If AI Does the Junior Work, Where Will Judgment Come From?
There is a growing argument that artificial intelligence will not eliminate entry-level careers so much as improve them. AI will take over the repetitive work, junior employees will move more quickly into interpretation and decision-making, and organizations will develop capable professionals faster than before.
It is an appealing idea. It is also based on a questionable understanding of how professional judgment develops.
The routine work undertaken at the beginning of a career was never valuable simply because it was repetitive. Its value came from the exposure it provided. Researching companies, reconciling data, reviewing documents, preparing presentations, producing an initial analysis and observing experienced colleagues exposed people to patterns. They learned what looked normal, where exceptions occurred, which details mattered and how apparently minor errors could alter the eventual conclusion.
Over time, that accumulated exposure became experience. Experience gradually became judgment.
Removing the work and moving an inexperienced person directly into interpretation does not automatically accelerate that process. How can someone reliably interpret an analysis they do not yet understand, verify work they could not produce themselves or challenge a recommendation without the experience needed to recognize that it is wrong?
Consider recruitment. AI can produce a shortlist that appears closely aligned with a job description, but assessing whether those people can succeed requires far more than matching their previous titles and employers. Someone needs to understand the client’s culture, leadership dynamics, market position, compensation, internal politics and the real reasons the role exists. A candidate who appears perfect on paper may be entirely wrong for the situation, while the strongest person may have transferable experience that an automated comparison overlooks.
A junior recruiter who has never personally mapped a market may not recognize that the AI has missed an entire category of potential candidates. If that person has never conducted a detailed interview, managed a reluctant candidate or seen a seemingly ideal appointment fail, on what basis are they expected to challenge the system’s recommendations?
This is equally relevant in law, accounting, consulting and technology. Junior lawyers historically spent time reviewing large numbers of documents, while trainee accountants worked through transactions and reconciliations. Much of that work was laborious, and there is no compelling reason to preserve unnecessary manual effort. However, it also helped people recognize inconsistencies, understand how evidence fits together and develop an instinct for when something requires further investigation.
If technology removes the process entirely, the organization needs to be certain that it has not also removed the training ground. Asking someone to interpret the final output without understanding how it was assembled is not necessarily career acceleration. It may simply conceal the gap in their knowledge until they encounter a situation the system handles badly.
AI can produce exceptionally convincing material. Its answers are often polished, confident and entirely plausible. This can make its mistakes more difficult to identify. A visibly incomplete first draft invites scrutiny. A professionally written but subtly incorrect answer can pass through an organization unnoticed.
I use AI regularly in my own work. It can help me structure information, test an argument, review material and develop an initial draft more quickly. However, I know the outcome I want to achieve. I read what it produces, question it, correct it, improve it and frequently ask it to approach the problem again. I can recognize when an argument is repetitive, when something lacks commercial context, when a conclusion is unsupported or when the tone is inappropriate for the intended audience.
Sometimes the first answer is useful. Sometimes it is generic, factually questionable or based on an incorrect interpretation of the request. The value comes from being able to tell the difference.
That is not because the technology has given me judgment. It is because I bring decades of professional experience to the technology. An experienced person can automate a task because they understand it. A junior employee may be automating the opportunity to learn it.
AI in experienced hands can improve productivity. AI in inexperienced hands can improve the presentation of a mistake.
There is also a danger that overreliance on AI will gradually weaken the very human capabilities organizations will increasingly need. Research conducted by WSJ Intelligence in partnership with Philip Morris International surveyed more than 2,500 professionals across five countries and organizational levels ranging from entry-level employees to the C-suite.
Twenty-nine percent identified the decline of critical thinking, judgment and creativity as the greatest risk to human capabilities. Only one in four strongly agreed that their organization had a clear and effective process for checking and verifying AI-generated output.
The danger, therefore, is not only that AI may produce mistakes. Overreliance on it may weaken the human capabilities needed to detect them. Technology can create more room for thought, but it can also remove the productive struggle through which people develop intellectual independence, original ideas and sound judgment. Human cognition does not remain strong simply because it is human. Like any capability, it has to be exercised.
We can already see what happens when organizations place too much confidence in polished outputs. In 2025, Deloitte Australia agreed to refund part of the fee for a A$440,000 report commissioned by the Australian government. The report contained fabricated references, citations to nonexistent academic research and an inaccurate quotation attributed to a federal court judgment. A revised version subsequently disclosed that Azure OpenAI had been used in producing it.
