The jobs most protected from AI aren't necessarily the most credentialed ones. They're the ones where tasks are so tightly bundled together that pulling one thread unravels the whole thing. In this episode, Ben sits down with Luis Garicano, LSE professor and author of Messy Jobs, to dig into what makes some work durable, how organizations need to reconfigure around AI, and why Europe may be in serious trouble.

  • Clean vs. messy jobs: why single-task, easily reinforced work is what AI takes first
  • The unbundling productivity trade-off: why breaking jobs apart unlocks more growth but also drives wages down as labor supply floods a narrower set of tasks
  • Why technology without business process reconfiguration produces almost no productivity gains
  • The junior professional problem in the current landscape
  • Why reduced job mobility might inadvertently push firms back toward German-style on-the-job training
  • Europe's stagnation trap

Ben Zweig: Not all jobs are equally vulnerable to automation and the difference may have less to do with skill level than with the structure of the job itself. My guest today is Luis Garicano, professor at the London School of Economics, former member of the European Parliament and author of "Messy Jobs", one of my favorite recent books. We discuss what AI does and doesn't do to work. Here's our talk. Welcome to the Economics of Work. I'm excited to welcome Luis Garicano, professor at LSE, former member of the European Parliament and author of "Messy Jobs", which in my opinion is the best and most important recent book on AI and work. Luis, welcome to the show.

Luis Garicano: Ben, that's really kind of you to say. I appreciate it and it's really a pleasure to be here.

00:40 What is a messy job?

Ben Zweig: Yeah, it was really excellent and very timely. So I got to start by asking you, what is a messy job and why would someone want one?

Luis Garicano: Okay, so to easily define it, let's define the opposite, which is a clean job. A clean job is a job that is single task, that is verifiable, that you perform, basically that task repeatedly and it's subject to reinforcement learning. The more you do the task, the more it gets improved and eventually AI kind of takes over that simple.

Ben Zweig: Do clean jobs exist?

Luis Garicano: I think clean jobs exist. Many jobs are clean to different degrees. Think of trucks, for example. I had read a lot of articles about trucks, and I always thought the truck drivers did quite a bit of the back end and front end, but they actually don't. Somebody loads the truck, somebody tells them the route, they get the route on the screen, they just drive through the road. So the truck seems as close to a single task clean job as you can get and-- One that is not protected by the messiness, is protected by all the things that we can discuss. So the single task jobs are only protected by regulation and potentially consumer taste or consumer demands or the technological demands on that single task. On the other hand, the messy jobs are bundles of tasks which include relational, which include commitment, which include judgment that cannot be easily split apart. So we talk about the strong bundles and weak bundles and strong bundles are when this coordination cost between the different tasks is sufficiently big that having automation on one task doesn't make it kind of split.

Ben Zweig: Wait, so let's say there's a job that has lots and lots of tasks, but they are easily split apart. So they're weakly bundled. Is that messy or clean? Like how would you characterize the example? I mean, it's a--

Luis Garicano: It's on the spectrum. We wouldn't say it's a messy job, but it won't remain in that way. So it's not the kind of jobs that we are recommending the prototypical person you were asking me to ask. So it's a, these are all tasks, it's a weak bundle. And as a result, messy jobs are strong bundles. So as a result, these will kind of be split by AI. It's not surviving. We talk about the key concept to think about this is the autonomy threshold. So the autonomy threshold that we describe in the book is for a certain level of autonomy, for a certain level of AI capability, how well can AI perform this job on its own without human assistance? The following-- The autonomy threshold increases in these things. It increases in the messiness. So when it's a clean job, the autonomy threshold is relatively low when it's a bundle, but it's kind of easily splittable. It's higher when it's a strong bundle, it's higher when it has very many relational elements, when it's a bundle of relationships, then it's even higher to the point that some of these jobs you would expect never to be replaced by AI. And we can discuss all that, of course.

04:02 Is unbundling good for the economy?

Ben Zweig: Yeah, so I understand for a person, for a worker, they would wanna be in a messy job. What about for society? Like, do we want, I mean, I guess society has two objectives here. I mean, obviously a lot of objectives, but the two objectives I'm thinking of are productivity and growth, and employment, healthy labor markets. So I would think if we want maximum productivity from technology, we'd want technology to be used more widely, we'd want to be able to embed it into processes in the organization maximally. So the ability to kind of unbundle maybe gets us more productivity. If you think about, let's say we imagine jobs are completely weakly bundled, then anything that AI can do, it does, anything that it can't do, it doesn't do, and jobs kind of reconfigure flexibly, and that seems to be the way we get maximum productivity. Is that a fair thing to say, and then?

