Think about the best senior people in your business. Years ago, each of them reconciled the accounts, drafted the first version, answered the routine tickets, or checked the same forms a hundred times. That work was never only output. It was how they learned what a good result looks like, which exceptions matter, and when something that looks fine is actually wrong.
Now AI can do a growing share of that routine work in seconds. That is genuinely good news for productivity, and it raises a genuinely new question for succession: if the routine work no longer needs a beginner, how does a beginner become an expert?
This article sorts through what the 2026 evidence says about AI and entry-level work, which is more contested than the headlines suggest, and then focuses on what a mid-size European business can actually do about it. It's general information rather than HR or employment-law advice, and the details of apprenticeship and hiring rules vary by country.
What the evidence says, and where it disagrees
A fast-growing set of studies examines whether generative AI is changing hiring at the bottom of the career ladder. Most are working papers rather than peer-reviewed publications, and they use very different data, which is part of why they disagree.
| Source | Data | Main finding | Main caveat |
|---|---|---|---|
| Stanford Digital Economy Lab working paper, Brynjolfsson, Chandar and Chen, August 2026 | US payroll records from ADP covering millions of workers, through June 2026 | Employment of workers aged 22 to 25 in AI-exposed occupations is 19% below where it would have been had it kept pace with less-exposed peers, with no comparable gap for experienced workers and no evidence of widespread economy-wide displacement | The authors describe the results as descriptive, not causal, and the pattern weakens when education is controlled for |
| Harvard working paper, Hosseini Maasoum and Lichtinger, August 2025 version | US résumé and job posting data for tens of millions of workers | In the paper's headline estimate, junior employment at firms adopting generative AI fell 7.7% relative to non-adopters after six quarters, driven mainly by slower hiring rather than separations, while senior employment was largely unchanged | A working paper that is not peer reviewed, and other researchers argue some estimates may reflect broad hiring slowdowns rather than targeting of junior roles |
| Humlum and Vestergaard, NBER working paper, revised March 2026 | Danish administrative records linked to adoption surveys, about 25,000 workers in 7,000 workplaces | No significant effects on earnings or recorded hours, ruling out effects larger than about 2% roughly two years after ChatGPT's release, including in early-career jobs | It measures earnings and hours of workers and workplaces, and does not directly answer the question of hiring new entrants |
| Lambert and Schindler, 2026 | Job postings and hiring data from the US, UK, Canada, and Australia | Argue that remote work explains falling entry-level hiring better than AI exposure does | The two are highly correlated, and the Stanford authors report the opposite on payroll data |
| Aldasoro and co-authors, EIB and BIS working papers, January 2026 | More than 12,000 non-financial firms in the EU and US, in matched data | AI adoption raised labour productivity by about 4% with no adverse effect on firm-level employment in the short run, with the strongest gains in medium and large firms | It looks at firm-level employment, not junior hiring, covers the short run only, and does not show that AI never reduces employment |
Three details in the Stanford paper are worth spelling out, because they are easy to lose in headlines. First, the paper's own first finding is that it sees no evidence of widespread, economy-wide job displacement. The effect it documents is concentrated on new entrants in AI-exposed occupations. Second, it operates mainly through reduced hiring rather than layoffs, and it appears where AI is used to substitute for human tasks, while where AI complements workers, employment is flat or rising. In levels, employment of 22- to 25-year-olds in the two most exposed groups of occupations fell about 11% between November 2022 and June 2026, while it grew about 10% in the three least exposed groups. Third, the authors are explicit about limits: the ADP sample over-represents larger firms and AI-exposed occupations, the divergence is larger there than in national survey benchmarks, and some of the pattern predates ChatGPT. Earlier versions of the paper headlined regression estimates of 13% and then 16%. The 19% figure is a simpler descriptive measure, which the paper reports at 15% on July 2025 data and 19% on June 2026 data, so the numbers across versions aren't a like-for-like series. To be clear about what the 19% means: it is a relative employment gap, measured against where employment would have been had it kept pace with less-exposed peers, and it is not a count of jobs eliminated or of young workers who lost their jobs.
The Harvard paper is a working paper rather than a peer-reviewed publication, and its 7.7% figure is one estimate for a specific window, six quarters after adoption, not a universal effect. The Danish and European firm-level studies answer different questions from the hiring studies. The Danish paper looks at earnings and hours, and the EIB and BIS study looks at firm-level employment and productivity in the short run, so neither contradicts a pattern in new-entrant hiring, and neither proves that AI never reduces employment.
It's also worth saying plainly that some figures circulating online, including very large quarterly declines attributed to these studies, don't match what the papers report. The primary sources are more modest and more qualified than the viral versions.
What employer surveys add, in the UK and EU
Surveys of employers add context, with an important caveat: they record what managers say, not what causes it.