The client was paying for independent assurance and professional advice, not simply for the production of a lengthy document. The failure was organizational: inaccurate material survived the firm’s review process and reached the client. AI may have produced the errors, but the organization remained responsible for delivering them.
This is particularly damaging for a professional-services firm. Clients engage trusted advisers because they expect informed analysis, independent challenge and credible recommendations. If the adviser delegates too much of that work to a system and fails to identify obvious fabrications, the issue is no longer simply technological. It becomes a question of professional competence and trust.
The risks become more serious when AI does not merely help produce a document but influences decisions about people. A US federal judge has ruled that Workday must face claims alleging that its AI-powered recruitment software screened out applicants in ways that violated California law and federal disability protections.
The proposed class action alleges that apparently neutral factors, including employment gaps, may have operated as proxy indicators for disability or illness. Other allegations concern discrimination against Black applicants, women and candidates over 40. These remain allegations, and Workday denies that its technology discriminates or makes hiring decisions. Nevertheless, the case challenges the assumption that an automated system must be objective because it has not been explicitly instructed to consider protected characteristics.
A system trained on historical employment patterns may reproduce the effects of previous decisions. An employment gap could be interpreted as a negative indicator without accounting for illness, caring responsibilities, military service, redundancy or relocation. A career change may appear inconsistent when it actually demonstrates adaptability. International experience may be undervalued because it does not resemble the career paths contained in the training data.
A company may buy an AI recruitment product believing that the supplier has already tested it thoroughly. Its internal team may not know what data shaped the model, which variables influence its recommendations or whether particular groups are being screened out disproportionately. Yet if the process creates unlawful or damaging outcomes, the employer cannot reasonably argue that the technology made the decision. The organization selected the system, incorporated it into the recruitment process and remains accountable for the consequences.
AI does not explain itself to a rejected candidate, repair a damaged client relationship, accept professional liability or lose its career when an unchecked recommendation causes harm. People and organizations ultimately carry those consequences.
The assumption that every organization can safely build its operations around AI is equally questionable. The capabilities of a major global bank or technology company cannot be reproduced simply by purchasing access to a commercial product.
JPMorgan Chase expects to spend approximately $19.8 billion on technology in 2026. Its AI capability sits within an enormous investment in infrastructure, proprietary data, cybersecurity, governance, specialist talent and employee training.
Most businesses do not have anything comparable. They may subscribe to the same general category of technology, but buying a product does not create institutional expertise, reliable internal data or effective governance.
Imagine a medium-sized company using AI to forecast staffing requirements. The system recommends reducing a service team because automation should absorb much of its workload. On paper, the productivity case looks compelling. What happens if the model has not accounted for seasonal demand, a forthcoming regulatory change or the volume of complex cases requiring human intervention? By the time customer complaints rise and experienced employees have left, reversing the decision may cost considerably more than the original savings.
The same problem could emerge in a consulting business. AI may allow a smaller team to produce market analysis and presentation materials at considerable speed. Initially, margins improve. Over time, however, junior consultants have fewer opportunities to conduct original research, test hypotheses, interview clients and understand why one recommendation works while another fails. They become highly efficient at assembling an answer without necessarily developing the expertise on which the firm’s advisory reputation depends.
The immediate output may look better. The underlying capability may be getting weaker.
Organizations risk adopting the operating philosophy of the most technologically advanced companies without possessing their investment or controls. They may reduce entry-level hiring, automate foundational work and instruct a smaller number of junior employees to oversee the resulting process. What is presented as transformation may amount to cost reduction accompanied by a new vocabulary.
Some proposed solutions suggest using AI itself as a coach. It may have value within a structured learning process, but loading previous examples, frameworks and decisions into a system is not the same as giving it real-world experience or an understanding of why one decision succeeded while another failed.
A trainee could use AI to practise handling a difficult client conversation. The system might identify obvious omissions or suggest alternative language. What it cannot fully reproduce is the tension of a real conversation in which the client has competing priorities, does not disclose everything and reacts unexpectedly. Nor can it explain the internal history behind the client’s position unless someone with that knowledge contributes it.