Luis Garicano: I think that's fair. I think that's exactly how it would work. The productivity gains are bigger. If you can split jobs, then you keep it together, because when you keep it together, you're still bound by the supply of hours of the particular person, right? Yeah, so if the truck driver has a human in the loop element that you need somebody to be there just in case, then at the end of the day, the number of hours driven by trucks is limited by how many truck drivers there are, so the supply increase is not as big. Whereas if you can split it, indeed, and you have one guy up somewhere doing just the planning and the cars, trucks get driven on their own, then as you are arguing, indeed, the number of hours supplied gets unhinged or unlocked from human hours, and we have infinite or very large supply of trucking. So I see the trade-off, and I agree the trade-off exists. I would say, over the long run, what we have generally found is that technology is a good thing for wages and for growth. And as a result, I would expect that creating messiness on purpose, kind of polluting jobs by making sure that all these things stay, is not a good thing. I mean, I would think it's good. A human will look for the places where the messiness stays. It would be good to have those jobs where you're a consultant, you have to convince the client, or you have to extract information from the client about what is the...

Ben Zweig: What I'm really wondering about is, is messiness a good I guess like, I mean, the way I'm thinking about it is that there's two parts of messiness. One is that there's a large collection of tasks for each job, rather than singular tasks, and that those tasks are strongly bundled rather than weakly bundled.

Luis Garicano: That's correct.

Ben Zweig: And it's not obvious to me that jobs being strongly bundled is good for the economy. I mean, you mentioned we don't wanna put artificial barriers, artificial bundle artificial constraints on unbundling. But I wonder if we wanna go the other way. Like maybe is it good for the economy if we get more unbundling and allow jobs to unbundle and rebundle more flexibly? I mean, what do you think?

Luis Garicano: Yes, so let me say that we have a kind of theoretical agreement behind the theory you suggested. We have a paper, Jin Li, Yanhui Wu, and myself were the ones who wrote this piece together. And we show that indeed productivity growth is higher when the bundle breaks apart. There's no question about it. But indeed, because supply grows much more in that case, then you have lower wages.

Ben Zweig: Oh, lower wages? Yeah, why would you have lower wages?

Luis Garicano: So the wages drop by more than they would in a strong bundle case, simply because you have the worker who was doing task one and two is only doing task one. The task two gets unbundled and done by the computer. So the supply into task one increases very much.

Ben Zweig: So the returns accrue to capital, that makes sense. But what about the introduction of new tasks? I mean, if you allow for that, then-

Luis Garicano: Yes, I'm just talking in a very simple framework with two tasks. Let's just talk about the bundling part. You're right that all of these other things will be happening on the background. So essentially, the trade-off that you identify works. And if you wanted to increase productivity, could you eliminate the... So I guess the question that we are kind of finding tricky and interesting, right, is putting the sand on the gears, kind of slowing their recombination, their bundling and their unbundling, kind of preserves things as they are, gives you lower productivity growth, but leaves you with less disruption. Kind of making the mechanism very efficient is going to lead to more turmoil and more adjustment, but also more productivity growth. And I would say, we also talk in the book about the unbundling and rebundling process. I think that the rebundling and unbundling process that we're gonna see is very significant. We're gonna see a lot of tasks get packed together. We talked, for example, about something David Autor has discussed, the new middle-class jobs where the cognitive part gets more bundled inside the physical and emotional part, think of the registered nurse. He or she was already doing, holding hands, putting you the injection, et cetera, gets more diagnosis ability and can do more. And in some sense, productivity-wise, and in terms of middle-class kind of distribution of income is probably likely to be good, although we would have to go to the demand side. That is certainly one rebundling that can happen. There's also a rebundling that can happen which today the Wall Street Journal was talking about, the entrepreneur, the one-person company.

Ben Zweig: The solo entrepreneur, yeah.

Luis Garicano: Yeah, which who can, again, it's a rebundling of many tasks that we're marketing with finance, thanks to the machine, you can do them yourself. So I think the rebundling is a good thing, and I think we should allow the economy to adjust. I don't think it's a different problem than the ones we often face with technology, that is that allowing for adjustment to take place is going to lead to more disruption, but it's what brings you the productivity growth. So I would be more favorable to not kind of putting obstacles, this technological change that will happen. Maybe some slowing down of the recombinations, maybe, but a priori I would err on the side of allowing it.

11:37 Rebundling and business process innovation

Ben Zweig: Yeah, I mean, I guess it all rests on this idea of how disruptive it is. I mean, one way that maybe is helpful to think about it is that you can think of the automation as the offense in a way, that's like automating jobs, but the recombination is kind of defense. So yeah, let me sort of clarify what I mean by that. Like, I mean, you were a former consultant, so you may know this way that consultants get characterized, that people say consultants really do two things, bundling and unbundling.

Luis Garicano: I remember Andreessen, was it Andreessen who said this at some point? It was in a podcast, but I do remember this. Oh, interesting. Bundling and unbundling, yes.

Ben Zweig: Yeah, and it makes me think of this kind of idea that Erik Brynjolfsson and others have popularized that technological innovation doesn't really get your productivity, but it's technological innovation paired with business process innovation. And business process innovation kind of at its core, to me, feels like job reconfiguration. You're trying to figure out who can do things differently. And if you can reconfigure if you've got some consultants, some business process innovators, whoever they are, who are kind of who have the job of reconfiguring and reshaping who does what and how things progress, then if part of those jobs get automated, then the remaining parts can get rebundled in a way that helps workers remain relevant, be productive.

