Pod Insights Azeem Azhar's Exponential View AI is eating into entry-level jobs About this episode This is the single most important paper to come out in tech in recent weeks. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen investigated whether generative AI is leading to job losses in roles most exposed to AI – and how these effects differ by age and the way AI is used. In this episode, I break down these results and their implications.
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Azeem Azhar? Now, what I would like to do today is talk about what we're learning about how artificial intelligence may impact the labour market. There's been so many words spilt over the past decade and more about what it might mean for jobs, creating this lovely portmanteau, the job apocalypse. And we're looking for data for evidence of that. And this week, my friend Erik Brynjolfsson and two of his collaborators came up with a new paper, which I think is really robust and very, very interesting. Eric is a professor at Stanford University, where he runs a digital economy lab, where I'm a digital fellow as well.
0:45 Azeem Azhar? And in this paper, Eric and his collaborators, Bharat Chandar and Rui Chen, analyzed payroll data from ADP, which is a really large payroll processor globally, but very, very strong in the US, handling payrolls for millions and millions of American workers. And they were able to look at this data, correct for various confounders like COVID and seasonality. And they identified a really, really interesting finding. And that finding was that early career employees, people between the ages of 22 and 25, in the most AI-exposed roles, roles exposed to artificial intelligence,
1:30 Azeem Azhar? such as customer service or software development, experienced a 13% relative decline in deployment from 2022 onwards. So this is a really important key finding. It's a substantial decline in employment for those early career workers. And it was really focused, concentrated on AI-exposed occupations. In contrast, for workers who had more experience in those particular professions and areas, employment actually went up. You know, mid-career software developers, you saw a 10% increase in employment.
2:12 Azeem Azhar? In the case of these early career employees, 22 to 25, there was a 13% decrease. So it's really, really meaningful, that bifurcation. And if you looked at less exposed occupations, so you looked at sales and marketing or health aides or stock clerks.
2:29 Azeem Azhar? You didn't see the same suppression of employment, particularly for that early career employees. So if you look at health aides, for example, over that period of time, there was roughly a 20% increase in hiring for early career workers. We see that there is a pattern which appears in jobs where AI automates rather than augments work. It appears in jobs where today's AI tools can actually be used in meaningful, meaningful ways. The thing that is really consequential is that while this isn't causal proof, the researchers have done a really, really sterling job ruling out obvious alternatives.
3:11 Azeem Azhar? So trends that existed before artificial intelligence, COVID-19, education differences, firm and industry shocks. So where they end up is the most likely explanation is the impact of artificial intelligence. So the question is, what's going on in here? What's this picture? There's a reasonable intuition which is reflected in the paper, which is that artificial intelligence is supplanting the bookish knowledge that a 22 or 23 year old has. Those workers have learnt through high school and university
3:51 Azeem Azhar? formal technical skills, but what they haven't necessarily established is practical skills to work, what the workplace is like, and specifically the tacit knowledge that exists, not just in work in general, but in specific firms and specific departments within specific firms. So that tacit knowledge is something that isn't in your degree syllabus for the moment. It isn't in the company manual that you get when you you get a job offer. It's something that you learn over time that shows up when you go through the job selection process, through the interview process. And I think you can, you can see some additional evidence, support for that theory.
4:36 Azeem Azhar? Because in the highly exposed occupations, that is occupations that were really exposed to AI. Older cohorts added jobs, six to nine percent. So that deltas thirteen percent down for first early career workers and maybe nine percent up for more experienced workers. So that seems to be a strongish sign that judgment and experience is valued. And in a sense, that fits what we might have seen elsewhere and what our intuitions might have been. That artificial intelligence is taking over the production side of the equation. Large language models operate at scale today. They draft thirty percent of Microsoft's code,
5:19 Azeem Azhar? processing billions of lines a day through tools like Cursor and other coding tools. They're starting to automate the creation of slides for management consultancies. And, and I think that model of junior workers generating output and senior workers judging that output, it is no longer one that makes as much sense in a world of AI. Because you get the AI to generate the output and then you go off and judge it. And, and when you think about, if you've worked with software development teams and engineering teams, a really large part of working effectively in that team is knowing how that team works, right? When do you do a pull request? How do you note up your code?
