List tasks that look basic but teach useful instincts. Example: checking invoices may teach a junior accountant how suppliers behave, where mistakes appear, and why certain approvals exist. If AI now performs the check, decide how the beginner still sees patterns, exceptions and consequences rather than only the polished output.

Not every old task deserves saving. Some admin is just admin. The goal is to keep practice that builds skill, not preserve suffering for tradition. Ask which tasks help beginners learn judgement, accuracy, communication, customer context or system behaviour. Keep those as training reps, even if AI helps later.

Let juniors compare their attempt with AI’s attempt. Example: ask a trainee to draft a short analysis, then compare it with AI output and discuss differences. This creates learning instead of silent outsourcing. If the trainee only edits AI work, they may learn polish without ever learning structure.

A senior person saying “I’ll review everything” is not a system. Define what gets reviewed, how often, what feedback looks like and what good means. Example: one weekly session comparing human drafts, AI drafts and final decisions may teach more than twenty rushed comments saying “looks good.”

For each role, state what a beginner should be able to do alone after 30, 60 and 90 days. AI can assist, but the human still needs milestones. Example: by 60 days, they should spot three common data issues without prompting, otherwise the tool is progressing faster than the person.