Luis Garicano: Yeah, absolutely. I think that is better than the alternative. So two instances of this. One is the police departments who adopted CompStat in the 2000s and the 90s. CompStat was, for people who were watching us, who watched The Wire, they would remember that the police officers meet in front of big screen, and they agreed on how they're redeploying their officers. So there was a technological innovation. They caught technology that could geocode crime. And Bill Bratton in New York, the police commander at the time, under Giuliani, kind of introduced this new bundle of practices like this new job redesign. Everybody could be moved around geographically, all the officers, the commanders had to go in front of the chief of the police and explain what they were doing and so on. So indeed, what we find in this paper, it's an empirical paper with Paul Heaton from the RAND Corporation, is that the technology by itself almost didn't do anything. When you give the computers to the departments of police, they don't clear more crimes and they don't reduce crime rates. On the other hand, when you allow this whole reconfiguration and business process redesign, all this rebundling of jobs, you get a big productivity increase. So I think that is absolutely essential to get the productivity gains. You're gonna have to allow for these reconfiguration of jobs. And the other example of course, is Europe versus the US and COVID. The way we did in Europe, the COVID response was by allowing the job to keep you employed and paying you the salary from the state to know those months. So that when the job is putting everybody in the freezer, the economy opens up, everybody goes back to the exact time, T equals zero before COVID. And the US didn't do that. It said, everybody out, you get an insurance, you get this future configuration of the economy. And the truth is, since COVID, the US has had big productivity growth and Europe hasn't had any.

15:22 Why the US reshuffled after COVID and Europe didn't

Ben Zweig: So how, I did not know that was part of the theory. So let me just see if I understand. So we have this big shock, which is obviously a labor supply shock, but also a just way of working shock. So remote work and digital tools and Zoom and whatever. So, yeah, I guess like the way that I think about, post COVID recovery in the US is that, we had this like great resignation, great reshuffle, whatever it's called, where everyone sort of rematched to jobs that- And better. Did that not happen in Europe to the same degree?

Luis Garicano: No, it did not happen. The entire Europe had a furlough scheme that was coming from Germany. It was called Kurzarbeit in Germany has always existed. When a company has a problem, it puts people in a scheme from which they get paid and it takes them back. And in this case, that was useful COVID and it was spread to the continent. So in fact, after the COVID, it wasn't a big reconfiguration at all. There was no reconfiguration. So I think there are good reasons. I can see why people are anxious, but I think there are good reasons to allow for this redesign and this processes to change. That requires experimentation, that requires people to be able to make mistakes. And that requires as a result that some jobs are lost and some jobs are gained. But I think that's just the way that the economy keeps growing. That's creative destruction, right? Yeah, creative destruction.

17:00 Is AI a different shock than the China shock?

Ben Zweig: Yeah, I mean, one idea I've sort of been grappling with on this and it's half baked, so bear with me, is that sometimes people contrast technological disruption with the China shock. And I get a little uneasy about the comparison because in the China shock, you had products, there was a shock to like, relative product prices, which affect industries and those are crushed by geography. So you basically go from like industry geography. With a technological shock, it affects tasks, which affect occupations. And that's like a sort of many to many dispersed relationship. And then you go from occupations to industry. And that's also a many to many dispersed relationship. And then you go from industry to geography. And that's also kind of like more dispersed than it was before, pre-COVID. So it seems like you have a lot of dispersion at every step and that the shock is like kind of something that doesn't have very targeted disruption in the same way that the China shock had targeted disruption. So it seems like potentially just more uniform throughout the economy, which feels safer. You have like more ability to kind of adapt. I don't know, what do you think of that comparison?

Luis Garicano: I agree that geography matters. I think when I studied economics, geography was considerably relevant. It was like, well if people lose jobs in geography, they should move to B and get the jobs. The truth of the matter is we now understand much better, if not for anything, for the political consequences, the human costs of making very geographically targeted disruption. And you're right, technology could be broader. However, with AI, I think I am spending a few weeks in the north of Spain in summer home and you see all the people that you walk around all day watching from surf instructors to restaurant owners and waiters and fruit sellers and so on. And you see an entire economy which will not have any impact from AI. And I was with a friend whose kids are doing, they're doing jobs that all of them will be spared completely. Whereas on the other hand, in areas like London, where professional services are very, very significant and where knowledge work is very important, the disruption will be more significant. So I do think that compared to the information technology, which was more skilled biased than occupation biased, I mean, I do think, and your data kind of might speak to this in more clear ways than what I am guessing, but the knowledge work is often very clustered, the cognitive tasks. True, there are lawyers and doctors in every village and every place and there are, but a lot of the high end cognitive work in the US is in Boston, in New York, et cetera. And so I'm not sure that it will be more geographically, as geographically spread as laptops, which basically everybody ended up having one or mobile phones, which everybody ended up having one, whatever disruption was, could probably be considered more broad in the way that you're saying. I mean, that's potentially a concern in terms of where the kind of clean, verifiable cognitive tasks are, I would expect something like places where accountants are concentrated, places where lawyers are concentrated, places where I guess doctors are protected by all the other things we discussed, but like regulation, but those places might be geographically pretty concentrated.

Ben Zweig: Yeah, yeah, interesting. That's a good point. I wasn't really thinking about like these places where everyone's like a surf instructor, like certainly they're protected, you know?