Research into medical education offers a more credible use of the technology. In one experiment cited by McKinsey, medical students answered AI-generated case questions and received automated feedback. In another model, a chatbot played the patient while students conducted the interview, committed to a diagnosis and defended their reasoning before the system provided its critique.
The sequencing is important. The student attempts the diagnosis first. AI does not provide the answer and then ask the student to agree with it. The learner has to think, decide and expose the reasoning to challenge.
This approach may increase the number of scenarios someone can encounter and allow learning to happen more quickly. It does not remove the need for medical knowledge, clinical supervision or eventual exposure to real patients. A simulated patient will not perfectly reproduce incomplete symptoms, emotional distress, family pressure or the consequences of a wrong decision.
A similar model has been proposed in software engineering, where a senior engineer supervises a small group of junior employees working alongside AI tools. The senior observes what the junior accepts, rejects and misunderstands. This could be an effective modern form of apprenticeship because the technology increases the amount of work that can be attempted while the learning remains visible.
These models share an important principle: the learner attempts, decides and explains before receiving assistance. AI expands the opportunities to practise; it does not perform the thinking on the learner’s behalf.
Experience may be accelerated through greater exposure, deliberate practice, close mentoring and earlier responsibility. It cannot be removed from the process altogether. Junior employees still need to observe real decisions, see how mistakes are corrected, encounter ambiguity and understand the consequences of getting something wrong.
This creates a difficult economic question. The attraction of AI is frequently its ability to reduce cost, headcount and time. Yet replacing the developmental value of junior work requires more structured training and greater involvement from senior employees. It requires organizations to reinvest some of the efficiency gains rather than treating them solely as savings.
Senior professionals would need to spend more time explaining their reasoning, reviewing attempts and exposing junior colleagues to difficult situations. Managers would have to be selected and assessed partly on their ability to develop people rather than simply deliver output. Entry-level roles would need protected learning time, meaningful rotations and controlled opportunities to make mistakes.
That is a significant investment. It is also the opposite of the low-cost model that makes removing junior work attractive in the first place.
Many businesses already struggle to provide consistent management, mentoring and career development. It is difficult to believe that every company removing junior work will simultaneously construct a carefully designed professional apprenticeship in its place. The more likely outcome is that the work disappears first, the savings are recorded immediately and the effects on the talent pipeline emerge years later.
By then, the problem will no longer be confined to graduates trying to enter the workforce. Organizations may find themselves with fewer people capable of becoming managers, directors, partners and technical specialists. They will compete for a shrinking pool of experienced talent while wondering why succession pipelines have weakened and institutional knowledge has disappeared.
The problem could become particularly acute in professional services. Accounting, law, consulting and executive search all depend on a pyramid in which people develop through progressively more complex work. If firms reduce the base of that pyramid without redesigning how expertise is built, they cannot assume that the same number of credible partners and advisers will somehow appear at the top years later.
Harvard Business Publishing has described the instinct to automate away entry-level jobs as understandable but short-sighted. These roles are not simply inefficiencies waiting to be eliminated. They form part of the system through which organizations transfer knowledge, develop expertise and create future leaders. Remove the role without replacing that developmental function and the immediate productivity gain may conceal a much larger institutional cost.
Businesses could eventually find themselves paying a premium to acquire the expertise they decided was too expensive to develop.
The objective should not be to preserve repetitive work simply because previous generations performed it. Technology should remove unnecessary administration and improve productivity wherever it can. However, organizations must identify the knowledge embedded in a task before automating it. They can then decide which elements should still be attempted independently, which can be practised through simulation and where exposure to real situations remains essential.
The purpose is not to preserve inefficient work. It is to preserve the development that occurred through it.
The fashionable belief that young professionals can move immediately into interpretation, judgment and trusted advisory work avoids the central problem. Judgment is not a stage that can simply be inserted into a redesigned workflow. It is a capability developed through knowledge, repetition, mistakes, feedback, context and accountability.
AI may help an experienced person reach an answer faster. It may also help a learner practise more effectively when used within a properly designed development process. What it cannot do is confer expertise on someone by presenting them with an expert-looking response.
Organizations should therefore be cautious about celebrating the disappearance of the first rung of the career ladder. If they remove the work through which people learned, reduce the number entering their professions and substitute artificial outputs for genuine experience, they may achieve an impressive short-term improvement in productivity.
They may also discover, several years from now, that very few people remain who know whether the machine is right.
.png)
Comments