How do trainees learn the craft if AI drafts the first version?

What was the first draft actually teaching?

Ask a partner why trainees should write the first version and the answer usually arrives as a feeling rather than an argument. You learn by doing it badly. You learn by being sent back. That feeling is right, but it is worth taking apart, because only some of what the blank page teaches is worth protecting.

Four things happen when a junior drafts from nothing. They have to decide what the document is for and who reads it. They have to go looking for a precedent, which means judging whether an old matter resembles the new one, a skill nobody teaches directly. They discover the limits of their own knowledge, usually at the point where the drafting stops and they have to go and ask. And they get the marked-up version back, which is where the craft is actually transmitted.

Notice that none of those four is typing. A good deal of what juniors did instead of learning was formatting, renumbering, chasing figures between spreadsheets and rebuilding tables that broke. Nobody became a better adviser by reconciling a schedule for the third time.

Which part of that is lost when the machine goes first?

The honest answer is the second and third items: the hunt and the discomfort. If a complete, plausible draft appears, the trainee never has to decide which precedent applies, and they never hit the wall that tells them what they do not know. The draft hides the gaps rather than exposing them.

What survives, in principle, is the first item and the fourth. The question of what the document is for can still be asked. The marked-up version can still be returned. But both now depend on someone choosing to make them happen, where before they happened by default. That is the whole problem in one sentence. The old apprenticeship worked partly by accident, and accidents do not survive a change in the process.

This is the training version of the argument we make elsewhere about a tool not being a process. Drop a drafting assistant into a firm and the drafting gets faster. Everything that was attached to the slow version, including the supervision, quietly detaches.

Does reviewing a good draft teach as much as writing a bad one?

Not on its own, and the reason matters. Reading something competent produces a strong sense of understanding. You follow the structure, you agree with the reasoning, you spot a typo, you feel informed. Then you are asked to produce the same thing unaided and nothing comes. Recognising a good argument and being able to generate one are different abilities, and only the second is what clients pay for.

There is a second risk, which is that the junior corrects the surface and misses the substance. We have written about the risk in copy and edit in the context of reused documents, and the dynamic is the same here. A draft that reads fluently invites light-touch editing. The wrong test applied to the right facts survives review far more easily than a blank page does.

How do you design supervision that still builds judgement?

The fix is not to ban the draft. It is to put something in front of it. A handful of practices do most of the work:

• Commit before you read. Before the trainee opens the draft, they write five lines: what this document has to achieve, the two or three points that will decide it, and where they expect the difficulty. Then they compare their note with the draft and explain every difference.
• Mark up with reasons, not just changes. A tracked change teaches nothing. A tracked change with six words of explanation teaches the rule behind it. The reviewer is doing the work anyway, and it is the only part of supervision that transfers.
• Ask where it is wrong, not whether it is right. Juniors asked to approve a draft tend to approve it. Juniors asked to find the two weakest points will find them. The second question produces critical reading; the first produces a signature.
• Keep some work deliberately slow. Choose a category of matter where the trainee drafts unaided, and say openly why. Scarcity makes it feel like development rather than punishment.
• Teach the sources, not just the output. If the firm has built proper knowledge retrieval, make juniors use it directly. Reading two near-miss precedents and arguing about which applies is the hunt, preserved on purpose.

Running underneath all of this is the rule that a named person reviews and approves anything that leaves the firm. Our guide to reviewing AI output on client work sets out what that review should cover. Training is a reason to keep the review human that has nothing to do with risk.

What does the firm have to give up?

Time, and it has to be given up on purpose. When a process that took four hours takes forty minutes, the recovered time has a habit of disappearing into more matters. If supervision is the thing that no longer fits, the firm has traded a training system for a throughput gain and will notice in about three years, when the next cohort of senior associates is thinner than it should be.

So decide the split in advance. Some of the recovered hours go to fee-earning work. Some go to the review conversation that used to happen free of charge inside the drafting cycle. That decision belongs to the same conversation as utilisation targets after the admin goes, and it is better made deliberately than by drift.

There is a related gain worth naming. Firms that write down why a document takes the shape it does are also firms that stop losing knowledge when people leave, which is the problem behind when knowledge leaves. Supervision notes and good templates are the same asset viewed from two angles.

What should a training principal do this quarter?

Three things. Read the last ten pieces of work a trainee produced and ask what decision each one required of them. If the answer is mostly presentation, the development has already stalled, regardless of any software. Second, give juniors a piece of the job they were previously kept away from: running a client call, taking a first view on a scope, handling the queries on a set of file notes. The capacity exists now. Third, write down what the firm's documents are supposed to contain, because you cannot supervise against a standard nobody has stated.

The shape of this differs by sector. In law firms the first draft was attendance notes and straightforward correspondence. In accountancy firms it was working papers and the queries list. The underlying bargain is identical: the firm got cheap output and the junior got exposure. Remove the cheap output and you have to buy the exposure another way.

Our wider view of how these roles are changing is in what changes for junior staff. If you want to know which process is consuming the hours that should be going into supervision, the audit will tell you.