21:21 How AI reshapes hierarchies

Ben Zweig: So one thing that you've argued in the past, which seems so interesting and relevant is that in organizations ultimately the scarcity within organizations is kind of like knowledge and expertise about like how things work so that you can escalate. And in a world where it looks like maybe knowledge, expertise is becoming more abundant, what does that do to an organization? Like how does that manifest to the hierarchies and like how organizations look on the inside?

Luis Garicano: Okay, thanks for the question. I will bring people up to speed since you're making a reference to my work on hierarchy. So basically the idea of the knowledge hierarchy and the high organization of expertise, which I spend a lot of my career working on, was look, acquiring expertise is very expensive. What we do is we leverage that expertise by having the experts only work on the exceptions and having a lot of people work on the routine jobs. And that's a hierarchy which is a knowledge base hierarchy. Basically, the exceptions get escalated. Alfred Sloan liked to say he doesn't, he didn't work on the details because those never get up to him. And when something gets up to him, Sloan, CEO of General Motors, got the exceptions. He said the exceptions, not the petty details. So that's indeed the hierarchy of the organization of knowledge. And the question is what happens when you put an LLM in there? So one thing I would say is a lot of the knowledge that is important in organizations that has to do with implementation of decision, that has to do with what different aspects of the organization get affected by a particular decision, et cetera. That knowledge is not the clean, repeated, verifiable type. This is kind of messy, it's in people's heads, it's tacit, it's as Polanyi would say, it's also dispersed as Hayek would say. And the rarer it is, meaning because of the way the hierarchy works, it's going to be higher up, placed higher up, the exceptions, the less likely that the verification that the algorithm is going to eventually learn it. Because there are not so many of the exceptions that correspond to this company that is tacit within this company that people understand but not somewhere else, et cetera. So my expectations is that the work of a higher end managers will remain more or less undisrupted, that there will be exceptions for them to consider and decisions on those exceptions to be made. That the work of the frontline employees enhanced by AIs will change, it will shift, it will be bundled, it will move up, et cetera. And that the main organizational change will be, middle managers on both sides are probably going to be squeezed. A lot of what they do is knowledge processing, information processing, passing on things, solving the middle problems, et cetera. The machines can do a lot of that. So I would expect the hierarchy to compress, I would expect organizations to get smaller. I don't expect organizations to go away. I mean, the idea that agentic AI can somehow replace an organization, I just don't think that's likely. You know how I introduced the book with this example about the family? Yeah. Just to make the analogy to the CEO. So you have all the cognitive tasks for the following day in your mind solved. You know where you have to take kids, you know what you have to do, you know when you have to drive home, what the breakfast is, what the meal is, and all those things. Then you wake up and your kid says, "I don't want to go to a swimming pool." And that is a task that is a management task to try to figure out what to do with a kid, to try to persuade the kids, to try to convince them, et cetera. That is completely unaffected by AI, right? So the top managerial task of kind of getting your organization from A to B and getting people to want to go from A to B and to all those kind of exceptional decisions, I think stay.

25:36 What parenting teaches about management

Ben Zweig: Let me ask you, I mean, you're a father, right?

Luis Garicano: Yes, I'm a father of two boys.

Ben Zweig: Okay, and you also kind of like are sort of the manager in your life. You manage postdocs and junior authors and all that. So do you think being a parent has made you a better manager professionally?

Luis Garicano: I think those two things are quite connected. I think the kinds of things that make you a good father, setting expectations, living by example, not just what you say, but what you do is what matters. Letting people, delegating and having lots of freedom within those frameworks you have, as you have said, I think those things are being consistent, which is so essential to both. And I think both things are similar. Now, which direction does it go? I mean, it's funny, right? Now what it really tells you about that when you become a father, those things just happen and you hope that you know what you're doing. But I think both are reinforcing for sure.

Ben Zweig: Yeah, yeah, it's interesting. I mean, when you gave that example in your book, I really resonated with it because it is management. Like when you're getting your kids ready to go to school, you have to kind of manage a-

Luis Garicano: How many kids do you have?

Ben Zweig: I have three young kids. So I have three daughters.

Luis Garicano: So the management task is-

Ben Zweig: Oh, it's way more than managing Revelio Labs. Yeah, I mean, my youngest is just a month, so like very, very new. So that is disruptive in a way that organizations also have disruption and like new employees, new people. And like, yeah, it is, yeah, it's like surprisingly, it's surprising how much of a parallel there is.

Luis Garicano: Yes. Yeah. And we want to deploy it to counter this extreme rationalist view in Silicon Valley. When they talk about, you know the data well, when they talk about, oh, I heard a top executive from an AI company, say in a meeting I was, or we've solved coding, we think we'll solve computer tasks, meaning general office jobs within another 18 months. And that means they don't actually understand what is an office job. And that was in part the motivation to do. That's my first paragraph of the book. It's just what our first paragraph is like, what the hell? I mean, there is very little in there that is the exact same thing you did before. And that is kind of, and that is just very variable. And that, I mean, there's a lot of context in all of that. There's a lot of conversations in the corridors. There's a lot of temper tantrums, like with your kids. And that's why I really deploy the family example, because I think that's the way to get to the research, to that question of imagine all the information about the day is deployed in front of you magically and perfectly. Would that get the task of getting the three girls out of the door in the morning, like much easier? A little bit, but there's a lot of stuff that remains, right?

