# What Happens to Entry-Level Work When AI Does It?

Canonical URL: https://blog.theaieducator.io/posts/what-happens-to-entry-level-work
Publication: Dan Fitzpatrick Insights
Author: Dan Fitzpatrick
Topic: Future Skills and the Future of Work
Published: 2026-09-13T07:15:11.000Z
Modified: 2026-09-13T07:15:34.282Z

Entry-level work is being hollowed out rather than abolished. The risk is not that you stop hiring juniors, but that you keep hiring them into jobs that no longer teach them anything. A three-question filter for deciding what to automate.

## In brief

Entry-level work is being reshaped rather than eliminated: Stanford's August 2026 payroll analysis puts 22 to 25 year olds in AI-exposed occupations about 19 percent below their peers, driven by reduced hiring, while 87 percent of UK employers expect graduate roles to change and only 18 percent expect to cut more than a tenth of them. The leadership risk is not headcount. Junior tasks did three jobs at once: they produced the work, they produced the worker, and they produced the evidence that the worker could be trusted. Automation takes over the first and simply stops the other two. Before handing a junior task to AI, Dan Fitzpatrick's Practice Question asks what doing it badly taught the person who used to do it, where they will learn that now, and who would notice if they never did. A task that fails any of the three should be retired, rehoused or reversed, out loud, with a named owner.

## Key takeaways

- Stanford Digital Economy Lab's August 12, 2026 update finds workers aged 22 to 25 in highly AI-exposed occupations about 19 percent below where employment would be had it tracked less-exposed peers, up from 15 percent in July 2025, driven by reduced hiring rather than more separations.
- The Institute of Student Employers found 87 percent of 144 employers expect AI to reshape graduate and apprentice roles, but 40 percent expect to eliminate none and only 18 percent expect to cut more than a tenth, so the story is reshaping rather than replacement.
- The finding that matters most for leaders is that 43 percent of those employers said junior roles had already changed without any formal redesign: the job is being altered by accretion, not by decision.
- Every junior task did three jobs at once: it produced the work, it produced the worker through the attempt-and-correction loop, and it produced the evidence that the worker could be trusted with more. Automation takes over only the first.
- The Practice Question is Dan Fitzpatrick's suggested filter before automating any junior task: what did doing this badly teach the person who used to do it, where will they learn that now, and who would notice if they never did.
- A task that fails The Practice Question should be retired, rehoused or reversed. Reversing means the person attempts and the machine checks, preserving the order that builds judgment rather than the order that is merely faster.
- A competency list will not carry the weight on its own: in a KPMG and University of Texas field study of 523 early-career professionals reported in Harvard Business Review in July 2026, critical thinking, AI literacy and domain knowledge were not clear signals of who added the most value alongside AI.

Somewhere in your organization there is a task that a twenty-three-year-old used to do badly for about six months, and then do well. It was slow. It produced work that someone more senior had to check. Everyone was quietly relieved when a machine took it over. That task was your training program, and nobody ever wrote it down.

Here is the answer before the detail. Entry-level work is not vanishing so much as being hollowed out, and the real exposure for most organizations is not that they stop hiring people at the start of their careers. It is that they keep hiring them into jobs that no longer teach them anything. The output was always the smaller half of what that work produced.

## What the evidence actually says about AI and entry-level work

The honest read of the data is narrower and more interesting than the headlines: there is no economy-wide displacement, but young workers in the occupations AI is best at have fallen behind their peers, and the gap is still widening.

Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the [Stanford Digital Economy Lab](https://digitaleconomy.stanford.edu/news/canariesaug26/) published their August 12, 2026 update to "Canaries in the Coal Mine?", built on ADP payroll records. Workers aged 22 to 25 in highly AI-exposed occupations now sit about 19 percent below where employment for that group would be had it tracked their less-exposed peers, up from 15 percent in July 2025. Two details matter more than the number. The adjustment is coming through reduced hiring rather than more people being let go, and the declines concentrate in roles where AI automates the work rather than assists with it. The authors are careful, and so should we be: they call these descriptive patterns, not causal estimates, the effect is larger in the ADP sample than in national survey benchmarks, the gaps shrink once education is controlled for, and some of the divergence predates generative AI.

