The work AI removes was also how people learned

EY has just put $100 million behind business acumen, professional judgement and adaptability. The easy interpretation is that one of the Big Four has decided human skills matter because AI cannot replace them. That is not quite what happened.

EY’s own announcement also rewards technology adoption, innovation and people who use advanced technologies to turn disruption into opportunity. This is not a retreat from AI. It is an attempt to work out what becomes more valuable once AI is embedded deeply enough into the work.

What interests me is the timing.

Two weeks ago, I ended Raising the floor is not building a ladder with a question:

If production was doing the teaching, and production has been absorbed, who is designing the replacement?

My answer was: nobody, as far as I could tell. EY is the first clear counterexample I’ve found. And it makes the ladder problem much more concrete.

The junior work had two jobs

A lot of entry-level professional work looks inefficient if you measure only the immediate output.

A graduate auditor checks evidence, follows procedures, reviews transactions, prepares schedules and investigates exceptions. A junior accountant cleans records, works through month-end processes and answers apparently simple client questions. A new consultant builds analyses a more experienced colleague could probably finish faster.

AI and automation can absorb a growing share of that work. From a productivity perspective, that can be an obvious gain.

But the work was doing two jobs.

It produced the output. It also exposed the junior person to hundreds of small cases from which pattern recognition, professional scepticism, business understanding and judgement accumulated.

The second function rarely appeared in a business case because nobody had to buy it separately. The apprenticeship was embedded inside the operating model.

That creates a problem for automation programmes: a task can be low-value as production and high-value as learning infrastructure.

Remove it and the organisation may save time now while weakening the mechanism that produces experienced people later.

This is the same distinction I was trying to make in the ladder piece. AI can raise the quality of early-career output. That is a claim about production. It tells us much less about whether the person is accumulating the judgement the output used to reveal.

EY gives us a useful place to see what happens when that stops being theoretical.

EY is already talking about the apprenticeship problem

The $100 million announcement is only half the story. Earlier in August, EY published a discussion with Fortune 100 chief accounting officers and controllers about how human and agent workforces may develop together. One of the issues raised was exactly this: if AI changes the entry-level work, the apprenticeship model has to change with it.

The participants distinguished between aggregating information and understanding it. The important part was that some of that understanding had historically been built by doing the aggregation. Accountants increasingly need to become business leaders and advisers, but the work through which they used to build the underlying knowledge is changing.

Then, on 17 August, EY announced its Career Residency.

Instead of a conventional eight-week internship, the programme runs for eight to 12 months and combines client work with coaching, critical thinking, professional judgement, curiosity, AI fluency, collaboration and communication.

Put the three things next to each other:

Routine work is increasingly automated.

Judgement and business acumen are becoming more valuable.

The organisation is building a more deliberate mechanism for producing them.

That is more interesting than another list of supposedly irreplaceable human skills.

EY is beginning to redesign the ladder.

The old pyramid contained a learning system almost by accident

Professional-services firms made this unusually visible because their organisations were built around a pyramid.

Large numbers of junior people did large volumes of work. More experienced people reviewed it. Repeated exposure moved some of those juniors upwards until they became the people capable of making the higher-stakes calls.

The model had obvious problems. It could be repetitive, exhausting and inefficient. It should not be romanticised simply because it trained people.

But it did train people. AI changes the economics of the pyramid because it can make some of the work at the bottom cheaper, faster or unnecessary.

The temptation is to treat the removed work as pure waste. The harder question is what else disappeared with it.

That is where I think the usual AI productivity discussion is too narrow. We calculate hours saved from completing the task. We are much worse at calculating the developmental value of having completed it 200 times.

The benefit appears immediately. The cost can take years to become visible.

Fewer people may also be getting onto the first rung

The labour-market evidence is still too messy to turn this into a simple AI-causes-fewer-jobs story. But the direction is worth watching. The Financial Times analysed more than 50,000 Big Four job listings and found AI-related roles had risen to almost 7% of postings while audit roles were below 3%.