Ben Zweig: Yeah. Yeah. Yeah, it's interesting because that even that, I mean, I think the mantra that you've introduced that I find myself saying all the time is the task is not the job. And that is so, it's so important to remember that. And even getting kids out the door for breakfast is not really a task. It's a bundle of tasks. Ties some of these shoes and get their lunch ready and take them to the there's a lot of stuff that has to be done.

Luis Garicano: Exactly. And somebody says, oh, I have it's like when they talk about a project management job. I mean, it seems like they must be thinking that the project manager sits there all day looking at the Gantt chart. And the thing's so much more messy. I mean, the guy, if your electrician doesn't come, you need to get the painter to come earlier because the electrician didn't come and the painter doesn't want to come. And I mean, the Gantt chart goes out of the door like Moltke said in the first minute of the battle, right?

29:58 Software engineers have messy jobs too

Ben Zweig: Yeah, totally. I mean, that's also true for engineering. Like I think people talk about software engineering is like, oh, it's clearly well defined. You just need to like push your code and like the end. But no, you're also owning something and dealing with all the complexities that come up and dealing with how things fit together. It's like very hard to foresee, but yeah, engineers have messier jobs than I think we are sometimes led to believe by the AI companies who think of that just as the coding task.

Luis Garicano: Exactly, exactly. And the truth of the matter is, and you can tell me what you see in the very, very last batches of the data, but it doesn't seem like the job openings are dropping even in those positions, correct?

Ben Zweig: Yeah, I mean, I think you do see some.

Luis Garicano: You see less growth, right? Less growth in programming and coding and otherwise, right? That's what the research studies show.

Ben Zweig: Yeah, yeah, but it's small. I think the magnitude is something like 5.5% since 2022. So it's like pretty slow actually.

Luis Garicano: And that includes COVID reorganization and all these other things of course.

31:18 A case for middle managers

Ben Zweig: Right, right, that's true, that's true. So let me push back against one point you made about sort of the organization getting squeezed and kind of like squeezing out middle management. So I think the argument you put forth makes sense that you have this kind of like better synthesis of information between the bottom and the top and like potentially cleaner escalation. But let me make a case for middle management being the real heroes of AI. So see if this kind of makes sense. But if we think about work reconfiguration, one way to think about that is that you have things at the top, you have the demand for work activities, you have the demand for tasks as a kind of fluid process. A client needs something, someone there's some priority, some reconfig so you have the people at the top who like have changing needs all the time needs change a lot. So you have some like fluid demand. And then at the bottom you have workers who are executing on things and there's fluid movement there too people decide they don't like something, someone quits, they want to pursue something else so there's like, there's changes at the top, there's changes at the bottom. And the middle manager is kind of the liaison the middle manager has to kind of do this bargaining between like the needs at the top, the capabilities at the bottom. And that is just a complicated thing to continually manage. And one thought is that as technology grows and maybe accelerates, then there's more there's more variance, there's more volatility in what's required at the top, more volatility in what's required at the bottom. And you just have to have this kind of constant task bargaining. And that is the role they're the mediators of these needs. So maybe they can like play an even bigger role. I mean, what do you think?

Luis Garicano: So it's similar to the last phrase of my article, I just wrote a Substack, I just wrote on what are consultants getting paid for, which is like, hey, maybe we have a different effect on managers, which is if we have more proposals and cheaper proposals, and a lot of decisions need a lot of input from many people and information is dispersed and we need to collect all this information and get people, then it could be that many more easy proposals coming up require many more meetings. And at the end, that's more consulting and more management. So I see what you're saying. And I think it's not just legitimate, but actually completely correct. I also think, however, that part of what's going on is like always with productivity effects, you are gonna have, so you have both of what you're saying and what I'm saying. So we have more problem solved per minute, and that squeezes down the ranks of problem solvers. But at the same time, more problem generation as generating problems becomes very, very cheap. And that makes those people who are trying to kind of get the consensus and so on, they will need to be more involved. So I think that there could be a Jevons effect, a Jevons paradox of the kind you're suggesting for managers and for consultants, as there are lots of possibilities. Like we were talking before we started the chat about how I and you and many others are working more than ever before AI. And there is this again, Jevons paradox for effort that even though we are more productive each hour, we are consuming, we are producing much more and doing many more hours. You could produce the same in less hours, but every project seems attractive. And so we generate more and more projects and then everybody else has to also work more and filtering those projects. So I can see that paradox as a possibility and it certainly is not impossible that's what we'll see.

35:50 Is AI making jobs greedier?

Ben Zweig: Yeah, it really makes me think of Claudia Goldin's characterization of greedy work, where you have jobs that just need to take up all your time and you have to constantly be engaged with the work, you can't really step out. And I wonder if technology is making jobs more greedy. And then, of course, as Claudia Goldin pointed out, that has major implications for families. Yeah, I mean, do you think that is likely that jobs become kind of greedier with AI?