In the United Kingdom the picture from employers is calmer. The Institute of Student Employers surveyed 144 organizations for its [Student Development Survey 2026](https://ise.org.uk/knowledge/insights/562/entrylevel_work_reshaped_not_replaced/), published May 7, 2026. Eighty-seven percent expect AI to reshape graduate and apprentice roles, but only 18 percent expect to eliminate more than a tenth of them, and 40 percent expect to eliminate none. What is going is routine admin and basic data and writing tasks. ISE's own summary is "reshaped, not replaced," and on the evidence they have, that is fair.

Then there is the finding that nobody built a headline out of. Forty-three percent of those employers said the roles had already changed without any formal redesign.

Put the two studies side by side and the leadership problem comes into focus. The junior job is changing. In more than four organizations in ten, nobody decided that it should.

## Entry-level work was never mainly about the output

Every junior task did three jobs at once, and automation takes over only one of them.

Think about a first-year analyst reconciling a spreadsheet, a trainee teacher marking a set of books, a junior solicitor reading through a bundle of documents. The first job that task does is obvious: the spreadsheet gets reconciled. The second is the one that shows up nowhere in a workflow diagram. By the fortieth reconciliation, that analyst can smell a wrong number before they have found it. That is calibration, and it is built by attempting something, getting it wrong, and being corrected by someone who has seen it before. The third job is quieter still. Their manager watched them do it, and formed a view about what they could be trusted with next.

Hand the task to a machine and the first job still gets done, faster and at three in the morning. The second and third do not relocate. They stop.

This is the argument Bryan Hancock and Charlotte Seiler make in [McKinsey Quarterly](https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/building-expertise-in-the-age-of-ai-who-trains-the-next-generation), July 14, 2026, and they have numbers behind it: about a third of employers report that AI has already reduced the foundational, skill-building tasks early-career staff used to do. They reach for medical education to explain why that matters, where the attempt-then-check loop, the trainee having a go before seeing the right answer, is what produces expertise that lasts.

My reading goes one step further than theirs. Organizations have been treating junior tasks as work with a training side effect. They were the other way round. The work was the side effect. The expensive thing they produced, the thing you cannot buy in, was people who had been wrong often enough to develop judgment.

## The Mistake I See Most Often

The mistake is treating this as a hiring question when it is a design question.

When this lands on a leadership agenda, the question in the room is almost always some version of "do we still need to recruit at this level?" That question is not just hard to answer. It is aimed at the wrong thing. It assumes what is at stake is how many people are in the building, when what is at stake is whether anybody in the building is still learning to do the work well.

Across the leadership teams I work with, I have yet to sit in a room where someone could name the person responsible for the learning that used to happen inside a task they had automated. The task had an owner. The output had an owner. The learning never did, because nobody had noticed it was there.

I spent years as a classroom teacher and then as an assistant headteacher, and the thing that turned a new teacher into a good one was never the training day. It was the first two years of getting it wrong in front of thirty children and having somebody more experienced tell them what they had missed. Take that away and you do not get a teacher who needs less development. You get one who arrives at year five with a year one's judgment and a great deal of confidence.

## The Practice Question

Before you hand a junior task to a machine, ask what that task was teaching and where the teaching goes instead.

The Practice Question is three questions I ask before handing any junior task to AI: What did doing this badly, and being corrected, teach the person who used to do it? Where will they learn that now? And who would notice if they never did? A task that survives all three can be automated with a clear conscience. A task that fails any of them is not a chore. It is a classroom, and you are deciding to close it.

