The End of the Entry-Level Job?

Suppose you hire a 23-year-old analyst. On Monday morning, you ask her to read a stack of reports, compare three competitors, clean up a spreadsheet, draft a client email, and prepare the first version of a presentation. For decades, that was normal entry-level work. It was not glamorous. Much of it was repetitive. Some of it was boring. But hidden inside those assignments was something more important than the finished spreadsheet or presentation. The young analyst was learning how the business worked.

Now imagine that an AI system can complete most of those first drafts in minutes. The employer gains speed. The client may get an answer sooner. The senior employee no longer has to wait for a junior person to assemble the raw material. That looks like an obvious productivity improvement. It may be one. But it creates a second question that is much harder to see on a quarterly earnings report: if the machine does the work through which beginners used to learn, how does the beginner become experienced?

That is the real issue behind the phrase “the end of the entry-level job.” The harder question is whether the old bargain that created those jobs is beginning to break. Employers once accepted lower productivity from beginners because beginners could become valuable experienced workers. AI changes that calculation by making many beginner tasks cheap, fast, and available on demand.

The danger is therefore larger than a missing first paycheck. It is a missing first rung.

The Ladder We Forgot Was a Ladder

Entry-level jobs have always had strange economics. A new lawyer is less skilled than a partner. A new accountant is slower than a veteran. A new financial analyst lacks the judgment of a managing director. A young software developer will make mistakes that an experienced engineer catches. Employers hire these people anyway because work has never been only about producing today’s output. Work is also one of the ways a society produces tomorrow’s expertise.

The junior employee reads the documents, checks the numbers, writes the first draft, listens to the client call, watches the senior person change the draft, and slowly begins to understand why those changes were made. Repetition builds pattern recognition. Mistakes build caution. Feedback builds judgment. A person who spends years doing the smaller pieces of a profession eventually learns how the pieces fit together.

This is why the dull parts of a job can have value that is easy to miss. Nobody becomes a great lawyer because photocopying was magical. Nobody becomes a great investor because formatting a spreadsheet contains secret wisdom. The learning happens around the task. The beginner sees the facts, the exceptions, the bad assumptions, the impatient client, the number that does not reconcile, and the senior person who notices what everyone else missed.

AI can remove a task without removing the profession. That sounds reassuring, and in many cases it is. But it can also remove the task that trained people for the profession. That is the part we have barely begun to price.

The First-Rung Problem

The important unit is the task. Chapter Five argued that a job is more than its tasks and that automation often begins by separating pieces of work from the employment relationship. Entry-level jobs are unusually exposed to that process because they contain many of the tasks that are easiest to describe, repeat, check, and standardize. Those are exactly the kinds of tasks organizations are most tempted to hand to AI first.

The result can arrive without mass unemployment. A company can keep the same number of employees and change what the youngest employees do. It can hire fewer beginners. It can hire the same number but expect more from each one. It can ask a first-year worker to use AI and produce work that once required a third-year worker. It can also decide that some junior tasks are worth preserving because they train judgment. Different firms will make different choices. The common feature is that the first rung of the ladder is being redesigned.

Consider a law firm. Junior lawyers have traditionally spent enormous amounts of time researching cases, reviewing documents, comparing contracts, checking citations, and preparing early drafts. Much of that work is expensive for clients and not especially exciting for the lawyer doing it. It is also how a new lawyer learns what good legal work looks like.

Generative AI can now assist with many of those tasks. That can be a genuine improvement. A junior lawyer who spends less time searching for a routine clause may spend more time thinking about the client’s actual problem. A young associate may see more matters, get feedback faster, and learn more quickly. The optimistic case is real.

But there is another possible path. If the firm concludes that one experienced lawyer with AI can do work that once required several junior lawyers, it may simply hire fewer junior lawyers. The immediate economics look good. Salaries disappear before the need for future partners does. The problem arrives later, when the firm discovers that experienced professionals are not a natural resource. They are manufactured slowly from inexperienced ones.

The same pattern can appear in accounting, consulting, banking, software, marketing, insurance, journalism, and corporate management. In each field, AI is most attractive when it can take a repeatable piece of work off a person’s desk. In each field, some of those repeatable pieces were also practice. The employer can save the cost of the practice while still expecting the expertise to exist later. That is a very human accounting trick: remove the expense now and act surprised when the asset is missing later.

A Real Labor-Market Signal

When researchers describe work as exposed to AI, they generally mean its tasks could be performed or assisted by the technology. The measures differ across studies. Exposure describes a possibility; it does not establish that an employer adopted AI or that a particular job disappeared.