Luis Garicano: I feel like that's been the case for me. I don't know if it's a general thing. I definitely, the hours are encroaching on many other family activities. And I don't know if it's a general thing. For me, it has to do definitely with this feeling that, A, so there's the context switching, which means that you have multiple things going on and you can just kind of do many more different things at the same time, you have things in parallel. Yeah. B, the idea that everything that you have just a very preliminary idea becomes feasible so you can explore it, maybe you draw it, but you explore it. And also the sense that, hey, I have this possibility now that I didn't have before, I wanna try it out. So all of those things are different elements in this enormous labor, individual labor supply elasticity that I didn't expect, which is that it's definitely kind of increased for me and for many people I talk to.

Ben Zweig: Yeah, well, in fairness, I mean, you and I are probably really bad examples here because, study organizations and labor markets, and this is like our moment. So it could be that you're so excited and in demand just because this is the world.

Luis Garicano: This is the shock my whole life has been waiting for in some sense.

Ben Zweig: Yeah, yeah, exactly.

Luis Garicano: It's interesting, I left academia to go to politics and I did a few years in the European Parliament and then I go back and then my work is more relevant than ever and all of these things are happening, which I have this prism to watch and to understand. So it's been a really interesting, it's been a very interesting period, you're right.

Ben Zweig: Yeah, do you think, I mean, it's like, I'm glad you're back in academia because you're writing so many interesting things, but do you think you would be doing as many interesting things if you were still in politics?

Luis Garicano: No, no, politics is very hard to actually produce. I mean, politics is the ultimate messy job where you need the coalition of people, you have very, very short bandwidth for each problem. So things come in your desk and there's been people already working on it and they tell you like, okay, if you want you to check or to verify or to approve and you do 2 million things a day, none of them in depth. It's almost the opposite, it's the least cognitive and most relational and most messy and intricate job possible where every day, every means of every day is completely different and you have no idea what's gonna hit you the following day with some big crisis, some big thing that didn't work the way it should and so on. It's very, I can see why people who are adrenaline junkies get hooked on it. I didn't get hooked because I was more stressed than excited by this sensation that everything can explode every minute of the day. I mean, I guess you're an entrepreneur, so this sensation is more present in you than me, this feeling that everything can go bust in five minutes, but in politics, it really is like that. Somebody leaks something to the newspaper about you that might be completely wrong or might be completely invented. Suddenly you're in the middle of a huge crisis that you thought you were on holiday, you thought you were doing something else. So it's very difficult in that world to be intellectually productive.

Ben Zweig: Yeah,

40:04 Is the career ladder broken?

Ben Zweig: Yeah, interesting. Okay, good to know. I wanna make sure to talk about early career stuff because I think that's a question on everyone's mind. And it's people are freaking out, young people are freaking out. So what is, I mean, I think the narrative that I hear a lot is that sort of like entry level work is being automated. People don't rely on the early career people to do kind of the menial tasks, like getting a PowerPoint deck ready. And so the career ladder is kind of broken. And that presents a challenge for young early career professionals. What do you make of that? Is that a real thing? Do you think there's some like-- Oh,

Luis Garicano: I wrote a paper with Luis Rayo on careers in the age of AI, where we have a little bit of a different spin. So let me tell you what we think about apprenticeships and what we think of how they change. The bottom line is it's not all bad. So apprenticeship is an exchange, is a relational exchange between a master and an apprentice. And you basically are solving the following problem. You want to receive general training. You want to learn to be a lawyer. You want to learn to be a doctor. But you don't have the money to pay for that. And as Becker said, the way you solve that is by paying you too little. But the amount of money that they will have to pay you by too little is just not enough. There's a constraint there. I mean, wage, let's say. So the way you do it, the way you solve the bargain is the menial tasks are the currency the junior uses to pay the senior. So I will pay you to be at Revelio and you ask me to do basic programming. And the way I pay you is I do the basic programming in exchange every day I learn from you interesting stuff. So the problem, the way we frame the problem a little bit differently than what usually it's framed at is the AI ability to do those simple, verifiable tasks, including PowerPoints and Excel spreadsheets and all the things you mentioned is devaluing the currency which the juniors use to pay for the training. So the juniors pay for the training in spreadsheets. Now spreadsheets are what's the, I can do millions of spreadsheets in the next five minutes. So don't pay me with that. There is nothing.

Ben Zweig: It's a very nice framing. It's a very elegant framing of this. Yeah, no go on.

Luis Garicano: So the currency is devalued. And what we argue is, look, I mean, clearly to the extent, so there are two things that are going on. Things that you use to pay. So the saleable knowledge is being reduced because all these things are done by the AI. But at the same time, the ceiling, what the expert can do with this knowledge goes up. To the extent that the ceiling goes up sufficiently fast and faster than the floor, that the minimum that AI can do. So what an expert with AI can do relative to what AI alone can do, then there's still a gap that can fill by training and there's still value in this expert training. Starting you higher up, but there's still value. To the extent that the bottom is growing, but the ceiling is not, then it really kind of shuts down the training channel and there's not enough saleable knowledge. There's not enough currency that I can deploy in order to pay you. So two consequences.