This is a filter I suggest, not a validated instrument. It is deliberately quick, because the decisions it applies to are made quickly and usually by one person.

Take a hypothetical that will feel familiar. A secondary school lets staff generate the first draft of parent reports with AI. Question one: what did writing those reports badly teach a new teacher? It forced them to put in plain language what they actually thought about a child, and then to discover at the parents' evening that followed whether they had been right. Question two: where does a teacher learn that now? Nowhere in the current design. Question three: who would notice? Nobody, for about three years. Then the school has a group of mid-career teachers who have never had to defend their own judgment about a child in writing, and it will be very hard to work out why.

## Retire, rehouse or reverse

Once a task fails The Practice Question, there are three honest things to do with it, and the point is to choose one out loud.

**Retire it.** Some junior work taught nobody anything. Chasing missing forms was never an apprenticeship. Automate it, and stop describing it as development for the people who no longer do it.

**Rehouse it.** The learning was real and still matters. Name where it happens now, who owns it, and how you will know it happened. If you cannot name all three, you have not rehoused the learning. You have mislaid it.

**Reverse it.** Keep the order that builds a person: they attempt, the machine checks. Most implementations run the other way round, because that way is faster, and it produces someone who is skilled at spotting a machine's mistakes and has never made enough of their own to recognize the shape of one. This is the same distinction I drew about oversight in [why a human in the loop is not enough](https://blog.theaieducator.io/posts/what-ai-agents-mean-for-work): checking is not the same activity as doing, and it does not build the same person.

## Why a skills list will not carry the weight

The standard answer to all of this is a list of competencies, and the largest field study to test those competencies did not find them predictive.

ISE's employers name four things they want more of from graduates: critical thinking, AI literacy, communication and adaptability. Sensible choices. But in [Harvard Business Review](https://hbr.org/2026/07/research-why-some-junior-employees-work-well-with-ai-and-others-dont) on July 23, 2026, Ashish Agarwal, Anitesh Barua, Anu Puvvada, Fangchen Song and Wen Wen reported a field study run by KPMG and the University of Texas with 523 early-career professionals using a purpose-built AI agent on real business tasks. Participants sorted into three groups by how much value they added, which the researchers called AI apprentices, AI delegators and AI amplifiers. Critical thinking, AI literacy and domain knowledge, they found, "weren't clear signals for who fell into which group."

Hold that finding at its actual size. It is one study, inside one firm, with one kind of agent, and it does not show those capabilities are worthless. What it punctures is the assumption underneath most current responses: that you can recruit or train your way through this with a better competency list. And the screen itself is getting noisier at exactly the wrong moment. Two-thirds of ISE's employers now worry that candidates misrepresent their skills using AI during recruitment, up from around half a year earlier.

If the attributes you select for do not predict who thrives, the lever you have left is the design of the work. That is my interpretation rather than a finding, but it is the one that follows.

## What this means if you run a school or a college

If you educate the people who will fill these jobs, the sharper question is not what to add to the curriculum but what you are now letting students skip.

The pipeline argument applies inside education as much as to the employers on the other side of it. The tasks that turn a novice teacher into a competent one are being automated in schools this term, by individual staff, one sensible decision at a time. Nobody is tracking it, because each decision looks like relief.

For students, the ISE data is worth reading closely. The gaps employers named were adaptability, contextual understanding, planning and organization, and 29 percent reported more performance issues with new hires than they used to, against 12 percent in 2022. Those are not knowledge gaps. They are what a person acquires by owning something whole, carrying it to the end, and being judged on the result. That is a design choice a school makes every time it decides how much of a task a student is allowed to hand over, which is the through-line in [what schools should be preparing children for now](https://blog.theaieducator.io/posts/what-should-schools-be-preparing-children-for-now) and in the scarcer skills I set out in [what skills matter when everyone has access to AI](https://blog.theaieducator.io/posts/what-skills-matter-in-the-age-of-ai).

