Professional firms rarely had to approve a budget for turning beginners into experts, because clients covered it. Clients paid for first-pass work: matching a company's cash records to its bank statements, reviewing documents, drafting a rough analysis. A first-year employee did that work, ideally sitting near someone senior who could explain why an odd entry mattered. Over time the newcomer took on larger pieces of the same real work and became a practitioner by doing it.
The arrangement was never as good as nostalgia suggests. Years of rote checking without feedback produce tired employees more often than skilled ones. Its durable virtue was that learning travelled inside work someone was already paying for, so no manager ever had to defend it as a separate expense.
AI happens to be very good at exactly those beginner tasks. A firm can automate them and record the saving this year. The missing experience shows up a decade later, when senior people retire and the layer beneath them has never worked a messy reconciliation from scratch.
The entry point already looks narrower. Stanford's updated "Canaries in the Coal Mine" paper finds that employment of 22-to-25-year-olds in the jobs most exposed to AI has fallen 19 percent behind the path of their peers in less-exposed jobs, mostly because firms are hiring fewer of them. The authors call these early descriptive indicators, not proof that AI is the cause. A separate Census working paper estimates that interest-rate shocks could explain up to a quarter of a similar decline, and New York Fed researchers point to remote work, which makes mentoring harder. Whatever the mix of causes, each one erodes contact between a newcomer, real work and someone able to teach.
Once learning stops riding along with billable work, training becomes a line item, and line items get cut when business slows. Each firm also has a sensible alternative, which is to hire experienced people later, after another employer has paid to train them. One firm making that choice is acting rationally. If most firms make it, the pool they all plan to hire from shrinks, and the firm that keeps training generously ends up partly training its rivals' future staff. It is a classic collective-action problem.
No single firm can fix that, which points toward the institutions all of them answer to. The professional bodies that certify accountants and auditors already decide what experience counts toward a qualification. They are well placed to check whether that experience was actually supervised, in lean years as well as good ones. When every firm faces the same requirement, skipping it no longer pays.
Deloitte has shown what such experience might look like now. In its redesigned audit training in England and Wales, trainees perform cash reconciliations themselves in mock audits, and they also review a reconciliation produced by AI. They get 22 more days of study leave, 135 in total. The pairing matters, because producing work with AI does not by itself teach the judgment needed to check it. Whether this produces lasting expertise is not yet known.
A check like that needs something concrete to look at, and so does any training budget meant to survive a slow year. Two measures would serve both. The first is the senior employee's hours, since a firm can promise mentoring while setting targets that leave no time for it. The second is what trainees can do: explain what is wrong with the machine's reconciliation, handle an unfamiliar case, take on harder work.
Firms will describe their commitment to talent in announcements. Their study days, unbilled practice hours and graduate hiring numbers are a more reliable guide to what they actually do.

