One way of thinking about this is that all knowledge work varies along one important spectrum: messiness. On one end, there is one defined task to execute, say helping clients fill their taxes. You get the expenses and payslips on email, you use some rules to put them on a form, you obtain a response. Over time, you become better at this task, and get a higher salary. On the other end of the spectrum, there is a wide bundle of complex tasks. Running a factory, or a family, involves many different tasks that are very hard to specify in advance.
The result is that workers with simple tasks will become continuously more productive (and richer), until their work is worth nothing. A junior customer support agent gets more and more effective while the AI provides her the accumulated knowledge of senior customer support agents, as in the recent Brynjolfsson, Li; Ramond (2025) paper, until the AI is good enough that she can be replaced. (...)
The end of work? Not so fast
The other option is to go for a messy job, where the output is the product of many different tasks, many of which affect each other.
The head of engineering at a manufacturing plant I know well must decide who to hire, which machines to buy, how to lay them down in the plant, negotiate with the workers and the higher ups the solutions proposed, and mobilise the resources to implement them. That task is extraordinarily hard to automate. Artificial intelligence commoditizes codified knowledge: textbooks, proofs, syntax. But it does not interface in a meaningful way with local knowledge, where a much larger share of the value of messy jobs is created. Even if artificial intelligence excelled at most of the single tasks that make up her job, it could not walk the factory floor to cajole a manager to redesign a production process.
A management consultant whose job consists entirely of producing slide decks is exposed. A consultant who spends half of her time reading the room, building client relationships, and navigating organizational politics has a bundle AI cannot replicate.
In 2016, star AI researcher Geoffrey Hinton leaped from automation of reading scans to the automation of the full radiologist job, and gave the advice to stop training radiologists. But even fields that can look simple from the outside, like radiology, can be quite messy. A small study from 2013 (cited in this Works in Progress article) found that radiologists only spend 36 percent of their time looking at scans. The rest is spent talking to patients, training others, and talking with the nurses and doctors treating the patient.
A radiologist’s job is a bundle. You can automate reading scans and still need a radiologist. The question is not whether AI can do one part of your job. It is whether the remaining parts cohere in a manner that justifies a role.
To me, a key characteristic of these “messy jobs” is execution. Execution is hard because it faces the friction of the real world. Consider a general contractor on a building site. Artificial intelligence can sketch a blueprint and calculate load-bearing requirements in seconds. That is codified knowledge. But the contractor must handle the delivery of lumber that arrived late, the ground that is too muddy to pour concrete, or the bickering between the electrician and the plumber.
Or consider the manager in charge of post-merger integration at a corporation. Again, the algorithm will map financial synergies and redraw org charts, but it will not have the “tribal” knowledge required to merge two distinct cultures and have the tact to prevent an exodus.
Corporate law is increasingly vulnerable to automation because contracts are essentially code, but I would expect trial attorneys to subsist.
AI implementation itself could be the ultimate messy job. Improvements will require drastically changing existing workflows, a process that will be resisted by internal politics, fear, and legacy business models. For instance, law firms have always relied on “billable hours” to charge clients, a concept that will be useless in an AI world. But this organizational inertia is a gift: the transformation will be messier and more delayed than the charts suggest and it will require a lot of consultants, managers and workers, well versed in what AI can do, but with sufficient domain knowledge to know how to use it and how to redefine the process.
In the extreme instances, the feared AI transformation may not take place. Jobs defined by empathy, care, and real-time judgment will become the economy’s ‘luxury goods.’ In these fields, artificial intelligence is not your competitor; it generates the wealth (and lowers the costs of goods and services) that will fund your higher wages.
[ed. See also: What is there to fear in a post AGI world? (pdf/NBER)