The Open University's 2026 Business Barometer surveyed 1,500 UK business leaders in April and May, through the research firm Opinium. It found that 51% said AI was changing how they hire and 19% said they had reduced entry-level recruitment, with 42% of that group citing wider AI adoption, which works out to roughly 8% of all respondents. That 8% is our own calculation from the two reported percentages, not a figure the survey reports directly, and it should not be read as 42% of all employers cutting entry-level recruitment because of AI. Reporting on the same survey adds the broader context: 76% of organizations said economic uncertainty had made recruitment or training harder, and 43% had hired fewer staff overall in the past year. Entry-level hiring is changing against a general slowdown in recruitment, not only because of AI.
A separate Survation poll of 1,001 UK businesses for Lancaster University's Work Foundation, conducted in May and reported by Reuters on 26 August, found that 36% of employers had cut entry-level roles over the past year and that 73% believe it has become harder for young people to find work than five years ago. Larger employers were considerably more likely than small firms to say AI and automation had reduced available jobs. Surveys like this can't establish causation.
Across the EU, Eurostat's release of 1 October 2026 put youth unemployment, covering people aged 15 to 24, at 15.4% in the EU and 15.0% in the euro area in August 2026. In the same release's revised series, the figures a year earlier, in August 2025, were 15.3% and 14.8%, and the number of unemployed young people rose by 19,000 in the EU and 38,000 in the euro area over the year. Eurostat doesn't attribute the change to any cause, overall unemployment rose only slightly over the same period, and the youth rate expresses jobless young people as a share of the youth labour force, not of all young people. It's a backdrop, not evidence about AI.
A Klarus survey of 500 UK and Irish mid-market leaders, reported by CNBC on 23 September, points the other way. It found that 45% said AI was being used to help junior employees work better and faster, and almost a quarter said it was creating new roles and opportunities.
On the European evidence overall, an April 2026 box in the ECB's Economic Bulletin notes growing evidence of negative effects on junior workers in highly exposed occupations, with the evidence concentrated in the US, and cites initial EU firm-level evidence of productivity gains without short-term job replacement. We did not find EU-wide, firm-level data on junior hiring specifically in the sources reviewed, so conclusions for Europe rest on adjacent evidence.
Why the pipeline question is real even if the debate isn't settled
Whatever the cause of recent hiring patterns, there's a structural point that stands on its own. A 2025 working paper by economists Luis Garicano and Luis Rayo, updated in 2026, frames apprenticeship as a bargain: juniors pay for their training by doing menial work. AI now does an increasing share of that work, which puts the bargain at risk. In their model, whether training survives depends on how much more valuable a fully trained expert is than what AI can do on its own. This is a theoretical model, not an empirical finding, but it names the mechanism precisely.
A related model-based paper by Enrique Ide, "Automation, AI, and the Intergenerational Transmission of Knowledge," argues that technology which reduces junior workers' access to the most productive mentors can erode the future supply of the very expertise it complements. Like the apprenticeship model above, it is a theoretical contribution: it describes what could happen under its assumptions, not effects already observed in the data. And the Stanford authors suggest a possible reason the effects show up at the entry level: AI may substitute for codified knowledge, the kind taught in textbooks and documented procedures, while complementing tacit knowledge, the kind built through practice, mentorship, and exposure to real cases. They label that mechanism suggestive rather than proven.
The practical reading doesn't depend on which study is right, and it isn't an argument for keeping routine work manual so beginners have something to do. The routine work was the vehicle through which beginners saw how experienced people decide. If you automate the vehicle without replacing it, you may keep your costs down this year and have fewer experienced people in five to ten years. The efficient response is to automate the routine work and rebuild, deliberately, the learning it used to provide.
How juniors use AI matters
A second line of evidence concerns learning itself. In a randomized experiment published in 2026, Judy Hanwen Shen and Alex Tamkin, researchers affiliated with Anthropic, had 52 professional and freelance software developers learn an unfamiliar Python library, with or without an AI assistant, and then tested them without AI. The assistant in the study was built on GPT-4o. The AI-assisted group scored about 4.15 points lower on the quiz, a gap on a 27-point quiz that the authors describe as about a 17% score difference or roughly two grade points. Completion time didn't differ significantly between the groups. The largest gap showed up on debugging questions, and the result depended on how people used the tool: those who asked conceptual questions or requested explanations alongside generated code preserved their learning, while those who delegated coding or debugging to the AI scored much lower.
It's one small, short-term experiment on a single task, using a one-hour session rather than months of work, so it doesn't show that AI causes long-term deskilling. Its useful lesson is practical: AI use isn't automatically good or bad for learning. The design of how a beginner works with it makes the difference, and that design is something a business controls.