6:03 Azeem Azhar? When do you go and talk to your senior developer or your engineering manager? These are all softer skills that you build with judgment and with, with experience. So Eric and his collaborators frame this research as a canary in the coal mine. You know, we used canaries in, well, we, our forebears used canaries in the nineteenth century in the coal mining industry because the canary was really susceptible to the leak of, you know, poisonous gases. Canary keels over, miners know, it's time to get out of that, out of that mine. So I think that there is something really suitable with that analogy. But I also, as I'll discuss over the next ten or fifteen minutes,
6:47 Azeem Azhar? think that we're not quite out of oxygen because it does take time for firms to reconfigure around a new technology. It's only August twenty twenty five. This fieldwork was done for a couple of years. Running through the earlier part of this year. It's very, very early days for large numbers of companies to do anything but dabble in the technology. And at these moments in time, it's much easier to do non-hires, that is, stop hiring people. And where do you stop hiring? You stop hiring in the graduate entry funnel than it is to get rid of people.
7:29 Azeem Azhar? Right, a firing is much harder than a non-hiring, especially at these moments where things are quite turbulent. And so I think one of the things I've been thinking about is that we're quite early in this. So both this survey and the data they used being incredibly robust. I mean, we're lucky to have this sort of depth of the survey and piece of research at this point in time, but also a recognition of where we are in the cycle. So with that, I just want to dig into how I had been critically analyzing and assessing this report. And I did have a chance to congratulate Eric by WhatsApp and have a little bit of a back and forth with him, but I won't represent what he said to me in that WhatsApp discussion because I may not do justice to it.
8:13 Azeem Azhar? But here is some of the things that I think are worth adding as not perhaps not qualifications, but just lenses on the research and what it could be telling us. The first is, who does ADP serve? I mean, ADP is this big, well-established business and it really has a middle market focus, serving smaller and regional firms with strengths, particularly in the northeast of the US. So these may well be firms that some of you have heard of, but many of us won't because they aren't the big global franchises necessarily that we think of, that, you know, the Starbuckses or the Microsofts. They are by and large smaller and more regional firms. So these findings may reflect a different tier of the labour market.
8:58 Azeem Azhar? And what do I mean by that different tier? You know, a lot of the focus on hiring. Obviously is around nameplate firms who we all recognise. It just makes life a lot, a lot easier. But this may not be reflecting those class of firms. It may also be reflecting companies that are sufficiently small that they outsource their payroll. Right. They don't run these capabilities themselves. And it made me wonder about the extent to which companies like this are actually quite sensitive to cost savings compared to really big firms that have big funnels of hiring. Right. How sensitive would they be shaving off a few new headcount may make
9:43 Azeem Azhar? a meaningful difference for those companies. They also might be able to respond more quickly. I mean, this is a, this is a canary in the coalmine moment, right? It may be that a mid-sized regional firm can respond and retool around AI far faster than a big, huge, regulated national company that's just going to take a little bit longer. In which case what we are seeing is a, a, a pattern that could get repeated across the economy. I've been thinking a little bit about the geographic concentration, that geographic issue and, and what it tells us. And the extent to which, you know, what we're seeing is a large number of mid-market, mid-sized firms coming to a similar conclusion about how they should think about youth workers and in the space of, of AI
10:27 Azeem Azhar? enabling higher productivity levels, you know, with people who are more experienced. So I, you know, I'm still playing around with that, that idea. I think the, the second point I would make is the two main sectors that were impacted because of their exposure to AI were customer service and software development. And those two are quite different classes of work. Customer service is very much closed-end class of tasks. You know, you have a range of customer service issues. We already know that AI systems can, through chatbots, answer 18, 19, 95% of queries.
11:05 Azeem Azhar? But you have a very different class of work because you're ultimately in a kind of responsive mode to customer input. Whereas software development is a much more open-ended task, right? There's no company that doesn't have a backlog of development projects that they will never finish because as they finish one piece, they add a new requirement. And so there's a question there about whether the same underlying dynamics are at play for customer service as are at play for software development. So it could be that it's the same dynamic and that is a, that is a kind of complicated position for us to find ourselves in because the argument there would be that artificial
11:49 Azeem Azhar? intelligence tools are sufficiently generalized, they know about everything from, you know, the World Series to the MSCI world benchmark and everything in between. And so what we're seeing is the generalizability of that technology impacting two very different classes of, of work. Or it might be that we're seeing two different etiologies, right? Two different paths to affecting customer service versus software development hiring. And I think it'll be interesting to, to unpick that because, you know, one speaks about, you know, a tidal wave that comes in and sort of affects all boats. And another speaks about much, much more tactical sector level granular effects that
12:33 Azeem Azhar? we need to unpick and understand. I, I did wonder why companies reduced hiring rather than cut, cutting graduate salaries, right? A typical economic analysis would be, well, if you get into this situation, you, you have two strategic choices. You could hire the same number of people and offer them, you know, lower salaries or you could just keep the same salary and hire, hire fewer people, which is what firms seem to have done. That's completely understandable in a sense, right? Internal pay structures are a little bit rigid, but I think there's a really interesting dynamic, which is that it allows firms to exhibit preference for higher quality, smaller workforces over pure cost optimization. So let's think about this, right?
