How to train junior staff when AI does the grunt work

The work that used to teach people

There is a version of this story that is only about efficiency. AI takes the background research, the first draft, the initial pass at the analysis. The work gets done faster and with fewer hands. That version is true as far as it goes.

What it leaves out is that those tasks were never only tasks. They were the training ground. The new hire who spent three days pulling together background on a funder was not only producing background on a funder. They were learning what mattered in it, what a partner actually cares about, and how long that kind of work really takes.

Scott Hutcheson made this point in Inc., writing about KPMG's decision to put nearly a thousand interns through its Lakehouse training centre in Orlando for hands-on work on critical thinking, judgement, communication and interpersonal skill. The size of that investment is the interesting part. A firm does not build something like that unless it has concluded something has gone missing.

Responding to KPMG's Investment in Interns Reveals an AI Problem Companies Can't Ignore — Scott Hutcheson, Inc.

What the repetitions were actually for

Anybody who has brought on somebody junior knows the shape of the first few months. There is orientation, there is training, and then there is a stretch where the work keeps coming back not quite right.

The first draft is bad. The research took two days longer than it should have. The analysis arrives at a conclusion that needs correcting. None of that is failure. That is the mechanism working exactly as intended.

Nobody learns to spot a weak assumption by being told to watch out for weak assumptions. They learn it by making one, having somebody point at it, and feeling the cost of having missed it. The correction sticks because the mistake belonged to them.

That is how most people in this sector learned. By running a programme badly the first time. By sitting through a partner meeting underprepared and knowing it in the room. By writing a grant application that did not get funded and then finding out why.

Take away the bad first draft and you have not only removed a task. You have removed the thing the coaching was attached to.

What KPMG's interns said about it themselves

KPMG surveyed 906 of its US summer interns in July 2026, and the answers are more interesting than the usual survey about attitudes to technology.

76% said future career success will require both human skill and knowing how to direct AI agents well. Both, not either. Their top concern about AI's effect on leadership development was over-reliance limiting critical thinking, named by 43%. Only 5% gave job displacement as their main reaction to AI agents.

Then the finding that should stop any manager reading it. Asked what the most valuable way they learned during the internship was, 57% said direct coaching from a person. AI-assisted learning finished last, at 3%.

They are also clear about what that implies for the people above them. 59% said managers and mentors become more important in an AI-driven workplace. Only 9% said less important.

Derek Thomas, who leads university talent acquisition for KPMG in the US, described them as "guarding those reps" — the hands-on work that builds the instincts. It is a striking thing for people entering the workforce to be saying about themselves. They can see the trade on offer and they are not comfortable with it.

Survey data from Summer Intern Pulse Survey 2026 — KPMG LLP

The question nobody has answered yet

So the question is not whether AI should be doing entry-level work. It already is. 64% of those same interns said more than a quarter of their current assignments are AI-assisted, and that number is not going down.

The question is what replaces the repetitions.

If a junior employee no longer spends three days on background research, what are they spending three days on that builds the same judgement? If the bad first draft never gets written, where does the correction happen? If nobody ever produces an analysis that needs fixing, what exactly is being coached?

Most organisations have not answered this. They have taken the efficiency, which is immediate and easy to see, and left the development gap open on the assumption that it will sort itself out.

Five replacements that fit inside an ordinary week

None of this needs a training centre in Orlando. The replacements are conversational, and most of them fit inside meetings that are already on the calendar.

1. Make people explain their thinking. Not the answer — the route to it. "Walk me through how you got here" is the cheapest diagnostic available to any manager. Somebody who cannot narrate the route did not take it, and you find that out in about ninety seconds.
2. Give the situation, not the task. Scenario practice builds judgement without needing a real crisis to practise on. What would you do if the funder pulled out in week two? Rehearsing decisions is most of what judgement is, and rehearsal costs nothing but the time in the room.
3. Critique in the room, not in the margins. A tracked-changes comment teaches one person quietly. The same correction talked through openly teaches everybody listening. That is how apprenticeship worked before it became a document workflow, and it is the part that got lost.
4. Have junior staff present and defend a call. Not a status update — an actual recommendation, with questions afterwards. Standing behind a decision under mild pressure from people who know more than you is a repetition that no tool can absorb.
5. Protect the coaching time. 57% of those interns named direct coaching as the most valuable thing in their internship. It is also the first thing deleted when the calendar tightens, because it never looks urgent on the day it gets cancelled.

Point AI at the reasoning, not the output

The answer is not to ban the tool. That is unrealistic, and it is also wrong. The same interns who worry about over-reliance expect to have to direct AI well, and they are right that both things are needed at once.

The useful move is to aim it at the thinking rather than at the deliverable.

• Ask it to explain the decision. Not what the answer is, but why, and what it weighed against what. Then check whether that reasoning actually holds up, which is a repetition in itself.
• Ask it to disagree with you. Hand over your conclusion and ask for the strongest case against it. Much harder to dismiss than a colleague being polite.
• Ask it to argue the other side of a decision already made. A pre-mortem that takes four minutes instead of a meeting, and it surfaces the objection nobody wanted to raise out loud.
• Ask it what it is least sure about. Wherever the confidence is thin is usually exactly where a person needs to go and look properly.

Used that way the tool creates repetitions instead of removing them. The person still has to make the call. They just get to the hard part faster, which is close to how those interns described using it themselves.

Nobody can actually predict this. It is possible that new kinds of repetitions appear on their own, and that in ten years all of this reads as a fuss about nothing.

But that is not a plan. The safer assumption is that judgement still gets built the way it always has been — by doing something imperfectly, being questioned about it by somebody more experienced, and adjusting. If AI has taken the first part, the second and third have to become deliberate. They used to arrive for free, attached to work that no longer exists.