Most firms assume the apprenticeship is still intact and the agent just made it faster. They have it backwards. The first draft was the apprenticeship, and once software owns the first draft, the classroom disappears unless somebody rebuilds it on purpose. The memo nobody writes anymore is the memo nobody learns from.
What follows is a short teardown of the assumptions keeping firms from noticing the problem, and what actually works once you stop pretending the on-ramp is still there.
Myth: The Agent Just Took the Boring Part
The first-draft work was rarely only boring. It was the deliverable and the training program running on the same clock. A junior associate writing the intake summary learned the client. A first-year analyst building the model learned the business.
The partner's redline was the lecture, and the next draft was the exam.
Bloomberg Law's Carlson and Ma make the point sharply: a human kept in the loop to review AI output can learn to spot errors without ever building the judgment required to produce the work in the first place. Reviewing a draft and writing one are separate skills, and they always have been.
Myth: Juniors Will Learn by Watching the Agent
Watching a competent output appear is spectating, not learning. The old loop worked because the junior had to make the choices themselves, defend them, and then see which ones the reviewer kept and which ones they cut. The reasoning lived in the delta between draft one and draft two.
That delta is where training has to move. The agent's draft, the reviewer's edits, the comments explaining why a clause got rewritten or a function got refactored, the whole exchange is now the richest teaching material a firm produces. For the fuller argument about what happens to the on-ramp when the on-ramp is the part software is best at, see The Chicago Journal on agentic AI and entry-level work.
Myth: The Hiring Pipeline Will Sort Itself Out
The usual assumption is that firms will keep hiring the same number of juniors and give them harder work sooner. The early signals point the other way. Harvard Business Publishing's warning on this is worth reading in full: organizations are already thinning entry-level roles instead of redesigning them, and early-career employment in AI-exposed fields has fallen measurably as a result.
Cut the bottom of the pyramid for two or three years and the middle thins out behind it. There is no shortcut to a senior.
Myth: You Can Bolt Training on After the Fact
A workshop series does not replace a thousand small corrections on real work. If training sits next to the delivery pipeline as a separate program, it will lose every scheduling fight with billable hours. The firms rebuilding the apprenticeship are putting it inside the workflow, not beside it.
- Review the agent, not the output. Make the junior responsible for the critique: what did the draft miss, where did it over-reach, which source did it misread? The review becomes the assignment.
- Keep the transcript. Store the prompt, the draft, the edits, and the reviewer's notes together. That record is the new training manual, and it only exists if the agent runs inside systems the firm controls.
- Hand juniors the harder half. Push them toward the parts the agent does worst: client judgment, edge cases, system design, the awkward conversation. Those are the skills that compound.
- Score the second draft. Measure how much a junior improved the agent's output, not how fast they shipped it. Speed is the agent's job now. Judgment is theirs.
The Firms Keeping Their Bench Are Already Doing This
The pattern is consistent across firms handling this well. They treat the agent as the first-year associate that never gets tired, and they treat the human junior as the person who learns by reviewing it, correcting it, and gradually taking on the work the agent cannot do. The review transcript is what makes that progression legible to everyone involved.
None of this works if the agent is a black box bought off a shelf. It works when the agent is wired into the firm's own repositories, templates, matter history, and reviewer tooling, so every exchange leaves a trace someone can learn from.
That is a build problem before it is a training problem. The firms that get it right will have a bench in five years. The ones that don't will be hiring laterals and wondering where the mid-levels went.