What redesigned junior roles look like in practice
Some companies are already redesigning junior roles rather than cutting them, though outcome data is thin, and what follows are company statements rather than measured results.
IBM's chief human resources officer said in February 2026 that the company plans to triple its US entry-level hiring in 2026, after redesigning those roles. According to reporting on her remarks, junior roles moved away from routine production toward client engagement, judgment calls, testing, and oversight of AI systems. In HR, for example, entry-level staff step in when a chatbot's answer is insufficient, correct it, and work with managers. She also said cutting entry-level hiring may save costs in the short term but risks leaving too few people to develop into future mid-level managers. This is an announced plan, not a measured outcome, and IBM did not disclose baseline numbers, so the size of the increase can't be verified from outside.
In the CNBC piece of 23 September, executives described a similar idea. One suggested treating entry-level work as exposure rather than output: juniors check and correct AI-generated work while senior colleagues deliberately teach the judgment around it. London-based FDM Group said it is training people on real business problems using agentic AI tools, shifting the focus from learning individual tasks to understanding the work those tools complete.
None of this proves the redesigns work. It shows that organizations are treating the question as a design problem, which is the useful reframing.
What this means for a European mid-size business
Four points are specific to your situation.
First, the EIB study found that the productivity gains from AI were concentrated in medium and large firms, and that complementary investments in software, data, or workforce training were important for unlocking them. Training isn't an optional extra in that finding; it's part of how the gains materialize.
Second, Europe has an institutional asset worth using. Cedefop the EU agency for vocational training, has issued a joint call with the OECD for abstracts and research contributions on apprenticeships in the age of AI, with a submission deadline of 15 October 2026, for a joint apprenticeship symposium planned for May 2027 in Thessaloniki. The call notes that AI is taking over baseline tasks once performed by apprentices and entry-level graduates and that apprenticeships may be exposed to fewer entry-level openings, though the contraction so far appears concentrated in white-collar roles and likely to be uneven. It also describes apprenticeships' learning-by-doing and two-venue structure as a potentially agile response. It's a research call and an institutional signal, not evidence that any particular redesign works. Germany's BIBB has also focused recent projects on integrating AI into apprenticeship training. Those are signals of direction rather than proof of results, but they suggest dual-training traditions are an asset when redesigning junior pathways, and that apprentices are expected to learn to work with AI.
Third, the talent market may open opportunities for smaller employers. Ide's paper also notes anecdotal reports that some graduates are taking positions at smaller or mid-tier firms that previously struggled to attract them. It's anecdotal, and we mention it as a possibility rather than a trend.
Fourth, a junior pipeline is a continuity question as much as a hiring one. The same logic we describe in Founder Dependency: The Hidden Risk to Your Business's Value applies: expertise that lives in a few experienced heads needs a deliberate path to the next generation.
For context: the US picture
Much of the quantitative evidence on this topic comes from US data. The Stanford and Harvard studies both analyze US workers, IBM's announcement concerns US hiring, and the large payroll sample comes from a US provider. That doesn't make the findings irrelevant to Europe, but US labour-market flexibility, firm sizes, and hiring practices differ from those in most European countries, and as noted above, EU-specific evidence on junior hiring is thinner. It's reasonable to treat the US data as an early signal and a prompt for planning, not a forecast for your own market.
A framework for rebuilding the junior pathway
This is a Kubera AI planning heuristic, not a validated method. It draws on the evidence above, and the specific design choices haven't been tested in controlled trials, so we'd treat any implementation as a pilot to measure.
The Kubera Junior Pathway Model has four steps:
| Step | Question to ask | What to do | Example output |
|---|---|---|---|
| 1. Map the learning vehicle | Which tasks did beginners do that actually taught them judgment? | List the tasks of a junior role and sort them into drudgery, execution that AI now handles reliably, and tasks that were the learning vehicle | A one-page map showing which tasks to automate and which learning to replace |
| 2. Replace volume with exposure | How will a beginner still see real cases? | Set up review-and-correct lanes where juniors check AI output against a reference and a senior explains the reasoning | A weekly review session built from real, anonymized cases |
| 3. Capture the judgment | Where does senior reasoning live today? | Record why decisions were made, not only what was decided, and make that library searchable | A decision log and a case library, linked to your knowledge base |
| 4. Practice unaided and measure | How will we know a junior is actually learning? | Schedule deliberate AI-free practice, set autonomy tiers for what juniors may approve alone, and track simple milestones | Milestones such as errors caught in AI drafts and time to independent sign-off |
Step 4 draws on the learning study above. Unaided practice isn't nostalgia; it's how you check that the junior can do the thing the AI is doing, which is the skill needed to supervise it.
