AI and early careers: how leaders can read conflicting employment evidence

A leader can encounter three apparently conflicting claims about AI and work in a single morning: young workers face greater difficulty entering certain occupations; measured earnings effects remain small; a substantial share of professional tasks could change. Each statement can belong to a careful study. Before making a workforce decision, it is necessary to establish what each study measured.

I am particularly interested in entry into professional work. Hospitals, pharmaceutical companies and medical technology manufacturers need people who can take responsibility in the future. When entry-level tasks change, the route through which those people gain experience may also change. An aggregate employment figure answers only part of the question about that route.

Three sources measuring different things

The Stanford analysis revised in August 2026 examines US ADP payroll records through June 2026. Among workers aged 22 to 25 in AI-exposed occupations, it describes a 19% relative employment gap compared with the trajectory of less-exposed peers. The authors explicitly interpret this as a descriptive association. Educational differences, pre-existing divergent trends and differences from national benchmarks constrain interpretation. [1]

In their April 2025 working paper, Humlum and Vestergaard link surveys of approximately 25,000 workers in eleven occupations to Danish administrative records. They find no statistically significant average effects on earnings or recorded hours during the early period studied. These findings concern different people, a different period and different outcomes from the US analysis. [2]

The ILO's May 2025 index assesses potential technical exposure of occupational tasks. It combines task assessments, survey responses and expert input. The result is a structured view of activities that could change. The index does not count jobs actually lost because of AI. [3]

These sources are the starting points for the original interpretation below. They do not forecast staffing requirements for a particular German hospital or individual company.

Headcount can remain stable while entry becomes harder

Within an organisation, it is useful to distinguish employment levels from movements. A department can employ the same number of people from one year to the next while hiring fewer entrants. Vacancies might, for example, be filled mainly by experienced candidates. A stable total would obscure that change.

This is a thought experiment, rather than a claim about what is already happening in a particular business. It nevertheless indicates what additional information leaders can request: hiring by experience level, internal transfers, departures and the reasons for changes to job descriptions. These figures should be interpreted alongside demand and economic conditions.

The definition of entry also needs to be clear. Age, professional experience and tenure are different characteristics. A 35-year-old can be new to a profession; a 25-year-old may already have several years of experience. Translating an age-based analysis into a statement about professional newcomers requires explicit attention to that distinction.

A technically changeable task does not settle staffing

Several decisions lie between the possibility of AI assistance and a changed job. Is the application available and suitable for its intended use? Will the necessary information be supplied? Does the task fit the workflow? Will demand increase if completing it becomes easier? What responsibility remains associated with the activity?

In a fictional example, a new member of a scientific department regularly prepares initial collections of public technical information. AI assistance might perform part of that preparation. Several outcomes are possible: the same person investigates more questions, spends more time interpreting the information or receives different responsibilities. Which outcome occurs depends on decisions and conditions that a task-exposure index does not capture.

I would therefore require a short explanation for a proposed change to a job description. It should identify the activities being removed or changed, describe the new activities and explain how the resulting need for human work was estimated. This turns a broad expectation about AI into a testable assumption about the organisation itself.

Include the learning route in the same decision

Entry-level activities can have both a productive and an educational purpose. Someone delivers a piece of work while learning to interpret information, recognise familiar errors or find the appropriate colleague. When that activity changes, leaders should consider both purposes.

Preserving every existing routine indefinitely is unnecessary. The intended learning outcome needs to be stated explicitly. If an initial information collection is automated, a newcomer might instead justify the selection, investigate an unresolved question or interpret a conflicting source. Whether this alternative produces the desired learning must be observed locally.

For physicians in training, this question belongs within properly supervised professional education. It can be made equally concrete for managers and scientific employees: which decisions should this person be able to make independently in two years, and what opportunities to practise will get them there? Workforce planning that omits this route leaves an unanswered assumption about the next generation of professionals.

A small observation over the coming months

My proposal is a bounded joint review of one department whose work is already changing. Examine demand for its output, the composition of new roles, actual AI use in selected activities and the learning opportunities available. Repeat the review so that isolated fluctuations do not receive too much weight.

Include a discussion with employees at different experience levels. Where does assistance make entry easier? Where has an opportunity to work something out independently disappeared? Which activities now require more experience than before? The answers are qualitative signals. Record them accordingly and supplement disputed points with a concrete observation.

Decide in advance what action the review could lead to. Possibilities include a changed induction programme, additional supervised tasks or a revised job description. An explicitly named next decision helps the exercise move beyond a general debate about the future of work.

The same record should also capture uncertainty. If a department has too few recent hires to infer a trend, say so. If demand changed at the same time as AI use, keep both explanations visible. This makes the next review more informative and reduces pressure to force a complicated situation into a single account.

I believe international research deserves serious attention, with its limitations equally visible. It leaves organisational leaders with a demanding responsibility: recognise change early, examine staffing assumptions and deliberately design access to future responsibility. Entry into professional work deserves that attention well before a definitive economy-wide answer becomes available.