The tasks that generative systems handle most reliably share a profile: high volume, well-defined output, tolerant of review, and previously assigned to whoever had joined most recently. First-pass document review, summarising a call, drafting a standard memo, reconciling two lists, building the initial version of a model — this is the work that gets automated first because it is the work that is easiest to specify.
It is also, and not coincidentally, how people learned their jobs. Nobody assigned first-year associates to document review because it was valuable in itself. It was valuable because after four hundred documents a person develops an instinct for which one matters, and that instinct is the foundation everything senior is built on.
The pipeline breaks quietly and reports well
Firms removing this work see immediate, legible gains. The output arrives faster, costs less and is frequently better than a tired junior produced at two in the morning. Nothing in the current quarter argues against continuing, and the back-office agents already doing this work are among the least controversial deployments in the enterprise.
The cost appears on a delay long enough to escape attribution. A firm that stops giving juniors the repetitive work does not notice for two or three years, and what it notices then is not a hiring problem — headcount is fine — but a judgement problem. The people who should be running matters have never done the volume that produces pattern recognition, and no amount of seniority substitutes for it, because it was never seniority doing the work.
This is the same structural mistake as the one companies are now paying to undo a layer up. The middle management tier that was cut and is being quietly rebuilt was the mechanism by which people practised running a function at reduced scale before running it fully, and its removal produced a capability gap years after it was booked as a saving. Automating the entry level is that decision made again, further down, with better tooling and the same blind spot.
A few firms are treating it as a design problem rather than an efficiency one. The approaches vary — deliberately routing a share of automatable work to juniors as training rather than production, having them review and correct machine output at volume, or building explicit exercises that reproduce the reps the job no longer supplies. All of them cost something, and all of them require admitting that a portion of headcount is now overhead rather than delivery, which is a harder line to defend than it used to be.
The firms most exposed are the ones that hire on demonstrated capability rather than credential. Skills-based hiring passed its durability test partly because the entry-level rung produced observable evidence of what someone could do; if that rung stops generating evidence, the assessment problem returns in a form no assessment tool solves.
None of this argues for keeping juniors on document review out of sentiment. It argues that the reps were doing something, that removing them removes it, and that a decision this consequential is currently being made by default — one automated workflow at a time, by people optimising a different number.
