We Automated the Bottom Rung of the Ladder — And We're Still Asking People to Start at the Top

Employment for workers aged 22-to-25 in AI-exposed roles has dropped 13% since late 2022. Older workers? Steady. Same companies, same headcount targets — but the entry points vanished while we were optimizing productivity metrics.

I've been watching this pattern play out with clients for eighteen months now, and the math doesn't add up. We deployed AI to handle the grunt work. We celebrated the efficiency gains. Then we opened req after req looking for "judgment" in candidates who've never had the chance to build it.

We didn't just automate tasks. We automated the training ground where professionals learn to think.

The Apprenticeship We Stopped Calling One

Here's what judgment looked like in my world before AI: a junior analyst would pull apart 200 workpapers over two years. The first 50 were disasters — wrong risk classifications, missed control gaps, beautiful documentation of completely irrelevant details. By workpaper 150, something clicked. They could smell a problem three pages before the numbers went sideways.

That wasn't innate talent. It was repetition with feedback. Do the boring thing badly. Get corrected by someone who's done it 2,000 times. Repeat until pattern recognition becomes instinct.

One person I've been advising described how she learned to write funding profiles early in her career: "I produced 50 of them badly, until I knew what good looked like." That apprenticeship took months. Today, AI produces those 50 profiles in two minutes. The output survived. The training ground didn't.

You can't speed-run judgment formation. Medicine figured this out centuries ago — that's why residents spend years doing supervised grunt work before anyone lets them operate unsupervised. Law still makes associates spend thousands of billable hours on document review before they take a deposition. The trades never abandoned apprenticeships because everyone involved knew that watching a master electrician for two years teaches things a textbook can't.

Corporate work quietly dismantled this scaffolding over the past two decades. We called it "flattening hierarchies" and "empowering talent." What we actually did was reduce the tolerance for junior mistakes while simultaneously eliminating the low-stakes environments where those mistakes built expertise.

Now AI finished the job, and we're surprised the pipeline broke.

The Entry-Level Paradox Became Real

Remember when "entry-level position, 5 years experience required" was the running joke about clueless HR departments? We're not laughing anymore. According to Stanford's Digital Economy Lab, employers now expect day-one hires to evaluate and improve AI's work — to supply judgment we never gave them a way to build.

The operating model now assumes someone else trained them. But if every company outsourced that training to "somewhere else," where exactly is somewhere else?

I'm seeing this tension play out in real time. A financial services client spent three months searching for an audit associate who could "think critically about AI-generated risk assessments." Their talent team kept asking: where are these people supposed to come from? The firms that used to train them cut their junior programs by 40% because AI handles first-pass reviews now.

"We'll just hire judgment" isn't a strategy. It's a bet that someone else is still running the training program you shut down.

This reminds me of what happened when proprietary trading desks automated market-making in the 2000s. The junior trader role — the person who spent two years watching order flow and learning to read market microstructure — effectively disappeared. By 2012, firms were desperate for mid-level traders with intuition about unusual price action. But they'd eliminated the apprenticeship that built that intuition. The talent pipeline took a decade to reconstruct through quant programs and rotational assignments, because you can't surge-manufacture pattern recognition.

What Actually Builds Judgment (And What Doesn't)

I've been thinking hard about solutions, because complaining about AI eating entry-level work doesn't help the 24-year-old graduating this spring. Here's what I know from watching judgment develop — and fail to develop — over twenty years:

Judgment requires decision-making under uncertainty, with feedback. Reading about decisions doesn't count. Watching someone else decide doesn't count. You have to make the call, see the consequences, and have someone with more scars explain what you missed.

AI gives us the output without the struggle. A junior accountant used to spend a week building a financial model, make three structural errors, and learn from the senior who caught them. Now AI builds the model in minutes. The junior's job became "review the AI's work" — which requires the judgment they were supposed to build by making the mistakes the AI no longer lets them make.

That's the trap. We're asking people to evaluate work they've never done themselves.

The fields that still produce reliable judgment never left this to chance:

• Medical residents do supervised procedures for years before independent practice

• Apprentice electricians wire circuits under master supervision for 8,000 hours

• Law associates review thousands of documents, write dozens of terrible briefs, and get them shredded by partners before they run their first case

The structure isn't optional. It's the product.

The Uncomfortable Middle Ground

Here's where I lose both sides of this argument. The AI optimists say we'll find new ways to train people — maybe AI can be the teacher. The skeptics say we should slow AI adoption to protect training grounds. I think they're both half-wrong.

We're not going back to manual grunt work for nostalgia's sake. The efficiency gains are real, the competitive pressure is real, and no company is going to unilaterally disarm by rejecting AI while competitors scale with it. That ship sailed.

But we also can't pretend the invisible curriculum will figure itself out. The market solved this problem before by accident — grunt work created judgment as a byproduct. That accident stopped happening. Now we need intention.

What does that look like? I don't have a tidy answer, but I'm seeing experiments worth watching:

• Structured feedback loops: One accounting firm now requires AI-generated work to go through mandatory peer review sessions where juniors present the output and defend the methodology. They're not doing the initial work, but they're doing the thinking.

• Scenario training: A legal team rotates juniors through "decision theaters" where they handle compressed versions of cases partners actually managed, with AI available as a tool but judgment as the evaluation criteria. Residency for corporate work.

• Supervised automation: Instead of "AI does it, junior reviews it," some teams are trying "junior does it with AI, senior reviews both." More expensive in the short term, but it keeps humans in the learning loop.

Are these perfect? No. Do they scale easily? Not yet. But they're better than crossing our fingers and hoping judgment materializes spontaneously in 26-year-olds who've never had to build a financial model from scratch or write a brief without autocomplete.

Where Does Your Next Generation Come From?

Nobody gets fired the day AI arrives. The juniors just slowly stop developing, and five years later you can't find a qualified manager anywhere.

I've survived enough technology disruption cycles to recognize this pattern. When electronic trading hit the NYSE floor, the immediate story was efficiency and cost savings. The delayed story, five years later, was "where did all the people who understand market structure during stress events go?" The automation eliminated the apprenticeship, and it took a decade of intentional reconstruction to rebuild the knowledge base.

We're about to run that movie again, but faster and across more industries simultaneously.

The companies that figure out judgment formation in an AI-assisted world will have a structural advantage that won't show up in quarterly metrics for years. The ones that don't will spend the 2030s competing for a shrinking pool of experienced talent that nobody trained because everybody assumed someone else would handle it.

So here's what to ask your leadership team Monday morning: If AI does the work that used to train your junior staff, where does your next generation of judgment come from? Not in theory — specifically. What decisions are they making? What mistakes are they allowed to make? Who's reviewing their thinking, not just their output?

"We'll hire experienced people" only works if someone, somewhere is still running the training program. And right now, looking at that Stanford data on 22-to-25-year-old employment, I'm not convinced anyone is.

We paved over the field where judgment grew. Either we plant a new one on purpose, or we find out what happens when an entire generation enters management without the pattern recognition that used to come standard.

But what do I know — I've only watched this particular disruption cycle play out in three different industries over twenty years. Maybe this time really is different.

What's your organization doing to build judgment in an AI-assisted world? I'm collecting case studies of what's actually working — not theory, but real programs with outcomes. If you're experimenting with structured learning in the age of automation, I want to hear about it.