We should be careful here because a weak job market for young people can have many causes. Interest rates matter. Economic growth matters. Remote work changed hiring. College attendance changed. Industries rise and fall for reasons independent of AI. A serious argument cannot look at every unemployed 23-year-old and blame a chatbot.

Still, by 2026 the evidence had become difficult to dismiss. A U.S. Census Bureau working paper by economist Lee Tucker examined early-career hiring using matched employer and employee data. It found a sizable and persistent decline in hiring of workers ages 22 to 24 after the introduction of ChatGPT in the industries and states most exposed to AI. The decline was driven mainly by fewer hires rather than a sudden wave of young workers being fired.

That distinction matters. The first visible effect of AI on a profession may be a job that is never posted, a graduate class that is smaller than it would have been, or a manager who replaces two planned junior hires with software and one more experienced employee.

The Census study is evidence, not a final causal verdict. The author examines earlier labor-market trends, remote work, education, and monetary policy as competing explanations. His decomposition attributes at most about one-quarter of the relative employment decline through mid-2025 to monetary-policy shocks and finds no evidence that those shocks explain the rapid drop in hires at the most exposed firms. That measured caution makes the study more useful. The pattern is unusually strong where AI exposure is unusually high, and the timing is consistent with AI becoming part of the explanation.

The Pattern Gets Harder to Ignore

A separate Stanford Digital Economy Lab study used payroll data covering millions of U.S. workers. Its first release found a disproportionate decline for young workers in occupations highly exposed to generative AI. A later revision found that the young-worker employment gap had widened further, while more experienced workers in the same kinds of occupations showed no comparable gap.

The Stanford researchers also found no evidence of widespread economy-wide job displacement. That is crucial. The pressure appears uneven, showing up first among some of the people trying to enter occupations where AI can perform more of the routine cognitive work.

That is exactly what we would expect if the old entry-level bargain were weakening. Experienced workers bring context, relationships, authority, judgment, and knowledge of the organization. A machine may make those workers more productive. A beginner brings less of that accumulated value and historically compensated by doing more of the basic work. If AI can do much of the basic work, the experienced employee may become more valuable while the beginner becomes harder to justify on a spreadsheet.

This is not a law of nature. It is an incentive. Incentives can be redesigned.

The Entry-Level Job May Become More Senior

One of the stranger possibilities is that entry-level jobs survive but stop feeling entry level. Employers may still want young workers, but they may expect those workers to arrive with judgment, communication skills, initiative, and the ability to supervise AI outputs almost immediately.

That pattern was visible in 2026 hiring research. In NACE’s spring survey of 185 employers, overall hiring for the Class of 2026 was projected to rise 5.6%, with the strongest gain concentrated among employers with more than 5,000 employees, where planned hiring rose 8.7%. At the same time, employers were rapidly increasing their expectations for AI skills in entry-level candidates. PwC’s much larger job-posting analysis found that the most AI-exposed U.S. entry-level roles were 7 times more likely to require skills traditionally associated with more senior workers. Those “seniorized” roles grew 35% since 2019 while other entry-level postings fell 10%.

The phrase “entry level” may therefore survive while the meaning changes underneath it. A company may say, in effect: we still want beginners, but we want beginners who can perform like people with several years of experience because the machine is taking care of the simple work.

There is an obvious problem. Where does a 22-year-old get three years of judgment before receiving the first three years of work? Employers have always complained that young applicants lack experience. AI can turn that old complaint into a structural contradiction.

The Law Firm Test

The legal profession offers a useful real-world test because its training ladder is so visible. Reuters reported that firms with more than 500 lawyers employed 6,588 graduates from the Class of 2025, about 540 fewer than the prior class, a 7.5% decline and the first such decline since the Class of 2014. The class’s overall employment rate still reached 92.8%, in part because the graduating class itself had shrunk 7.0%.

AI is one factor among several. Government hiring, public-interest funding, firm economics, and normal business cycles all matter. The direction is still notable because the profession is simultaneously adopting tools that can perform legal research, document review, contract comparison, and drafting work that once consumed huge amounts of junior time.

Law schools and firms are already confronting the training problem directly. The debate has moved beyond whether young lawyers should use AI. The harder question is how they can use it without skipping the mental work that produces legal judgment. That is a more mature question than banning the technology or pretending nothing has changed.

Every profession will eventually face its own version of the law firm test: which beginner tasks are merely expensive, and which beginner tasks are expensive because learning itself takes time?

Efficiency Can Create a Training Debt

Imagine a company that automates 80% of the routine work done by its junior analysts. In the first year, the result may look excellent. Fewer people produce more output. Senior employees spend less time correcting formatting and searching for facts. The company celebrates a productivity gain.