Ben Zweig: Well, I'm just this idea of like, if you, I think that makes sense when there is kind of a floor on wages, then I understand that like the apprentice case is like a very valuable edge case. But in the case where there's a very highly paid entry level job, there seems like a lot of buffer for wages to just go down. Like if someone becomes an investment banker or straight out of college, they're doing menial work. And if that gets automated, then maybe instead of getting paid $150,000 a year, they get paid $80,000 a year. Like on the margin, there seems to be a lot of flexibility for highly paid entry level work. So is that just like an issue on wages and less an issue on like, this edge case of like needing a floor?

Luis Garicano: The entry level work rents are very, very high. So if you think of what a lawyer gets paid per hour when they enter, which is like 400 or something like that. And when they get a build up, which is like, I mean, at the top law firm, which is like several thousand, the gap is still very, very large. So the rents you're paying for the training are very significant. And I think they're very large relative to the wage. The wage looks high to us, but relative to the rents that are being captured are not as high. So, but you're going in direction that I did want to go, which is the following. The news is not all bad. So if these kind of very long, if the menial tasks are not the knowledge, if what you're learning is not really to do a spreadsheet, they're just using the spreadsheet to pay for your training. And that shortens a bit. And there's less of that menial part. That's not necessarily bad news. That means the training gets shortened, you get to the senior level faster. That's not very bad news. It's only very bad news when it's so short that then it's not worth it anymore. So there are two effects. One is shortening the training period. Another one is, which is a bad thing, which is a good thing, sorry, which increases your welfare because you don't spend as much time with menial tasks. The other one is potentially fully eliminating the intensive and the extensive margin, potentially fully eliminating the job. My sense is that a lot of those tasks that you were mentioning at the start of your question, the PowerPoints and the Excels and all of those things are indeed well accomplished by AI. And that the content of the training is going to have to be reconsidered and rethought. And it's going to have to move towards those messy tasks. It's going to have to be, those juniors are going to have to be learning more about presenting to the clients, interacting with the clients. I was talking to a consultant, recently it was telling me exactly that. What their juniors do is much more relational than it was for years. A second year junior is doing, first year junior is doing what a second or third year would do in terms of those tasks. So that might be a positive development in the sea of worry for all the juniors.

Ben Zweig: That sounds very nice to me. I mean, I think, when you step back, it seems crazy that these very talented, qualified people are formatting PowerPoint.

Luis Garicano: And really like, if you go to Cravath, so there are two interpretations of what happens in Cravath, right? You spend three years reading contracts because that's how you learned. I'm sorry, it seems too inefficient to me. The other is to spend three years proofreading contracts because it's very important for the managers and for the partners. And that's how you pay for getting trained. And that second is the way we see it. And if that second is true, shortening it is not such a terrible thing.

47:39 On-the-job training vs. higher ed

Ben Zweig: Yeah, yeah, I think that sounds right. I mean, you mentioned this Becker problem before that firms may not have an incentive to train you in generalizable knowledge and have more of an incentive to train you in firm specific knowledge. To me, that feels like, I mean, for sure that's kind of true. Although like you need some general knowledge to do firm specific knowledge. So you have to produce some positive externality. It would be very hard to avoid that completely. And it seems like--

Luis Garicano: I think our model, the way they do the general knowledge, the way you pay for your general knowledge is with these menial tasks. So the firm ends up doing, solving the Becker problem. And the way they solve the Becker problem in our 2017 AER paper is with this relational contract in which each period of time I do some menial tasks and you transfer me some knowledge. And the reason I stay in the job is because the net present value of staying is higher. And the reason you pay me is because you're, what I'm paying, what I'm producing for you in each period is sufficient to, that's an action that you have.

Ben Zweig: One thing, no, that's, it's a great point. I mean, one thing it's making me think of is that, as we think about net present value, like the key parameter in that is how long people stay in their job. It's like the attrition rate. But, one thing I've been saying a lot that I view as a bad thing, but now it's making me think maybe there's a good side of it is that, so right now, I feel like one of the most toxic characteristics of AI today is the job search and match process. That candidates can just apply to 100 jobs a minute and then there's like low information, information jamming, signal jamming, information overload. So it seems like a really bad thing because people are not able to find jobs easily. And I think, I suspect that it's resulting in like low hiring rates and attrition rates. At least in the US, like hiring and attrition, the rate at which people move jobs is like basically an all time low. And that seems really bad because people can't like match and find the right employer. But it does have this silver lining that maybe it incentivizes training more. And at the same time, higher ed seems to be becoming less valuable on the margin. I don't know, maybe that's too spicy of a take, but it seems like--

Luis Garicano: I think it's a likely pattern we will see is the one that you're going to define of a substitution between higher ed and on the job training for the reasons that we are discussing. So the higher education worker that doesn't know how to do anything on the job and doesn't get training relational aspects kind of loses value relative to someone who is being trained on the job and getting all these other exposure. If indeed, as you say, mobility declines, that's more easy for firms to recoup those investments. That was Acemoglu and Pischke's defense of the German training system that it induces more of an investment in training. So this would be a move in that direction.

