If AI Does the Job I Learned On, How Does the Next Person Learn It?

My first year in finance, a huge chunk of my job was building the same variance report every month, by hand, until I could do it without thinking. That sounds like a complaint. It wasn’t, in retrospect—it was how I learned what a “normal” number looks like, so that two years later I could glance at a report and immediately feel that something was off, before I could have told you why.

That kind of pattern recognition doesn’t come from being told the theory. It comes from repetition, from doing the boring version of the task enough times that your brain starts flagging anomalies automatically. It’s the same reason experienced pilots can sense something’s wrong before the instruments confirm it—thousands of hours of the routine version built the intuition.

Agentic AI is very good at exactly the repetitive part of the job that used to build that intuition. That’s the part of this conversation I don’t see discussed enough, because it’s not about jobs disappearing this quarter—it’s about a slower, harder-to-notice problem: the on-ramp for building senior judgment is getting narrower right as the industry needs more people with that judgment, not fewer.

This isn’t an argument against the tools

I want to be clear I’m not making a “kids these days” argument, or suggesting we should keep manual busywork around as a character-building exercise. Nobody should mourn spending less time formatting the same pivot table for the fortieth time. The tools are a real improvement in how much time gets spent on work that never taught anyone anything.

The problem isn’t that the mechanical work is going away. It’s that we haven’t yet built a deliberate replacement for what that mechanical work was quietly teaching. If we don’t design one on purpose, it doesn’t happen by accident.

What I think actually needs to change

Rotate people through the “why” question earlier and more often, on purpose. If an agent produces the first-pass variance commentary, the junior person’s job shouldn’t be to proofread it—it should be to argue with it. Does this actually explain what happened, or does it just describe the number? Make that a named, evaluated part of the role, not an assumed byproduct of “doing the work.”

Shrink the feedback loop on judgment, not just on output. The old model taught judgment slowly, over years, through volume. If volume is gone, the loop has to get tighter—more frequent, more explicit conversations about why a senior person made a specific call, not just what the call was.

Stop measuring junior performance by output speed. If the agent produces the draft in ninety seconds, “how fast did you turn this around” stops being a meaningful signal about the person. The signal that matters now is closer to “did you catch the thing the agent missed,” and that’s a much harder thing to build a performance review around—but it’s the actual job now, so it’s worth the extra work.

The honest part

I don’t think anyone, including me, has fully solved this yet. I’ve started deliberately giving newer team members the “poke holes in the agent’s draft” assignment even when it would genuinely be faster for me to just fix it myself. It’s slower in the short term. It’s the only thing I’ve found that seems to actually build the muscle instead of assuming it’ll show up on its own.

If you’re leading a finance team right now, I’d treat this as a real design problem, not a side effect to worry about later. The teams that figure out how to build judgment deliberately are going to have a real talent advantage over the ones that just let the on-ramp quietly disappear.