Five years later, it still needs people who know how to question a forecast, recognize a weak assumption, understand a client, negotiate a disagreement, and take responsibility for a decision. Those abilities do not automatically appear because the software produced thousands of good drafts. Human judgment grows through use.

If the organization has eliminated most of the situations in which young people develop that judgment, it has borrowed productivity from the future. It kept the output and stopped funding part of the learning process that made future experts possible.

This is training debt. A company can save money by removing work beginners used to do. If it does not replace the learning inside that work, the bill arrives later: too few experienced people, lost practical knowledge, and managers who can accept an AI answer but cannot recognize a bad one.

Chapter Three described permission debt: a system’s powers grow while the human approvals and checks fall behind. Training debt is a different bill, created by neglecting the learning that future judgment will require. Both can build up while the organization believes everything is going well.

The irony is obvious. The better AI becomes at producing acceptable work, the more important it may become to preserve humans who understand how the work became acceptable.

The Better Path: AI as an Accelerator of Apprenticeship

Preserving apprenticeship does not require preserving bad work for sentimental reasons. Young accountants need not spend weeks doing something a machine can do in seconds merely because older accountants once suffered through it. Suffering is not a curriculum.

In February 2026, PwC UK made that distinction unusually concrete. The Financial Times reported that the firm received about 60,000 applications for 2,000 entry-level places for its 2026 intake, while recent graduate intakes had been a couple of hundred smaller than in the past. PwC had also abandoned an earlier target to add 100,000 people to its worldwide headcount by mid-2026. Yet the firm had chosen to retain some routine junior work rather than automate or offshore it because younger staff needed opportunities to develop professional judgment, including the ability to recognize subjectivity in audit questions that can look simple to an AI system. Chief People Officer Phillippa O’Connor described the strategic temptation to remove a larger share of the bottom of the traditional consulting pyramid, then framed the longer-term problem: if firms remove that work now, how will future managers acquire the nonautomatable skills the business will still need? She also said recent UK job cuts were driven much more by market conditions than by AI. PwC was embracing AI while deliberately preserving part of the learning process. The useful principle is straightforward: preserve the apprenticeship, not the inefficiency.

The better goal is to keep the learning while removing unnecessary work. A junior employee can use an AI draft, then explain its assumptions, check the facts, and defend the recommendation to a mentor. A young lawyer still needs to read the relevant cases. A new programmer needs to test and explain the code. An analyst needs to identify the assumption that could overturn a forecast. The finished answer is only part of the assignment.

That approach can make apprenticeship faster rather than weaker. The beginner gets more repetitions, more examples, and quicker feedback. A good mentor can spend less time correcting punctuation and more time teaching judgment. The machine becomes a practice partner rather than a substitute for practice.

This outcome requires design. A firm measured only on this quarter’s cost has an incentive to remove junior work rather than rebuild junior learning. Somebody has to own the second problem.

Current Evidence: Signals in the Entry-Level Market

Reviewed September 9, 2026. Graduate-unemployment evidence added September 29, 2026.

The evidence at this date points to a mixed labor market rather than a simple employment apocalypse. U.S. Census research found a pronounced reduction in early-career hiring in the industry-state groups most exposed to AI, with regression-adjusted employment for workers ages 22 to 24 down 12% over the 10 quarters after ChatGPT’s introduction. The hiring rate largely recovered by early 2025, but from a smaller employment base, which is why the employment gap persisted. The author cautioned that other forces may explain part of the change and that some trends began earlier.

Stanford’s August 2026 revision, using ADP payroll data through June 2026, found no broad economy-wide displacement but reported that employment for workers ages 22 to 25 in highly AI-exposed occupations was 19% below the path implied by less-exposed peers. The gap had widened from earlier versions of the study and was not mirrored among more experienced workers in the same way. The result also persisted after controlling for occupational exposure to interest-rate increases.

At the same time, the U.S. graduate market remained far from dead. NACE’s spring 2026 survey projected 5.6% growth in hiring for new college graduates. The stronger conclusion is therefore narrower: AI appears to be changing who gets hired, what beginners are expected to do, and how many beginners some highly exposed organizations need. The old entry-level job is under pressure before the idea of entry-level employment itself has disappeared.

A September 2026 working paper by Robert Fairlie and Jane Wu provides a counterweight. Using U.S. household-survey data for June through August, the researchers found no statistically significant increase in recent college graduates’ unemployment relative to the comparison groups studied. That does not establish that AI has no effect on hiring. Graduate unemployment and employment within highly exposed occupations measure different things, and a short observation period cannot settle the longer-term question. Employers should examine their own hiring and learning opportunities rather than assume that a national headline explains what is happening inside their organization.

On October 16, 2026, Anthropic reported that the share of its technical hires in junior roles had fallen from 27% to 9% since early 2025, even as the company grew. It described moving existing staff into new roles, while acknowledging that this does not help people who never get hired. This is the company’s study of its own workplace, not a national estimate or proof that AI alone caused the change.

An employer’s AI plan should therefore account for people outside its current payroll. If beginner assignments disappear, where will new workers practice and receive feedback? Protecting current staff and creating a route into skilled work are separate responsibilities. Each needs a named owner and evidence that the promised learning actually happens.

The evidence should continue to be updated. If hiring rebounds broadly across highly exposed occupations, the argument should narrow. If the young-worker gap keeps widening while experienced employment remains stable, the training-ladder explanation becomes harder to ignore.

Accountability Analysis

The Delegation Doctrine does not tell an employer how many junior workers to hire. It asks a more basic question: who owns the consequences of the delegation? If a company transfers beginner tasks to AI, the decision owner also inherits responsibility for what those tasks used to produce besides output.

Principle 5 says that speed and architecture move accountability upstream. In this case, the architecture is partly the career ladder itself. A manager who redesigns a department around AI is responsible for more than this year’s productivity. The manager is also making choices about where future competence will come from, which employees will receive real practice, and whether the organization will still contain people capable of challenging the machine later.

Principle 8 is even more direct. Accountability decays when human competence is not renewed. A workplace that depends on human judgment at the top cannot be indifferent to how that judgment is created at the bottom. If the organization wants experienced humans in 2036, it must decide how inexperienced humans in 2026 are supposed to become them.

An employer can choose whom to hire. But it cannot assume that other employers will always train the experienced people it needs. If every firm saves money by cutting beginner opportunities, fewer people may gain the experience everyone later wants to buy.

The accountability leak appears when the productivity decision has a named owner but the training consequence belongs to nobody.

Accountability Verdict

AI is unlikely to end the entry-level job. It is much more likely to end many of the tasks that used to define it. That can be good. It can remove drudgery, reduce errors, lower costs, and allow young workers to contribute at a higher level sooner.

But the first rung cannot simply be deleted because the elevator got faster. Employers, schools, professional bodies, and workers themselves will have to build new ways to create experience. The best organizations will teach young people to use AI while developing the judgment, context, skepticism, communication, and responsibility that matter above the tasks AI performs cheaply.

For the individual entering the workforce, the practical lesson is equally clear. Do not build a career around being the cheapest producer of a first draft. Learn to direct the tool, verify the result, understand the underlying work, explain the decision, and take responsibility for what leaves your hands. The value of being human is moving upward from production toward judgment, but judgment still has to be learned somewhere.

The old entry-level job was never just a job. It was a training system disguised as cheap labor. If AI changes the labor, we must redesign the training. Otherwise, we may discover too late that we automated the path to expertise while assuming the experts would somehow keep arriving.

Keep the learning in the task

If you are learning a job, choose one AI-assisted assignment and explain the result to an experienced colleague. Show which facts you checked, which assumption matters most, and what would make you change the answer. Ask for feedback on your reasoning.

If you manage beginners, identify what one routine assignment used to teach. When AI takes over part of it, give the learner another way to practice that skill and receive feedback. Assign someone to check progress. That is how a plan to save time becomes a plan to develop judgment too.

Sources and Notes

1. U.S. Census Bureau, Lee C. Tucker, “You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators,” CES Working Paper 26-27, April 2026. https://www2.census.gov/library/working-papers/2026/adrm/ces/CES-WP-26-27.pdf. Accessed September 9, 2026. The paper reports an immediate and persistent decline in early-career hiring in the most AI-exposed industry-state cells and a 12% regression-adjusted employment decline for ages 22 to 24 over the 10 quarters after ChatGPT’s introduction. It also reports that the hiring rate largely recovered by early 2025 from a smaller employment base and estimates that monetary-policy shocks explain at most about one-quarter of the relative decline through mid-2025. Source record: DA-CH6-S1. Archive reference: pending.

2. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab, revised August 2026. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/. Accessed September 10, 2026. The revision uses ADP payroll data through June 2026, reports no widespread economy-wide displacement, and finds employment for ages 22 to 25 in highly AI-exposed occupations 19% below the path of less-exposed peers, primarily through reduced hiring. Source record: DA-CH6-S2. Archive reference: pending.

3. Kathryn Bonney, Cory L. Breaux, Emin Dinlersoz, Lucia S. Foster, John C. Haltiwanger, and Aditya A. Pande, “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks,” U.S. Census Bureau CES-26-25 and NBER Working Paper 35141, revised July 9, 2026. https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-25.html and https://www.nber.org/papers/w35141. Accessed September 9, 2026. Used as continuity evidence from Chapters Two and Five regarding actual business adoption, task augmentation, and the rarity of reported AI-related employment reductions among AI-using firms. Source record: DA-CH6-S3. Archive reference: pending.

4. National Association of Colleges and Employers, “Job Outlook 2026: Spring Update,” April 2026, with accompanying analysis “Demand for AI Skills in Entry-level Jobs Nearly Triples Since Fall 2025,” April 20, 2026. https://naceweb.org/research/reports/job-outlook/ and https://naceweb.org/job-market/trends-and-predictions/demand-for-ai-skills-in-entry-level-jobs-nearly-triples-since-fall-2025. Accessed September 9, 2026. The survey included 185 employers, projected 5.6% growth in Class of 2026 hiring, and found rapidly increasing expectations for AI skills in entry-level candidates. Source record: DA-CH6-S4. Archive reference: pending.

5. Reuters, “Entry-level hiring at large US law firms declined for the first time in a decade, data shows,” August 5, 2026, reporting National Association for Law Placement data. https://www.reuters.com/legal/government/entry-level-hiring-large-us-law-firms-declined-first-time-decade-data-shows-2026-08-05/. Accessed September 9, 2026. Firms with more than 500 lawyers employed 6,588 Class of 2025 graduates, about 540 fewer than the prior class, a 7.5% decline; the class’s overall employment rate was 92.8%, partly against a graduating class that shrank 7.0%. Source record: DA-CH6-S5. Archive reference: pending.

6. PwC, “2026 Global AI Jobs Barometer,” June 15, 2026. https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html. Accessed September 9, 2026. PwC’s job-posting analysis found that the most AI-exposed U.S. entry-level roles were 7 times more likely to require traditionally senior skills; those roles grew 35% since 2019 while other entry-level postings fell 10%. This is employer-advisory research rather than independent administrative labor-market data and is treated accordingly. Source record: DA-CH6-S6. Archive reference: pending.

7. Financial Times, “PwC UK applications jump 35% in graduate jobs drought,” February 13, 2026. https://www.ft.com/content/915a32cf-d62d-4caf-b539-8525e52257ad. Republished by The Irish Times: https://www.irishtimes.com/business/2026/02/13/pwc-uk-applications-jump-35-in-graduate-jobs-drought/. Accessed September 10, 2026. The report says PwC UK received about 60,000 applications for 2,000 entry-level places for its 2026 intake; recent graduate intakes were a couple of hundred smaller than in the past; PwC had abandoned an earlier target to add 100,000 people to worldwide headcount by mid-2026; and the firm deliberately retained some routine junior work so early-career staff could develop judgment. Chief People Officer Phillippa O’Connor described the long-term risk of shrinking the bottom of the traditional pyramid too far and said recent UK job cuts were driven much more by market conditions than by AI. Source record: DA-CH6-S7. Archive reference: pending.

8. Robert W. Fairlie and Jane Wu, “The Early Impacts of AI on Employment among Recent College Graduates,” IZA Discussion Paper No. 18945, September 2026. https://www.iza.org/publications/dp/18945/the-early-impacts-of-ai-on-employment-among-recent-college-graduates. Accessed September 29, 2026. The authors’ abstract reports Current Population Survey microdata for June, July and August 2026, with older graduates and young nongraduates as comparison groups. This working-paper finding concerns relative unemployment, not proof that AI has no hiring effects. The measurement distinction and employer recommendation are author analysis. Source record: DA-CH6-S8. Archive reference: pending.

9. Nathan Wilmers, Regina Catipon, Maxim Massenkoff, and Peter McCrory, “How employment is changing at Anthropic,” The Anthropic Institute, October 16, 2026. https://www.anthropic.com/research/employment-at-anthropic. Accessed October 17, 2026. Company study using internal employment and work records and employee interviews. The hiring statistic is a share of technical hires, not a count of jobs lost. Observational evidence from one rapidly growing company does not establish economy-wide effects or isolate AI causation. Used for the reported hiring composition and distinction between internal reallocation and opportunities for people not hired. The recommendation about responsibility for entry routes is the author's analysis. Anthropic is also used in this book's external critique workflow; this report is not independent validation. Source record: DA-CH6-S9. Update 101. Archive reference: pending.