Messy Jobs: The Work That AI Cannot Reach, by Luis Garicano, Jin Lee, and Yanhui Wu (Upriver Press, 222 pp., $22)
As the political salience of artificial intelligence keeps rising, a new book brings much-needed evidence and rigor to an increasingly polarized debate.
Finally, a reason to check your email.
Sign up for our free newsletter today.
Messy Jobs: The Work That AI Cannot Reach, by economists Luis Garicano, Jin Li, and Yanhui Wu, deserves acclaim for refusing both boomerism and doomerism, while still landing, correctly, closer to the optimists’ side of the debate. The book’s contribution is less a forecast than a framework, drawn from organizational economics, for thinking about what jobs actually are.
Jobs, on the authors’ account, are best understood as bundles of tasks rather than indivisible units. Some tasks are routine and separable; others are entangled with accountability, trust, judgment, and the small “p” political work of getting things done inside organizations.
The framework’s power lies in its central variable: the cost of separating tasks from one another. When separation is cheap, AI takes the routine component, and the human role narrows. (Think of the travel agent whose booking function is detached easily from the rest of the job.) When separation is expensive, the human keeps selling the full service. The accountant, for instance, survives not because AI can’t reconcile ledgers but because he or she interprets tax law for a client, signs the audit, and carries the legal exposure.
The “mess,” as Garicano et al. put it—competing interests, unpredictable demands, decisions someone must own—is what machines can’t reach. For workers worried about AI-induced displacement, the framework is reassuring. The messier the job, the safer it is.
The authors’ work rests on what they call “the autonomy threshold,” the point at which machine performance meets the standard a competent professional currently supplies. Below it, AI complements whoever operates it. Above it, the question changes from whether a job survives to which humans still add something the machine doesn’t.
All of this concerns the supply side: what AI can take over and what it can’t. The book is equally convincing about the demand side. As legal services, software, design, and analysis get cheaper, society will consume vastly more of them. In the authors’ telling, the right historical analogy isn’t the defunct elevator operator but the “Jevons effect,” in which efficiency expands markets rather than exhausting them.
Add the democratizing effects—AI hands entrepreneurs the capabilities of consultancies and erodes the gatekeeping that made expertise artificially scarce—and the optimistic case gets even stronger. On the aggregate economics, the book is persuasive, and the endorsements from David Autor, Tyler Cowen, and Patrick Collison are earned.
But the authors leave unasked the question of how gains are distributed within the professional class itself. The book’s unit of analysis is the job, and jobs are analyzed by their strongest version: that accountant who signs the audit, the manager who holds the coalition together, the partner who owns the client relationship. But every occupation contains enormous variation beneath the top performers. A framework that asks whether lawyers generally survive AI risks missing the more consequential question: What share of lawyers survive?
There’s a structural reason to suspect that the share will be smaller than professionals assume. The knowledge economy conflated two scarcities—cognition and judgment—because for 50 years, they traveled together. Organizations needed armies of educated workers, and the resulting expansion of professional employment suggested that economically valuable judgment was widely distributed. Yet judgment was mostly inferred and never quite tested. Credential inflation was the market’s admission that a degree no longer guaranteed judgment.
Field experiments show that AI compresses performance, lifting the weakest workers most. Optimists read compression as leveling, and below the autonomy threshold, it indeed is. Above it, workers who merely match the machine become replaceable by it, and the wage premium shifts to those who exceed it.
The plausible result looks less like the hollowing-out that the doomers predict than superstar dynamics arriving in ordinary professions. Good judgment used to come with a hard limit: the person who had it could only oversee so much work, so applying it at scale meant hiring dozens of people to execute. Now AI does the executing, which means an editor with genuine taste can direct output that once required 50 subordinates. That’s a world of extraordinary productivity and wider dispersion at once.
The authors have a serious rejoinder. Many roles, according to their framework, exist for unpredictable demand: the job looks routine most of the time and exists for the exceptions, which can’t be separated in advance. The rejoinder is right about the function but perhaps incomplete about the headcount. AI that manages the routine and flags the hard cases for a human is precisely a technology for separating the routine from the complex. The compliance operation that needed 500 may need 30.
It must be emphasized that none of this overturns the book’s optimism. But it may push it in a different direction. Judgment in the relevant sense means deciding what matters, owning outcomes, and knowing when the machine is wrong. If those capacities are only loosely correlated with credentials, then AI should over time rebalance both labor demand and the social value attributed to different professions and skills.
That’s why the book’s analysis is most compelling exactly where the knowledge economy was most dismissive. Nurses, for example, along with electricians and other skilled trades, perform work that mixes thinking with presence, accountability, and trust. No machine can pull that bundle apart. For two generations, economies steered ambitious young people away from that work because cognition was scarce and paid accordingly. If AI changes the returns, talent will flow back the way it always has.
These are testable propositions, and the coming years will adjudicate them. The book’s task-based account predicts occupation-level effects, with incumbents broadly along for the ride. The compositional account predicts sorting within occupations—the same job title shedding its lower tier while its upper tier gains leverage. The displacement data will look quite different under the two stories.
Messy Jobs is the best book yet written on AI and work. Its core insight—that judgment will survive cognitive abundance—is a major contribution to the policy debate. A friendly amendment is that survival isn’t distribution.
For half a century, advanced economies organized status around expensive cognition. Cheap cognition puts that hierarchy up for revaluation. The transition will be hard for many who built lives on the old signals, but an economy that prices judgment, taste, trust, and skill on their merits rather than their packaging is both healthier and more egalitarian than the one it replaces. That revaluation is the question AI is about to ask on everyone’s behalf.