses AI to improve productivity while deliberately rebuilding the ladder into professional work.

South Africa needs the second path.

We cannot afford a professional economy where only the already prepared can enter. We cannot afford to tell young people that the first rung has become too expensive, too slow, or too inefficient to maintain. We also cannot afford to preserve outdated tasks merely because they once served as training.

The article clerk of the future should not be a cheaper version of yesterday's junior. Nor should they be replaced by a dashboard and a prompt. They should be trained differently, with better tools, clearer judgement standards, stronger supervision, and more explicit evidence of competence.

AI can remove drudgery. That is good. But if it removes apprenticeship, we will have solved the wrong problem.

The future of entry-level work in South Africa will not be decided by the capability of the technology alone. It will be decided by whether firms, universities, and professional bodies are willing to redesign the first rung before it breaks.

Sources Consulted

• Statistics South Africa, Quarterly Labour Force Survey Q1 2026: https://www.statssa.gov.za/publications/P0211/P02111stQuarter2026.pdf
• Statistics South Africa, youth labour market presentation Q1 2026: https://www.statssa.gov.za/publications/P0211/Presentation%20QLFS%20Q1%202026.pdf
• SAICA, CA2025 competency framework: https://www.saica.org.za/initiatives/competency-framework/
• SAICA, Training Regulations effective 1 January 2025: https://saicawebprstorage.blob.core.windows.net/uploads/resources/Training-Regulations-2025-Clean-version.pdf
• World Economic Forum, How AI is Changing Early Careers, January 2026: https://reports.weforum.org/docs/WEF_Briefing_AI_and_Entry-Level_Jobs_January_2026.pdf
• PwC, Africa Workforce Hopes and Fears Survey 2025: https://www.pwc.co.za/en/publications/global-workforce-hopes-and-fears-survey.html
• International Labour Organization, Generative AI and Jobs: A Refined Global Index of Occupational Exposure, 2025: https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure