Where did all the entry-level jobs go? (With Revelio CEO Ben Zweig)

0:00Today, I want to talk about jobs. There are lots of worrying headlines about the job markets. We've heard about big layoffs at firms like Amazon and other companies going into hiring freezes like Walmart. We've heard that entry level jobs are under pressure from AI, that the amount of graduate hiring there is, is going down. So what should we make of this? Is this artificial intelligence and its early effects on human work? Or is AI just a convenient excuse for bosses looking to manage their costs? To explore this, I've asked Ben Zweig of Reveglio Labs to join me. He is one of the more thoughtful founders and thinkers on the topic of the labor market.

0:48And in this moment in time when the BLS, the Bureau of Labor Statistics, has gone dark and become unreliable, Ben's work at Reveglio Labs is providing a private sector take on what's happening in the job market at large. So, Ben, I hope you're going to be able to bring some clarity to this confusion and join me as we jointly explore what's happening to the world of work. Yeah, thanks for having me. Let's get started with what the headlines say. The headlines for a decade have talked about the threat of automation to the world of work. I think it's now been 12 years since my friend Carl Fry and his collaborator Michael Osborne published that famous paper about the way in which machine learning might affect human tasks.

1:38And as we move into 2025, we are starting to hear stories. The entry-level job market in particular seems to be collapsing. When you look across your data at Reveglio and give us a moment to understand what data Reveglio looks at and gathers, what is the most accurate description of what's happening in the labor market? I think you're largely right that the job market is not very good right now. So hiring is at a low. Also, attrition is at a low. So people are not leaving jobs as much, even despite increases in layoffs. But, you know, separations from jobs are not happening to the same degree as they used to be.

2:21And that is particularly affecting young workers. But what's interesting about this is I think you kind of touched upon this Brunjolfson paper and others out of Stanford that highlight that there is an interaction between entry-level workers and AI exposure. And that's research that we've done at Reveglio as well early on. But I think the researchers have gone a little deeper into it. And I think it is credible that AI seems to be affecting entry-level work. But notably, it's not causing technological unemployment like was predicted by Osborne and Frey or was predicted 100 years ago from John Maynard Keynes.

3:00You know, in 1930, he wrote, you know, economic possibilities for our grandchildren and predicted in 100 years we'd all be working two day work weeks. I think it was 15 hours. So it was a 15 hour, 15 hour week. It hasn't quite been 100 years. So maybe maybe in 2030, we'll see that he's right. But it doesn't look like he'll be right. So we're not seeing technological unemployment, but we are seeing decreases in demand for entry-level workers. And that is particularly concentrated among AI exposed roles. I think you could say, oh, well, you know, AI, you know, seems to be affecting the execution of tasks, which, you know, young people do.

3:34And rather than the orchestration, which favors more senior workers, people with experience in management, because, you know, in the age of generative AI, we're essentially managing these bots to go and... to go and do things and execute on some subtasks. And I think that that is a read. Personally, I think I'm not quite sold on that yet because we're not seeing so much substitution. Like we're not seeing enough substitution that you would think would result if that were the main effect. I think what we're also seeing and what's conflating this result a little bit is that there's a lot of uncertainty in the labor market, partially due to AI, partially due to policy.

4:13But I think as a result, employers have a kind of high discount rate. They're optimizing for the short term more than the long term. And young workers generally, entry-level workers, are workers that are more uncertain. They require more of an upfront investment, and maybe they'll work out, maybe they won't. So it's a kind of high-risk, high-reward type of bet, whereas more experienced hires are the safer bets. So if someone's just looking to optimize for the short term, I think they'll generally favor more experienced workers. That is such a fair summary. And I'm in danger of agreeing with you too much.

4:49So you're going to allow me to push back on a few of those observations. But let's unpick some of the points that you raised. You talked about Eric Brynjolfsson's paper. You've also talked about this idea of managers becoming more risk averse. and if you're more risk-averse, taking on board a millennial who might need their kombucha and early break on a Friday afternoon to go for a meditative stroll is higher risk than whipping the corporate wage slave in their 50s for another few hours in the week. That's a decent picture. But the other thing that you drew up was just how complicated and multi-factor labor markets are.

5:34And one of the things that strikes me about labor markets is that there are so many forces at play from interest rates to the effect of the pandemic to political uncertainty impacting how long bosses are willing to make investments for that it's really difficult to unpick those causes. And the starting point for that needs to be the data that you can look at and that you can make sense of. So we would typically go off and use BLS data or similar data like that in the United Kingdom or Yale University has their budget lab and they run surveys. What is the way that you look at to make sense of what workers are doing and what hirers are doing?

6:20So give me a sense of the sense of the sensing that you're doing, the data that you're gathering. We started Reveglia Labs with the goal of producing high quality data that can be used for any downstream analyses. And right now, we are not seeing government data. You know, the BLS hasn't put out their employment data in over two months. The government's been shut down and there's a lot of turmoil there. UK public employment data is not fantastic either. I think what we're really trying to do is kind of start from very micro level data. So actual individuals, you know, try to understand like the actual firm dynamics at any given company and then aggregate that up globally.

6:59I mean, how do you do that? What are you going to show us? So you're going to show me because you've crawled LinkedIn and you've figured out how many customer service reps there are at United Airlines compared to field engineers. And that gives us a number or is it not as granular as that? Or is it more comprehensive than that? That's exactly right. Yeah, we start at the micro level data, you know, individual profiles and resumes and stuff like that. And then we can see headcounts, start dates, end dates. We can see inflows, outflows. So we can see the stocks and the flows of people. Of course, there are biases in the data.

7:29There's, you know, there's sampling bias and that can be adjusted for. There's lags in reporting and that can be adjusted for. I mean, not easily, but, you know, it's possible to, you know, that's why we have all these data scientists and data engineers kind of working hard. But yeah, we need to adjust for issues with the data. We have to classify it to the proper occupations and industries and, you know, all these different dimensions of employment. Once the data is properly de-biased and categorized, then it becomes possible to see actual stocks and flows of employment in the whole world. So I think one real limitation to government statistics is that they're done domestically.

8:04So the U.S. has the BLS and every country's got sort of their version of it and they do things differently and have different assumptions and it's hard to standardize between those things. So if we are looking to analyze some policy or even just like technological adoption that happens unevenly throughout the globe, like comparing countries to each other seems really important. Public statistics are really limited for analyzing things like policy and technological change. So let's get a sense of then go back to what we really think is happening. You talked about the Brynjolfsson paper. And the Brynjolfsson paper was something that came out in the summer.

8:46It was in late August, early September with one of Eric's postdocs. And they looked at a lot of hiring data over a couple of years, two or three years. And they found that if you look to AI exposed occupations, computer science or programming and customer service, the level of hiring for entry level jobs was much lower than in occupations that were not exposed. But in those same occupations, mid and senior level hiring or employment levels had not declined. In some cases, it had increased. And the argument was, and I think you put it earlier in one of your answers, it's challenging to hire somebody young because while they're cheaper, they need training.

9:31They may not have decided that what they want to do is be a COBOL developer or a CSS jockey. And so they might flake out after six or nine months. So again, why not just, you know, whip the 45-year-old for a few more years? So that was Eric's argument. And the paper was called The Canaries in the Coal Mine. It did really well. Understandably, I thought it was a great piece of work. the buts. And I think this is why it becomes complex, this discussion about what are we actually seeing? So the buts from my side were, number one, it felt really early for ChatGPT and Gen AI, which are not even three years old, to effect hiring.

10:09I mean, I've worked with big companies. They do not move with the degree of alacrity that you might expect. So that was my first challenge to Eric. The second challenge was simply the idea that sometimes, especially if you're having to sail uncertainty. You may have overhired in previous years during COVID. You may have cost concerns. It's much easier to not hire a 24-year-old than it is to fire a 44-year-old. So as a manager, you like water, you'll take the path of least resistance. I suppose the third argument was really what are we seeing here? Because a few weeks later, the Yale Budget Lab or the Yale Economics Lab came out and said, well, we've looked at you know, widespread data, and we can't see these effects at all.

10:54Now, of course, Eric and Barat's effects, as they described in Canaries in a Coal Mine, was highly, highly plausible, right? It fits lots of mental models we might have had. But equally, there are mental models and arguments that I've put forward that might be as credible. So when you look at that through your lens, how would you disentangle what was really going on? We did the same analysis at Reveglio, and we found actually before before either of these papers came out and we we found that this phenomenon that eric and co-authors did document is actually happening you know we had different measures of ai exposure but found the same the same pattern so so what was that pattern that revelio found like specifically it was that the difference in ai exposed roles and non-ai exposed roles was more negative for entry-level workers than it was for more senior workers so the idea is that you know AI exposure is more harmful to more junior employees than it is to more senior employees.

11:52And what does AI exposure mean? I mean, who's got an AI exposed job? Does a graphic designer have an AI exposed job? Does a paralegal have an AI exposed job? What is the measure by which you assess AI exposure? And how do you control for factors relating to the industry or the sub-industry or the firm itself? The most important thing to understand is that exposure is not something that it really doesn't happen at the job level. It happens at the work activity level. You can think of jobs as collections of work activities, collections of tasks. And, you know, AI does not automate jobs wholesale.

12:29No technology really automates jobs wholesale, but automates components of jobs. There really is no job that has zero AI exposure. Every job is somewhat exposed, but some jobs are more exposed. So, you know, you mentioned like like paralegals and stuff like that, there's a higher prevalence of tasks that have high AI exposure and a lower prevalence of tasks that are more immune to AI, specifically generative AI. So this is not a commentary on like, you know, technology at large. This is really this specific technology. I believe the pattern, but I also believe that it's not so plausible that these companies are just substituting labor for AI that quickly.

13:07I mean, you know, you said you've worked at large organizations. I've worked at large organizations. since I used to run workforce analytics at IBM. And I know that like nothing gets done the way they want it to get done. I don't know when they'll start actually adopting AI technology, but you know, it's not going to be overnight. And I don't really believe that firms are as sensitive to this technology in how they can substitute labor, which leads me to believe more that they're just taking safer bets and they're discounting the future more. And I think what we're actually seeing is less of a commentary on what AI is doing, and it's more about what the expectation of AI is doing.

13:46There is some thought that, you know, oh, AI is very powerful, it demos very well, you know, you can do all, you know, you dazzle people, and someday it will have a big effect on workflows in a large organization. Maybe not today, but since we expect that that might happen, maybe it's better to take safer bets while we wait and see how this thing plays out. We have this ad that goes out around Christmas, which is a dog is for life and not just for Christmas because lots of kids get bought puppies and they have to keep them for a long time. So a young grad hire is for three years, not just for three weeks.

14:21And your argument here is that even if AI can't do what people say it might be able to do, the fact that the managers believe it might be able to do that has them looking at these long, long -term decisions, which is hiring a person, making a commitment to them, being willing to work with them for two, three, four years. They're putting that off. And this actually, Ben, is super fascinating. because a mismatch between expectations and reality can be thought of as exuberance. And that mismatch, when we think of stock markets and the stock prices, is the exuberance that leads to bubbles. It leads to the idea of ultimately disappointment and the wrong allocation of capital and resources.

15:14So what I'm hearing from you and sort of drawing a thread out is, if there are these expectations about what AI can do, encouraging companies to take decisions about 2027 and 2028, which is what the hiring of a grad in Q4 2025 is actually about, and those expectations are not met, that is actually a kind of a fundamental problem, not just for the graduate, but actually also for the company. How do we test and identify that it really is the expectation mismatch hypothesis that you've laid out here and not something else. It is still very early. And, you know, you mentioned the title of that paper was Canaries in the Coal Mine, really just because, you know, we're looking at very early signals.

15:56But I think as more data comes out, I think what I would want to look at is, you know, what actually happens to adoption within firms. So are firms adopting this technology in their workflows? I think we'd be able to see that from, you know, what people say on their resume ascend profiles. Like, you know, what do they say in their bullet points, you know, like on their resume? What are the work activities that show up? Even in job postings, I think we'd be able to see it. By that, you mean that I'm a large company, I'm hiring, and I say must have advanced prompts and context engineering on at least two different LLMs as a requirement.

16:32And that is an indicator that you attract. I think you could see it in qualifications and requirements. I think you could also see it in the sections on responsibilities. So if they say, you know, your job, you're being hired to do these, you know, six responsibilities and they outline, you know, responsibilities that are really about, you know, interacting with AI systems, we would reasonably conclude that this actually is a responsibility that they're giving to people in the actual job. And it actually is a core part of their workflow. So I think if we start seeing more of that, I would say we have a lot more evidence that there is some adoption going on, which, by the way, adoption doesn't necessarily mean that this will lead to labor substitution.

17:13Like you could have adoption of AI that augments certain workers and creates more employment opportunities. I mean, we should touch on that point as well, right? Because of course, Excel showed up 35, 40 years ago, and it's created more work than it's taken away. What would you look out for? And where have you been able to see places where AI is enabling more work through lifting productivity rather than going into this task substitution modality we've been discussing? I think here's how I would kind of frame up this question. So we mentioned, you know, jobs are collections of tasks and activities.

17:51As some tasks get automated, let's say, you know, you have 20 different tasks or activities or responsibilities in your job. And let's say, you know, six of those get automated. then you're left with 14. Do you double down on those 14 and get extra good at those things, or do you take on more? That's a big question. How quickly can jobs kind of reconfigure? So I think that depends heavily on the organization. So while we're on the topic of good academic papers, there's this wonderful paper called Adaptive Organizations, where it goes through this model of firms as being either adaptive or procedural.

18:30So in adaptive organizations, I mean, I say this like as a founder of a startup where, you know, people's jobs change all the time. So, you know, and this has nothing to do with technology usually. Sometimes, you know, we get a client request and, you know, we say, oh, now there's extra work to do. Who could do it? Who's interested? Who wants to take this on? And, you know, we figure it out. Or someone leaves and we have to reallocate their work or someone says they're not interested in something or we automate something and we like don't have to do that anymore. And we kind of we're constantly reconfiguring what people do in their jobs.

19:03And I think it's a very common pattern for someone to enter a job thinking they're going to do one set of things. And then as the job evolves, they find themselves doing totally different things. And that is the mark of an adaptive organization, an organization that can be fluid, that can think about jobs fluidly, where people, you know, do what the business needs and business needs change all the time. So if you're listening to this and wondering, when are they going to tell me what I need to do to get a job in this tough market? Just hang in there because at the end, Ben and I give some really specific advice.

19:36This point about the adaptive organization is a really, again, it's one that we need to unpack. So when I go off and talk to C-suite execs, I've had a discussion with a chief human resources officer in a firm with 47 ,000 employees, so of a reasonable size. And she was saying, see jobs as bundles of tasks, units of work that need to be done rather than as single lumpy activities. And we're now going through a process of unpacking what each job is into a series of tasks. And they're having to do this from both the top-down and centralized team, but also enabling individual managers to look at their teams and teams of teams and say, what are the tasks that comprise which jobs?

20:21It's a big effort. And of course, economists have been talking a long time about jobs as bundles of tasks and tasks having certain skills and activities attached to them. It's super theoretical. It is super theoretical. I mean, I have managed teams of my biggest teams, you know, hundreds of people. We never thought in those terms. So perhaps let's break this down with some really simple step-by-step. Give me a real example of converting, you know, a specific job that people listening might recognize into its set of tasks and help us understand how would an HR manager do that and how many of them are doing that?

20:59Organizations generally don't have good taxonomies of work activities and tasks. And that's something that, you know, we're trying to do. You know, I wrote this book about job architecture, which really is about like, you know, helping create taxonomies of what people do in their jobs and how to think about jobs as bundles of tasks that change. So I think having the right data behind this and the right taxonomies and categorizations is really critical if we are going to manage this purposefully. Now, of course, we can manage it, you know, reactively. And, you know, that's not always terrible for a small company.

21:31For a big company, I think that becomes much harder. But, I mean, I'll give you a personal example from my company. We have a team of economists. And, you know, we have 12 economists. You know, they all have PhDs in economics, and that's their background, you know. And the term economist is really, it's more of an identity than an occupation, because what they actually do is whatever needs to be done. You know, they write newsletters and we produce content and they do, you know, we work with journalists and they do media relations. And we just started producing public labor statistics. And that was just a few months ago.

22:04And now they are involved in creating public labor statistics. That was not in the job description. And sometimes we talk to a client and they want to understand the data and talk to an expert and they get on client calls and they are the trusted expert of what we need. And sometimes those responsibilities shift. Sometimes they do more client work. Sometimes they're producing reports or newsletters. They sound very productive. Are you sure they're economists? Yeah, yeah. Not like economists I've met, to be fair, but yes. You know, we call them economists because that's their training. But, you know, what we actually have them doing is, you know, some mixture of client success and media relations and data engineering and data science.

22:45Their title is really is really more a function of the talent pool, less a function of their occupation. That's a good example because you've you've unpicked the different things they do. But what that means, most people probably not working in firms that are as dynamic and adaptable as a technology startup like your own. They'll be working in small enterprises where perhaps the HR manager doubles as the finance controller or as the head of legal, or they're working in these enormous organizations where change management works up hierarchy up to hierarchy. I'm really looking at this notion of how do you turn the theory, which is a job isn't a job, it's a bundle of tasks that a person does, and that bundle might be a little bit dynamic or a bit more static.

23:35And it's the understanding of that bundle of tasks that helps us make most sense of how you develop that individual in the shape of a series of automation and augmentation technologies. You know, that is the theory and there's endless white papers. Entire parts of the Amazon have been cut down to print out the white papers from think tanks about this over the last decade. But let's turn that theory into like really practical, example-driven reality. At these very large organizations, there's a lot more process. And if you have an organization that is procedural, you may have a job within that organization where someone's job is very narrow and they are a specialist in, you know, one or two subtasks, you know, one or two tasks where they are hired to do kind of like singular things really well.

24:33You know, you can think of them as like on some sort of assembly line. They do the same thing day in, day out. That's work that I think is getting less and less common, you know, work that is like super, super specialized. But it still happens in large organizations. And if that is how we think about work and workers, then it really is possible for that work to get displaced when those tasks get automatable. So that is concerning. And I think when there's a lot of rigidity around what people do in their job, you know, either from occupational licensing or just organizational like standards or just, you know, wherever rigidity comes from in these like big bureaucratic organizations, that that can inhibit our flexibility to respond to technological shocks.

25:20But not just technological shocks, also just needs of the business. You know, people get cut because, you know, businesses need different things as they evolve. I am concerned for these big organizations. I do think that their work can become more adaptive. So I think in the 60s, Peter Drucker wrote about how workers are going from these Taylorist workers where they're super specialized in one thing to knowledge workers, where their job is actually more about connectivity and coordination and navigating complexity. And in that world where you're navigating complexity, I think you do have more diversity of tasks.

26:00And the work is more about orchestration among people that execute subtasks. So I think we are becoming more like managers and knowledge workers. I mean, Drucker was a brilliant, brilliant character, right, who foresaw a lot of this. There still feels to me like there is a series of complexities and nuances that are hard to capture. You've built a taxonomy of tasks in this job architecture, which is the theory encapsulated in your book. That idea that, well, we can construct a taxonomy and we can perhaps look at an organization and figure out the propensity for particular tasks to feature in a job bundle.

26:41and therefore if Alice has this job and Bob has that job, we can get a sense of what bundles of tasks they look at. But of course, that's a correlation and that correlation might be high, but these people are real people with real experiences, real things that they do that might not show up in the organogram. There's a whole bunch of relationality. So Alice is one of these hyperproductive economists, but what you don't realize is she's phenomenal interpersonally and the head of sales always likes taking her along because she can open that account there and then and where is that captured and it's not captured in the job architecture it's not captured in the statistical top-down process and it's that nuance you know maybe it's tacit knowledge as Polanyi might have talked about it's also relationality how do you operationalize this given those complexities and what is your advice to companies thinking about that?