AI Is Removing the Work That Trained Your Next Generation of Advisors

Quick answer

AI has automated the routine tasks that junior accountants once used to build judgment: reconciliations, first-pass reviews, schedule prep and basic research. PwC’s 2026 analysis of 2.4 million US entry-level jobs found that the entry-level roles most exposed to AI are seven times more likely to demand traditionally senior skills than the least exposed. The profession removed the training ground and kept the expectation. The macro evidence on entry-level headcount is genuinely contested, but the design problem underneath it is not, and it belongs to firm leadership rather than to economists.

The routine work performed a second function

For most of the past two decades, the profession has treated the bottom of the pyramid as low-value work. Reconciliations, first-pass reviews, schedule prep and basic research were the tasks firms pushed offshore, pushed down, or pushed into software, and each of those decisions was defensible on its own terms. What went largely unexamined was the second function that work performed alongside the first.

It was the mechanism by which a 24-year-old learned what good looked like. Working through it, they saw two hundred versions of a thing before anyone let them near a client, and they developed the instinct that something in a file was wrong before they could articulate why. No firm designed it as a development system. It simply was what the work happened to produce, and because it produced it reliably for the better part of a century, nobody thought to write it down.

Over roughly eighteen months, the profession automated it. When PwC published its 2026 Global AI Jobs Barometer in June 2026, Global Workforce Leader Pete Brown described the mechanism precisely: “AI is removing some of the routine work that once acted as an apprenticeship.” No managing partner voted to eliminate the training ground. It disappeared as a byproduct of a hundred efficiency decisions, each individually correct, none of which accounted for the second function.

What the data shows

PwC analysed more than one billion job advertisements across 27 countries and territories, and within that dataset ran a targeted analysis of 2.4 million US entry-level jobs. Three findings bear directly on how firms develop people:

• Junior roles are being asked to do senior work. Entry-level roles most exposed to AI are seven times more likely to demand traditionally senior skills such as leadership and strategic thinking than the least AI-exposed entry-level roles. The comparison matters: this is not seven times more than five years ago, but seven times more than comparable junior roles AI has not reached.
• Entry-level postings are splitting into two tracks. Openings for these “seniorised” roles grew 35% since 2019, while non-seniorised roles within the same most-exposed group declined 10%. The junior job is not disappearing so much as bifurcating, with one track expanding and asking for capabilities a 22-year-old has no plausible route to having acquired.
• The premium moved toward human capability, not technical capability. PwC’s full report applies the EPOCH framework developed by Loaiza and Rigobon, which scores how far a task depends on empathy, presence, opinion, creativity and hope. New tasks added to AI-exposed roles since 2022 score 2.5 times higher on human intensity than new tasks in the least exposed roles.

That third finding is the one that catches most firms off guard, and it has an immediate practical consequence. The work AI leaves behind is not more technical; it is more human and more demanding of judgment. A firm whose entire response to AI is a technical training programme is therefore training for the wrong thing, and will produce people who can operate the tool competently while remaining unable to do the work the tool created.

The evidence is contested; the design problem is not

We are not going to tell you the macro picture is settled, because it is not, and any advisor presenting it as settled is running ahead of the data. Two credible sources point the other way, and firm leaders should know about both before they act.

The Federal Reserve Bank of New York published analysis on 14 May 2026 examining whether job postings show early labour-market effects of AI. Economists Richard Audoly, Miles Guerin and Giorgio Topa ran an event study comparing junior and senior postings within highly AI-exposed occupations and found no clear divergence between them, concluding that demand for the two is moving broadly in parallel. Anthropic’s labour market research from March 2026 reached a similar place from a different direction, finding no systematic increase in unemployment among the most exposed workers. It did detect roughly a 14% decline in the job-finding rate for workers aged 22 to 25 entering high-exposure occupations, though the authors flagged that result as barely statistically significant themselves. One caveat that most commentary misses is worth carrying into any partner discussion: the New York Fed study uses Anthropic’s exposure metric, so the two are not independent, and treating them as separate confirmation overstates the case.

None of this changes what a firm should do this quarter, for a reason that survives whichever side eventually wins. Whether or not the entry-level job disappears, the entry-level task is disappearing, and that point is not in dispute anywhere in the literature. The argument is about headcount, not about the work. Since your development model was built on those tasks rather than on those headcount numbers, you can either wait five years for the economists to settle the question or accept that the mechanism your firm used to turn junior people into senior people has been quietly disassembled. It has been disassembled either way.

The cost arrives four years late

This is a problem that is easy to defer precisely because deferring it produces no symptoms. Nothing breaks this year. Your seniors are still senior, your managers still manage, and the people who learned under the old model are still in the building with their calibration intact, so work continues to leave the firm at the standard clients expect. The gap opens in the cohort behind them, and it opens slowly.

In four years, a firm that changed nothing will have a group of people holding the title of manager who were never actually made into managers, because the mechanism that used to do it was switched off and nothing replaced it. They will be bright, well-credentialed and fluent with AI from their first week, and they will lack the calibration that came from producing four hundred versions of something badly and being corrected each time. Leadership will notice it as review time creeping upward, as partners quietly pulling work back, and as a widening gap between what the engagement letter promises and who is actually capable of delivering it.

By that point AI will be unremarkable, so nobody will connect the symptom to the cause. The firm will conclude that it has been hiring badly. It has not been hiring badly. It stopped training, without ever deciding to.

Four decisions worth making now

1. Write down what the apprenticeship actually taught. The outcome was never the reconciliation itself; it was pattern recognition at volume, exposure to what wrong looks like, and calibration against a partner’s correction. Those outcomes can be produced by other means, but only by a firm that can articulate what it is trying to produce. Most cannot, because the system was never designed. This costs an afternoon and everything else depends on it.
2. Decide which work stays human on purpose. The reason is not that AI cannot do it. The reason is that a person needs to do it in order to become someone who can evaluate it later. Almost no firm is making this decision consciously, because tasks drift toward automation by default whenever nobody names a reason to protect them. Naming the reason is the entire intervention.
3. Budget for it and defend the budget. Protected development work carries a cost and will surface in somebody’s utilization. Left implicit, the first difficult busy season will delete it and no one will notice until the four-year mark. Treat it as you would any other investment with a long payback and an unattractive first year.
4. Stop treating judgment as a soft skill. PwC’s data indicates the market is repricing human capability upward while most firms’ internal language still positions it as what gets discussed after the technical review concludes. That gap is not semantic, because language of that kind determines what gets budgeted and what gets cut.

None of these steps requires a firm to hold a view on whether AI will eliminate jobs. They require only the more modest concession that the tasks your people learned from are gone, and that nothing has yet been built to replace them.

Apoorv Dwivedi is Founder and President of Fixyr and a speaker on marketing and growth for accounting and advisory firms.

FAQ