That does not tell us total headcount, and it does not prove AI caused a reduction in traditional roles. It does show recruitment capital moving towards a different capability base.

The broader issue is the one I added to the ladder article as more evidence arrived: fewer entry-level opportunities and weaker judgement formation are related, but they are not the same problem.

One removes rungs. The other changes what people learn while standing on them.

A company can therefore end up with two pressures at once: fewer juniors entering the system, and less routine work available to train the juniors who do enter. Meanwhile, it still needs experienced people at the top.

Accounting makes the problem unusually clear

This is one reason the EY example feels so relevant to the work around accounting and finance.

The standard promise around AI in accounting is plausible: automate more reconciliation, compliance, data preparation and workflow administration so accountants can spend more time advising clients and making better decisions.

Sage is making versions of that argument publicly too. Its work on AI operational skills in accounting describes accountants moving towards directing workflows, reviewing outputs and applying judgement. Its research on the AI skills gap says lack of skilled people is already a barrier to implementation.

There is nothing inconsistent about those claims. But there is a missing step between them. Removing reconciliation does not automatically create an adviser.

Someone still has to develop the commercial understanding, confidence, pattern recognition, communication skill and professional judgement required to use the freed capacity well. The technology can move the value of the role upstream faster than the profession can move the person upstream. That is the ladder problem in another industry.

It also connects to the broader argument in The editorial jobs are moving upstream. Different profession, same structural shift: as machines absorb more execution, the value of human work concentrates around judgement, interpretation, evidence and responsibility.

But saying the valuable work moves upstream is only half an answer.

You still need a route upstream.

The hidden cost may be apprenticeship debt

I’m using apprenticeship debt as a working term for the problem.

The organisation removes an old learning mechanism faster than it builds its replacement.

For a while, that can look like success. The experienced people are still there. Output is faster. Costs fall. Junior staff can use good systems to produce credible work much sooner.

The debt becomes visible later, when there are fewer people with enough accumulated experience to supervise automated systems, challenge unusual outputs or take responsibility for consequential decisions. It is similar to technical debt in one useful sense: the short-term gain can be completely real while still creating a deferred obligation. The answer is not to preserve bad manual work for sentimental reasons.

Making graduates spend thousands of hours doing something a system can do better is not a serious workforce strategy. The better question is:

If a task no longer needs a human to perform it, what experience does the human now need in order to judge, challenge and govern the system that performs it?

That turns automation into a workforce-design problem rather than a task-removal exercise.

Judgement has to become infrastructure

This also extends an argument I’ve been making elsewhere in Editorial Intelligence.

As AI makes competent output abundant, judgement becomes more valuable. But that formulation can become lazy too. It starts to sound as though judgement is a magical human quality we simply need to protect from the machines.

It isn’t. Judgement is built.

It comes from exposure, comparison, feedback, mistakes, coaching, evidence, accumulated context and having to make decisions whose consequences you can see. That means organisations increasingly need to treat the production of judgement as infrastructure.

A post-AI apprenticeship might involve deliberate exposure to exceptions and edge cases, senior review of model outputs, case libraries, simulations, decision journals, client conversations and explicit discussion of why a recommendation was accepted or rejected.

Maybe reviewing far more AI-generated cases becomes a better apprenticeship than producing everything manually ever was. I raised that possibility in the ladder piece and still think it matters. There is no reason to assume the old route was optimal simply because it was ours.

But the new route has to exist. That is the part EY makes more interesting. When I wrote Raising the floor is not building a ladder, I could see the disappearing apprenticeship but not anyone clearly taking responsibility for replacing it.

Now I can point to at least one organisation trying. The $100 million in awards tells us what EY thinks is becoming valuable. The Career Residency tells us something more important: it may have realised those capabilities have to be produced deliberately.

The real question is therefore not whether human skills survive AI. It is whether organisations can still produce enough humans with those skills once the work that used to train them has been automated. That’s a much harder problem than teaching everyone to prompt.