Ben Zweig: Yeah, that seems like a very positive thing. I mean, it doesn't solve the assessment problem that higher ed presumably solves. But it but--

Luis Garicano: We just stop solving this. We've been shooting ourselves in the foot by not giving informative grades and not using SATs and not using all of this information that we could be using. We are kind of not doing our job of making our credential signal value. And we're gonna have to, in a world where all the other credentials are depreciated, we should go back to a world where we'd say, "Hey come here, you get a credential and you'll get a grade and the grade will be meaningful." I would say the UK universities have done a better job at that.

Ben Zweig: I mean, my optimistic take is that I think the vibe is shifting a little bit. And I think people are starting to feel like, "Okay, this grade inflation is like getting too ridiculous and we have to treat higher education institutions as like more, like more of their purpose in providing assessment and signals. Yeah.

Luis Garicano: Yeah, it's a, I think it's a thing. At least professors are very aware of it. Administrators and parents and students, I know how, where are they? But it definitely is the case that, for example, in UC Berkeley, there was all this discussion and clearly the fact that you wanted to introduce admissions standards and so on. And I think that the vibe is changing, is probably true.

Ben Zweig: At least economics professors. I don't know, the rest of the field.

52:40 Europe's stagnation and the politics of aging

Ben Zweig: Yeah, so while we're on the topic of vibes, vibe shifts, one final question for you and is what is the feeling in Europe about technology and progress generally? I mean, I know you mentioned like since COVID growth in Europe has been weaker than it has been in the US. And that's concerning. How are people feeling in Europe? I mean, what is the vibe of--

Luis Garicano: Very concerned. I think the question is really very, very important. You probably know Stefanie Stantcheva and her coauthors' recent paper about zero-sum thinking. And basically what they find is that zero sum thinking, the idea that if you lose, I win, has spread out and spreads out when there is low growth. So this comes--

Ben Zweig: That idea seems to be spreading in the US quite a bit also.

Luis Garicano: Yes, so the idea that, oh, I control the rents and that way the landlords can't take more money. And without thinking, okay, but where do people get the money to do repairs in the house, they have to invest in housing, they have to build new houses, how are you going to get that? That idea gets lost and it's only about zero sum. It's indeed happening in the US as well. I would say in Europe right now, the progress movement that you see in the US emerging is really not emerging yet. I want to do my best to get it going. And my blog is, my Substack, Silicon Continent is a big part of my effort on that. The general kind of sense of Europe is one of stagnation and of inability to move forward. And it has to do very much with the mentality, but it also has to do very much with the demographics. Politics is very much driven by those who vote and those who vote are overwhelmingly older. They have a higher probability to vote and there are many. And you get a new dynamic Prime Minister in the UK. And what is his big idea? His care homes. And his big speech yesterday was about how it's unfair that people have to sell their house in order to move into a care home. And he's going to do everything to pay that with taxes. And that means a young person who's trying to build his life is going to have to pay so that his grandfather doesn't have to sell his house. To go to, I mean, every time you move from one house to the next, you sell the previous one. Why wouldn't you not do that when you are going into a care home? So it doesn't make sense from a policy perspective, but it makes huge sense from the politics perspective. And the politics of Europe is, we are doing everything possible, think the politicians, to make sure that we have the voters of the elderly. And that means transfers always come first, investments come next, and you see almost everywhere uniformly investment as a share of GDP is dropping while transfers are increasing. That's the European disease right now. And it has a difficult solution because if it was just the mentality in the ideology, you could see a kind of Thatcher-like figure coming and saying, let's free the markets up. But it has to do with demographics. And other people are like, Europe's pretty good. I live well, I have a good place to go for coffee, and I have a good pension, and I have good friends, and nothing needs to change. And that big constituency of nothing needs to change. It's not going to be persuaded because the way they see it takes a fight. And let me give you one example on that. Almost the entire GDP growth of Spain since 2008 has been dedicated to the 65 plus. The real wages have barely increased. The real pensions have increased very significantly. And that tells you like where the actual political power is.

Ben Zweig: Yeah, it's unfortunate because I mean, as we continue aging as a society, we are going to need growth to pay for that.

Luis Garicano: Yes, I think that's my main message from my Silicon Continent blog and a Substack. And that's the thing that I try to get Europe to realize and to accept, but it's going to be very hard. And I'm glad to see the US has this abundance movement, both in the left and the right, that thinks that growth is important and growth needs to come back. That's just not a debate here.

Ben Zweig: Okay, well, I mean, I hate to end on a more pessimistic note, but there was enough optimism talking about jobs and organizations. So thank you so much for being part of the show. I have to recommend your- Yeah, Messy Jobs, it's excellent. And your Substack Silicon Continent is also really great. Thank you so much.

Luis Garicano: Thanks, Ben. It was really my pleasure. I really enjoyed it. Thank you.

Ben Zweig: The Economics of Work is brought to you by Revelio Labs, workforce data for research and benchmarking. If you enjoyed, please remember to rate, review, and subscribe. It'll help us reach other curious listeners. I'm your host, Ben Zweig. Our producer is Cole Wagner. Thanks for listening.

More episodes

All episodes

Try Revelio Labs

Start understanding the world's workforce.