The AI Economy’s Hidden Bottleneck

The debate over artificial intelligence and jobs may seem like a thoroughly juiced orange. Most of such arguments begin and end with back-of-the-envelope automation estimates: how many tasks AI can perform, how many workers these models replace. (Less considered, unfortunately, is how many new tasks and whole jobs might be created.) Those things matter, of course. And we’ll see how things play out over the coming months and years. Still, a narrow focus on the aforementioned factors misses what may be the more immediate challenge. The next labor shock may come not from breakthroughs in AI, but from how quickly companies restructure around the considerable capabilities already in hand. When thinking about the economic impact of technological progress, history typically offers a useful starting point. General-purpose technologies—from steam to electricity to computing—don’t transform economies on arrival. Their impact comes later—typically much later—when businesses redesign production around them. Adoption can lag invention by decades. A classic example: Electric power was available for years before firms worked out what to do with it. Henry Ford’s moving assembly line at Highland Park in 1913-14 was really a story about factory layout, not about electricity—yet it transformed output. It also hammered workers. Annual turnover hit nearly 400 percent. Likewise, important new digital tech isn’t just plug and play. Firms must invest in intangible capital—the hard-to-measure assets that make the technology actually useful, such as retraining workers, redesigning workflows, building new software systems, and developing managerial know-how. In the early years, productivity can look weak because a) resources are diverted into this hidden investment and b) firms temporarily sacrifice output as they move away from processes they already know how to run efficiently. Economists describe this as a “productivity J-curve“: a dip followed by a surge as the benefits of intangible investments are finally realized. It helps explain today’s puzzle of powerful AI tools coexisting with modest productivity growth. The technology may look like it’s underperforming the hype, but really it’s being slowly absorbed—though hopefully at a more rapid pace than electrification. There’s a second, less-appreciated problem: What if the GPT absorption happens too quickly? That’s the concern raised by economist Eduardo Levy Yeyati, whose new analysis (from which I took that example about Ford) focuses on the speed of adoption. Workers displaced from their existing roles—whether through outright job loss or the erosion of discrete tasks—don’t instantly step into new ones elsewhere. They enter a retraining pipeline with limited capacity. If firms adopt AI gradually, the system can cope. If adoption accelerates, the pipeline clogs. Workers facing long waits and mounting uncertainty may simply exit the labor force. Two economies can arrive at the same technological frontier and end up with very different social outcomes, Yeyati concludes. In short, the J-curve describes why productivity gains are delayed. The adoption-speed analysis explains why the transition can be really bumpy. In both models, the constraint is reorganization rather than access to innovation. None of this is an argument against AI. But it does raise the issue of the role that smart public policy can play. Expanding retraining capacity and improving labor-market mobility aren’t just social policies. They’re also important growth policies, ones most valuable when built before the wave of displacement hits, not after.