13:14 Azeem Azhar? If firms are acting in a rational way when it comes to their hiring, they will very much try to hire the best people that they can.
13:26 Azeem Azhar? And let's say their selection process is perhaps not perfect, but very good at determining who's a great candidate, who's an average candidate and who's just a kind of good enough candidate. This result might suggest that what companies are able to do is slough off that bottom decile or bottom two deciles of new hires they might have otherwise done, because they feel they can achieve their commercial and business objectives with a smaller workforce, which through this selection process will end up being the higher potential group of people who get the offers or not. And that, I think, is quite an interesting shift. It's an interesting dynamic. I mean, of course, like you always have bottom, top and bottom quartiles, but what I'm suggesting is that
14:11 Azeem Azhar? the sort of, it could be the case that the absolute capability of the bottom quartile rises when you reduce your hiring and you can therefore be a little bit more selective. A quick note, if you want to support us in bringing more of these conversations to the world, please consider subscribing to the show. I think there's, there's another question, which is really about what can we say this early into the, the deployment of, of AI technologies in companies? I've always argued that what history has shown is that you start with relatively low hanging fruit and it takes time to re-engineer your, your processes.
14:54 Azeem Azhar? And my favorite example of this is to think about electricity back at the turn of the, the 20th century. So when electricity shows up, lots of factory owners end up buying into electricity early. I mean, they weren't slow, but buying in meant you put some pendant light bulbs in your workshops so that workers could work longer hours. You could perhaps extend the, the working day. But you didn't change the fundamental power delivery, which was a sort of steam driven power delivery, or in the car industry, very artisanal working, looking not too dissimilar to a, a coachworks or a, a blacksmith's workshop. And it took time before you re-engineered and rewired processes. Henry Ford in the car industry being the best example of somebody who believed
15:38 Azeem Azhar? in electricity and built a production flow that could only work with, with electricity.
15:46 Azeem Azhar? And so if we're at that early stage where this is about buy-in rather than being able to deliver belief, where existing firms, particularly mid-sized firms, which generally may be able to make decisions more quickly, but they don't have the working capital, they may not have the talent depth in the leadership, they may not be able to to make the investment for a full re-engineering and reorganisation, what we might be seeing is the effect of low-hanging fruit of simple productivity and quick cost savings, rather than a fundamental pattern of what might happen as artificial intelligence comes in and helps people rethink
16:30 Azeem Azhar? what work looks like. And you know, this tacit knowledge piece is is quite important for all of that, because tacit knowledge is often the je ne sais quoi that makes firms do what they do, be who they are, be as good as what they are in the sectors they are and serving the customers that they do. And that experiential workplace wisdom may explain why younger workers are going to be vulnerable. But ultimately, younger workers are cheap, they're probably very willing to learn, and they may actually be more AI au fait than older workers. So in my mind, there's also a question about, does this pattern hold over a few quarters?
17:14 Azeem Azhar? Or does that pressure on higher-paid, more experienced workers start to start to show up? There's another dynamic that I want to introduce here. So Darren Asomoglu, Nobel Laureate and his long-term collaborator Pascal Restrepo, talk about something that is probably best described as as business stealing. So when they looked at automation, over a few decades, this would have been robotic automation in in factories and so on, what they identified was that firms who which went early in automation actually ended up growing their labour forces, jobs that were lost as a consequence of of automation. Were lost in firms that didn't automate.
17:58 Azeem Azhar? And I write about this in my first book. You can see it over my shoulder there where my finger is. And what I describe essentially is my hypothesis around this is that to get involved in automation is a difficult, complicated thing. It's not just about understanding the technology. It's also about being able to manage that change and communicate it to your workforce, get them motivated. All of those things are signs of a good, great management. And so you could say that a proxy for great management might be firms that can roll out automation very, very well because those are difficult projects. And so where we end up is essentially through Restrepo and Asimoglu, and this is my interpretation
18:40 Azeem Azhar? rather than their conclusion, is that firms who are successful in automation are successful partly because of the automation, but also partly because they have better leadership. And so in traditional competitive dynamics, we'll outperform firms that have have worse leadership. And I think this is a little bit relevant to what we might be seeing as we broaden the lens from Erik Brynjolfsson and his collaborators' very tight and very, very well evidenced paper to asking how this might ultimately play out. Now, there are obviously some really clear challenges. You know, history rhymes, it doesn't repeat. And so let's talk about, I think, one of the hardest clear challenge, which is that
19:23 Azeem Azhar? the technology is improving rapidly, rapidly, rapidly, and maybe rapidly and rapidly-est. And you think about a deep search, deep research query you might put to Claude or ChatGPT, that's 20 minutes of work for the AI agent. It's sort of dozens of hours of work for a human. And it can generally complete that work unattended. And that 20 minutes is going to turn into 200 minutes and then 2,000 minutes. And you can parallelize the agents to go off and do that work. And it will go beyond deep research, which is feels like it's increasingly a solved problem, into other domains. And so that, I think, is a differentiator.