Where automation helps build the pathway
This is where automation does more than remove work. A well-built workflow can create the structures the model above needs. It can route routine volume to AI and send the exceptions, the cases that need judgment, to a review lane where a junior and a senior look at them together. It can log decisions and their reasoning as a by-product of the work, building the case library without asking anyone to write documentation on top of their day job. It can apply approval tiers, so a junior's sign-off rights expand as milestones are met, using the same logic we describe in How Much Autonomy Should an AI Agent Have?. And it can feed an internal knowledge base, the architecture we cover in How to Build an Internal Knowledge Base with AI, so a new hire can search how experienced colleagues handled similar situations.
The people side matters as much as the tooling. We cover how teams adopt new tools in How to Onboard Your Team to AI Automation, and how roles change more broadly in How AI Restructures Teams, Not Just Headcount. The distinction between routine execution and judgment that underpins this article is the one we draw in Can AI Replace an Employee?. A 12-month plan for putting this in sequence is in How to Build an AI Roadmap for a Small Business.
Where this plays out in practice
Illustrative scenario, not a specific Kubera client: a mid-size accounting and advisory firm in a European city used to start new staff on reconciliations, data entry, and first-draft client reports, and it was in those tasks that juniors absorbed what normal looks like. After automating most of the reconciliation and drafting work, the firm sees that its newest staff now have little to do and, more importantly, little to learn from. It maps the junior role, keeps the automation, and builds a weekly review lane in which AI-prepared client packs, including a few seeded known issues, are checked by juniors who flag and explain what they find, followed by a short session with a senior. Each resolved case is logged with the reasoning, building a searchable library. Sign-off rights expand in steps as juniors meet agreed milestones, and each month includes one task done without AI assistance. The firm treats the design as a pilot and reviews after two quarters whether juniors are reaching independent sign-off faster than before. We describe the design intent here, not a measured result.
FAQ
Is AI actually reducing junior hiring? The evidence points in different directions. US payroll data show weaker employment growth for young workers in AI-exposed occupations, mainly through reduced hiring, but the authors call their results descriptive and not causal, and other studies attribute part of the pattern to remote work or to broad hiring slowdowns. A Danish study found no effects on earnings or hours of existing workers. The honest answer is that it's likely part of the picture and not yet settled.
Should we stop hiring juniors? The evidence doesn't support a blanket answer, and some large employers are doing the opposite. IBM said it plans to triple US entry-level hiring in 2026 after redesigning the roles, though that is an announced plan rather than a measured result, and it didn't disclose baseline numbers. The more useful question is what a junior should do in your business now that routine work can be automated.
Which numbers should I trust? Prefer the primary papers and their stated caveats over secondary summaries. For example, the Stanford paper's headline measure has changed between versions, and some figures repeated online don't match the papers. When you cite a figure, state its source, its date, and whether the authors call it causal.
Are US studies relevant to a European business? As an early signal, yes. As a forecast for your market, treat them cautiously: the main studies use US data, and EU-specific evidence on junior hiring is thinner. The EIB and ECB work points to productivity gains without short-term job replacement at firm level, though that isn't the same question as junior hiring.
What does an AI-era junior role look like? Examples from companies describe less routine production and more checking and correcting AI output, client engagement, testing, and oversight, with senior colleagues teaching the judgment around it. These are company statements rather than measured results.
Does using AI hurt juniors' learning? It depends on how it's used. In one randomized experiment, AI-assisted learners scored lower on an unaided test, but those who asked conceptual questions and requested explanations kept their learning. It's a small, short-term study, so treat it as a design hint, not a rule.
How do we keep juniors from over-relying on AI? Build in deliberate unaided practice, require juniors to explain their reasoning rather than only submit results, and expand their autonomy as they meet milestones. These are design choices drawn from the evidence above, and worth testing in a pilot before scaling.
Can a smaller company afford to redesign junior roles? The principle scales down. Even a small team can list what its juniors learn from, decide which of those tasks AI should take over, and set up a weekly case review. The EIB study suggests training investment is part of how AI gains materialize, though its productivity gains were concentrated in medium and large firms.
How does this connect to an automation project? Directly. The same workflow that automates routine volume can route exceptions to a review lane, log decisions with reasoning, and apply approval tiers. We usually raise the junior pathway early in an automation project, not after it, because the automation changes what the role is.
What should we measure? Keep it simple and observable: how many errors juniors catch in AI-prepared drafts, how long it takes them to reach independent sign-off, and the quality of their explanations. Review the pilot after a couple of quarters rather than expecting immediate results.
If you'd like to work out which routine tasks in your business should move to AI, which learning they were quietly providing, and how to rebuild that pathway deliberately, that's exactly the kind of design work worth doing before the gap shows up in your senior bench.