PwC's internal numbers, reported by Business Insider in August 2025 and confirmed by the firm: junior associate hiring falls from 3,242 in the fiscal year ending June 2025 to a planned 2,197 by fiscal 2028. A 32% reduction across three years. The reasons given were technological change and unusually low attrition.
The Ladder Loses Its Bottom Rung
The more revealing line came from PwC's AI assurance leader, Jennifer Kosar, describing what the remaining hires will do. People will walk in the door and almost instantaneously become reviewers and supervisors.
That is a coherent plan for staffing. It is not yet a plan for training.
The work that AI removes first is the work that used to be handed to the newest person in the building. Tie out the schedule. Pull the support. Draft the first pass. Chase the missing document. It was tedious, it was cheap to supervise, and it was also the entire curriculum.
Most small and mid sized firms will never hire at Big Four scale, but they will inherit the shape of the change, because they buy the same software. When the workpaper prep and the first draft come out of a tool, the same rung disappears from a twelve person firm as from a firm of forty thousand.
The Pipeline Is Not Actually the Constraint
It is tempting to fold this into the talent shortage story, and the numbers do not quite support that.
Spring 2025 enrolment in two and four year accounting programmes reached 266,506 students, a 12.4% increase and the highest since 2020, according to figures reported by the Journal of Accountancy. Degrees awarded went the other way, 55,152 bachelor's and master's in 2023 to 2024, down 6.6%, with master's down about 15%. New CPA Exam candidates fell from 42,626 in 2023 to 28,082 in 2024, with 16,448 in the first six months of 2025. And 75% of participating public accounting firms expected to hire at least as many new graduates in 2025 as they had in 2024.
That is a mixed picture, not an empty one. There are people arriving. The question is what happens to them in year one and year two, and that question is about the firm, not the pipeline.
Review Is a Skill and It Was Learned by Repetition
Ask a partner how they learned to spot the thing that is wrong in a file and the answer is almost never a training course. It is some version of: I did four hundred of these and my manager marked up the first fifty.
That mechanism has two parts, and firms tend to remember only the first. The repetition mattered. So did the mark up, the moment where somebody senior showed their reasoning on a specific file: this is why that treatment is wrong, this is what I looked at first, this is the question you should have asked the client.
Remove the four hundred and the mechanism does not survive on the mark ups alone, because there is nothing left to mark up. A junior reviewing a machine's output is being asked to have the judgment before doing the thing that produced it.
The cost of getting this wrong is not felt for two or three years, which is exactly why it does not appear in anyone's business case. It shows up later as a group of five year qualifieds who are fast, competent and oddly unwilling to overrule the software.
What Actually Taught Judgment
Here is the part worth arguing about. The repetition was never the teacher. It was the delivery mechanism.
What a junior absorbed over four hundred files was a library of worked examples: what the firm did last time, why, what the reviewer caught, what the client came back with. Volume was simply how that library got transmitted, one file at a time, at enormous cost.
Most firms never built the library as an actual thing. The reasoning lived in a review note somebody deleted, a comment in the margin of a workpaper, a conversation in February. Volume worked as a substitute for documentation for a long time, and firms could afford not to notice.
Automation removes the substitute. A firm that can show a new hire what it did on a comparable client three years ago, and why, can train a reviewer without four hundred files. A firm that cannot is relying on a transmission mechanism it has just switched off.
The Shape of the Fix
Treat the client file as the teaching material, and make it findable in the time a curious person is willing to spend looking.
MetaWurks is built for that half of it. It ingests a client's returns, statements, contracts, workpapers and correspondence and lets anyone in the firm query the whole history in plain English, so a second year can ask what the firm did on a comparable matter and read the actual documents rather than interrupting a partner. Role based access controls govern who can open which client's records, audit logs record who opened what and when, and documents ingested into the platform are not used to train models or exposed to other users.
It does not write down why a position was taken. People still have to do that. What it changes is whether the writing down is ever found again, which decides whether it was worth doing.
The staffing plan is easy to model. The training plan it quietly assumes has not been written.
Join the Conversation
If a second year in your firm wanted to see how you handled a similar client three years ago, could they find it without asking a partner?