20:05 Azeem Azhar? And this technology compared to previous automation waves and things like electricity, which we go back to because we have to think about what happens in workforce where you can go off and get that level of work done. Most employees can manage to think about a three to four hour task. But when you start to think about three to four month tasks or three to four year tasks, that's where you get involved in senior management ultimately and and leadership. Today, the AIs can't can't be led in that way, but at some point, perhaps they might be. So that adds a sort of a degree of uncertainty. So I know we've got a question coming in about universities lagging behind providing a useful state of the art education. And I'm going to I'm
20:50 Azeem Azhar? going to touch on that because I think that gets to this key idea that actually what matters is policy solutions to all of this, right? What's very helpful with a paper like this and some other papers that have come out that have come out recently is that it gives us an indication as to what we are diagnosing and therefore what is the policy that is required. There is a trap here. There is a market failure. Firms have shown to some extent they're disinclined to hire graduates and train them and give them that tacit knowledge that makes them useful. Why? Why are they disinclined? Maybe it's an investment they can't afford. Maybe they they'll make that investment
21:33 Azeem Azhar? and then they're worried that those workers will will leave. But it's just too risky for them. So there is a kind of policy market failure here that will probably require interventions that are both carrots and sticks. So, you know, hire people, train people, get some kind of a tax credit is the type of thing that you might, you know, be thinking about. What is the education that workers need to come out with? They will clearly need earlier access to tacit knowledge, whether that comes through internships or apprenticeships. Whether that comes through formal university education. I have faith that universities can teach difficult, complex, structured and unstructured things.
22:17 Azeem Azhar? I mean, they they have done for for centuries. So I mean, there's no reason to believe they can't teach elements of these tacit skills. And I think there could be an issue for this cohort, which is that they may have graduated before ChatGPT was everywhere. They may not be as AI native. And therefore, as they go into the job market now, where employers are expecting that. They're falling down both on AI nativity and on tacit knowledge. Again, to be clear, I'm making an assumption about the AI nativity as it was. So it does feel that we will need a different mechanism to train workers. Whether it's universities or something else, of course, not everyone goes to university. The costs of university are very, very high.
23:02 Azeem Azhar? And I can imagine that splitting in in a couple of dimensions. So dimension number one would be really understanding that that tacit knowledge. And the challenge of tacit knowledge is that it's often not codified. So, you know, can you go out and codify that knowledge in some useful way? That allows you to bring people up to speed. And I've invested in a company, Sugarwork it's called, CEO Vanessa Liu, that attempts to do that. To identify and crystallize a tacit knowledge within within companies. And and and then if you can do that, can you then have that be that something that young people can get? Either through university or through their internships.
23:42 Azeem Azhar? I I think the second thing is how good are people really at using the AI tools? And using them to be able to develop and strengthen the business case for for hiring them. And again, this early cohort probably hasn't had a chance to formally learn it. And what are those skills anyway? You know, is it just about prompting or is it about meta prompting? Is it about understanding critical thinking and and analysis and more formal analysis? So that you can ask the right questions. It's probably a mix of all of those. I think there's another, you know, difficult pattern to all of this, which is that if you think about internships? And how do you gather tacit knowledge and experience in the world? Well, there we run into a a real, you know, equity.
24:26 Azeem Azhar? Problem, which is that ultimately people who come from particular backgrounds where perhaps there's space in the family home to discuss business will in the background implicitly develop that knowledge. They may be the ones who can afford to take unpaid internships. And so it raises another thorny question about how this is going to bifurcate across the workforce and it, you know, with a different cut. So it was a great paper. It's made me think a great deal. I've spent a lot of time, yeah, thinking it through and talking to people in California when I was there earlier this week. That is all we have time for today. Thank you so much for showing up. I'd encourage you to read the paper. We will put it out, of course, with some more analysis
25:07 Azeem Azhar? across Exponential View over the coming weeks. But in the meantime, I wish you all a wonderful weekend. Thanks for listening all the way to the end. If you want to know when the next conversation is released, just hit subscribe wherever you're listening. That's all for now and I'll catch you next time.
Previous AI & the future of media with The Atlantic CEO, Nicholas Thompson Next My OpenAI thought experiment: $500B to $1.5T? Page updated November 1, 2026 · transcript and summary are machine-generated