The Training Ladder Problem: AI Is Breaking Junior Roles

Why Luis Garicano Sees AI Dismantling Professional Training

Luis Garicano is a professor at the London School of Economics who has spent decades studying knowledge hierarchies and organizational economics. In this EPOCH AI interview, he presents one of the most sobering analyses of AI's impact on professional work—not through mass layoffs, but through the quiet disappearance of entry-level hiring. His research reveals a hidden crisis: the "training ladder" that turns juniors into experts is being dismantled by the very tools meant to make them productive.

On the human bottleneck: "As long as the AI needs your supervision because it makes lots of mistakes, then the bottleneck is the human." This is the core tension Garicano identifies—AI productivity gains are capped by human oversight capacity until AI crosses the "supervision threshold" into genuine autonomy. We're in a period of augmentation, not replacement, which actually limits the upside.

On seniority-based technological change: "In the AI exposed occupations, you don't see anything happen to senior employment. You see it growing. You see junior employment really dropping. And the way it's dropping is through hiring." Garicano cites two August 2025 papers analyzing 62 million workers showing this isn't anecdotal—it's systematic. McKinsey partners aren't being replaced; they're just not bringing on new analysts.

On the broken training contract: "The apprentice is paying not in dollars, it's paying in menial tasks... If the AI can do the basic research at McKenzie, can do the contract review at Cravath, then how do you pay for your training?" Junior employees historically exchanged grunt work for learning from experts. When AI does that grunt work, the economic logic of the master-apprentice relationship collapses.

On the superstar effect: "A very good AI programmer with lots of AI can have enormous leverage and can reach very large market size." Just as Messi can reach 500 million viewers, top-tier AI engineers can now command $100M compensation packages. The bifurcation is real: complements at the top, substitutes at the bottom.

On regulatory fragmentation: "We are in a game theoretical situation between China and the US. I don't think the possibility of slowing things down exists... Who is 'we' here? Is it China, is the US? Is it firms? Is it workers?" Garicano pushes back on Daron Acemoglu's optimism about directing technology—there is no unified actor who can coordinate a slowdown.

6 Insights From Luis Garicano on AI and Workforce Economics

• Juniors are being quietly de-hired, not fired - The shift shows up in hiring freezes, not layoffs, making it harder to detect until the pipeline of future experts is already broken
• The supervision threshold determines everything - AI moves from complement to substitute the moment it crosses the quality bar where humans can't improve its output
• Tacit knowledge transfer is at risk - Expertise that isn't in any manual—how to handle clients, navigate organizations—may stop being transmitted entirely
• Superstar economics are amplifying - Those at the top gain leverage while those below the autonomy threshold face wage compression
• Short-term GDP could fall even as welfare rises - If legal and medical services become free, consumer surplus soars but measured economic output may contract during the transition
• GDPR and EU AI Act may backfire - "Part of the risk is you try to control the technology and you end up without technology"

What This Means for Professional Services and Hiring

Garicano's analysis reveals a paradox at the heart of AI augmentation: the tools meant to help junior workers learn faster may actually eliminate the economic rationale for training them at all. When AI does the contract review, the PowerPoint slides, and the basic research, what currency does the apprentice have left to pay for mentorship? This isn't about AI replacing experts—it's about AI preventing the next generation from becoming experts. For organizations deploying AI agents, the question isn't just "what tasks can we automate?" but "what training pathways are we destroying?"