KPMG rebuilds early careers training for a generation AI never lets do the grunt work

---

KPMG has put close to 1,000 interns, 906 of them counted in the firm’s own July survey, through an early careers training programme it rebuilt almost entirely this year. The repetitive, low-stakes work that used to turn a nervous 22-year-old into an auditor with judgement is now mostly done by AI, so the firm had to find another way to teach it.

The Wall Street Journal sent a reporter to watch. Interns at KPMG’s Lakehouse campus in Orlando were put through a scavenger hunt and a whodunnit-style fraud exercise, hunting for the kind of inconsistency that used to turn up only after months of checking real invoices against real ledgers. Fortune, visiting a separate cohort of 600 winter interns drawn from 9,000 applicants across 146 schools, watched a different version of the same idea: three days of scenario work built around a new mantra, think, prompt, check.

Both early careers training ideas are trying to solve the same problem.

What early careers training used to teach by accident

For decades, the professional services apprenticeship worked in a specific way. Graduates spent their first two or three years doing large volumes of repetitive work: testing samples, checking invoices, building the same model forty times for forty different clients. They did it badly at first, then competently. Judgement arrived as a byproduct, not as a lesson anyone taught on purpose. The repetitive work was the curriculum, not just the output.

KPMG’s own description of the early careers training shift is the phrase Fortune reported: the “middle to middle” work, the routine analysis and cross-checking that used to eat three-quarters of a junior professional’s day, six hours out of eight, is now handled by AI. Which leaves a firm needing to teach directly what its juniors used to absorb by accident, and no obvious early careers curriculum for doing it.

It’s worth setting this against the wider early careers hiring picture, where employers currently sit on two opposite bets about what AI means for entry-level roles. KPMG’s bet is narrower than it looks: the old way of training graduates has gone, and something has to replace it fast, because the firm still needs auditors capable of spotting fraud a model missed.

KPMG’s own July survey of 906 US summer interns backs the theory up with numbers. Fifty-seven per cent said direct coaching from a real person was the most valuable way they learned this summer, more than any other method on offer. AI-assisted learning came last, at just 3%. Forty-three per cent named over-reliance on AI limiting their own critical thinking as their top concern, well ahead of the 5% who worry about AI taking their job outright.

One of the interns Fortune spoke to, K-Linh Nguyen, put the harder problem more plainly than any KPMG slide deck could. Judging when an AI output is wrong takes an eye for it, she said, and you can’t teach that eye: “it’s one of those things where you have to experience it to appreciate it.” That’s the sentence that should worry anyone trying to design an early careers programme to replace lived mistakes with simulated ones.

Can judgement really be taught in a classroom? The honest answer is that nobody knows yet, and KPMG doesn’t pretend otherwise.

Judgement has always been built from making mistakes on real client work and getting corrected by someone senior enough to catch the mistake before it mattered. A scavenger hunt can teach pattern recognition. It cannot yet replicate the stomach-drop of a partner flagging a genuine error in a genuine client file, and that discomfort may be doing more of the teaching than anyone wants to admit.

KPMG’s field study with the University of Texas at Austin, tracking 523 early-career professionals, found that people with near-identical skills got wildly different results once AI entered the workflow. What mattered was how well they guided and checked it, not how much they already knew going in.

That’s a real data point, not just a press release line.

It fits the broader argument that the way young people enter and progress at work needs rethinking well beyond one firm’s intern cohort. Most of what’s publicly known about whether KPMG’s redesign actually works, though, still comes from the firm’s own surveys and the access it granted two friendly newsrooms, so the results are promising rather than proven.

For any HR leader watching their own graduate scheme, the transferable question is which parts of your own early careers pipeline relied on juniors doing dull work badly before they did it well, without anyone designing it that way on purpose.

That mechanism is disappearing from every function AI touches, not just audit.

Building the judgement AI can’t replicate is turning into a design problem every graduate scheme will have to solve deliberately, whether or not the answer ends up looking anything like Lakehouse.