## Where to start this term

Take the last piece of junior work your organization automated. Not the biggest one, the last one. Run The Practice Question on it, write down retire, rehouse or reverse, and put a name against it. Then do the same for the next automation anyone proposes, before it goes in rather than after.

That one habit does something no policy has managed: it forces the learning to have an owner. Most organizations will find they can do it in ten minutes per decision, and that they have a backlog.

If your leadership team is working out what AI is doing to the shape of work and to the people who will be doing it in ten years, this is the kind of question I take on in [keynotes and leadership sessions](https://theaieducator.io/keynote).

## Sources and further reading

- Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, "Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence (August 2026 Update)", Stanford Digital Economy Lab, August 12, 2026. [Summary](https://digitaleconomy.stanford.edu/news/canariesaug26/)
- Institute of Student Employers, "Entry-level work reshaped not replaced", Student Development Survey 2026, May 7, 2026. [Link](https://ise.org.uk/knowledge/insights/562/entrylevel_work_reshaped_not_replaced/)
- Bryan Hancock and Charlotte Seiler, "Building expertise in the age of AI: Who trains the next generation?", McKinsey Quarterly, July 14, 2026. [Link](https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/building-expertise-in-the-age-of-ai-who-trains-the-next-generation)
- Ashish Agarwal, Anitesh Barua, Anu Puvvada, Fangchen Song and Wen Wen, "Research: Why Some Junior Employees Work Well with AI, and Others Don't", Harvard Business Review, July 23, 2026. [Link](https://hbr.org/2026/07/research-why-some-junior-employees-work-well-with-ai-and-others-dont)

*Dan Fitzpatrick is the founder of [The AI Educator](https://www.theaieducator.io/about), a Forbes contributor, and an international keynote speaker on AI, leadership and the future of work.*

## Frequently asked questions

### Is AI taking entry-level jobs?

Not wholesale. Stanford's August 2026 payroll analysis finds no economy-wide displacement, but a widening gap for 22 to 25 year olds in AI-exposed occupations, now around 19 percent, driven by slower hiring rather than more people being let go. UK employers report reshaping, with 40 percent expecting no role eliminations.

### Which entry-level tasks is AI actually absorbing?

The Institute of Student Employers points to routine administration and basic data and writing tasks: exactly the work that used to be handed to someone in their first year. McKinsey reports about a third of employers say AI has already reduced the foundational, skill-building tasks early-career staff once did.

### Should we still hire graduates and apprentices?

Usually yes, but the harder question is what they will do once hired. Hiring at this level only works if the role still contains attempt-and-correction work. A junior job stripped of everything difficult produces someone who has been employed for three years and practiced for none.

### How do you train junior staff when AI does the work they used to learn on?

Decide deliberately for each automated task. Retire it if it taught nothing, rehouse the learning somewhere with a named owner and a way of knowing it happened, or reverse the order so the person attempts the work first and the machine checks it afterward.

### What skills do employers say graduates are missing?

The Institute of Student Employers names adaptability, contextual understanding, planning and organization, and reports 29 percent of employers seeing more performance issues with new hires against 12 percent in 2022. These are not knowledge gaps. They come from owning a piece of work end to end.

### Is teaching critical thinking and AI literacy the answer?

It helps, but one large field study suggests it is not the whole lever. Researchers at KPMG and the University of Texas studied 523 early-career professionals and found critical thinking, AI literacy and domain knowledge were not clear signals of who added the most value working with an AI agent.

### What should a leadership team do first?

Take the last piece of junior work your organization automated and run The Practice Question on it: what did it teach, where is that learned now, and who would notice if nobody learned it. Then write down retire, rehouse or reverse, and put a name against the decision.

---
Source: [What Happens to Entry-Level Work When AI Does It?](https://blog.theaieducator.io/posts/what-happens-to-entry-level-work)
Publisher: [The AI Educator](https://theaieducator.io)