AI Exchanges: AI’s Impact on Employment

0:05Welcome to Goldman Sachs Exchanges. I'm Allison Nathan and I'm here with George Lee, co-head of the Goldman Sachs Global Institute. Together we're co-hosting a series of episodes exploring the rise of AI and everything it could mean for companies, investors, and economies.

0:22Allison Nathan:George, great to see you again. You too, Alison. Thank you. So George, today we're discussing one of the biggest anxieties, I would say, about the rise of AI, which is the impact on jobs. I would say even prior to the amazing advances in the technology that we've seen over the last few years and that you and I have discussed a lot on this podcast, there have been a lot of questions about the extent to which AI will ultimately replace workers. And if that happens to a large extent, what industries, what roles will be most impacted? It's now been almost three years, though, since we've seen the launch of ChatGPT, which is very hard to believe.

0:59Time flies when you're having fun. But that means that we're actually beginning to see some hard data that I think could finally reveal some questions, or at least start to reveal some answers to these questions. So to that end, I'm excited to welcome my colleague from Goldman Investment Research, Joseph Briggs, who leads our global economics research. He's done some really, truly informative and formative work, I would say, on the economic implications of AI more broadly and certainly on this topic. Joseph, thanks for joining us. Great to be here, Alison. Great to be here, George. And so, Joseph, before we just dive into all the details, as I just said, we've talked a lot about the advances of the technology.

1:40It's come a long way in the last few years, but I think we'd all agree that the adoption is still pretty early. So is it too early to see tangible signs on the labor market? Yeah. So when we look at adoption, we're currently tracking about 9 % of companies in the U.S. using AI for regular production. Now, this number is probably a little bit lower than some of the more eye-popping adoption rates that get picked up in the media. But the definition that we're using when we say 9 % is regular production for goods and services over the last two weeks. And I think this is the right definition to keep in mind because it's really what is going to be necessary.

2:19Companies using AI and regular production to drive significant productivity impacts. Now, given that adoption rates are only at 9%, it's not too surprising that we haven't seen a large impact or really any sort of meaningful impact in the overall labor market data yet. We have seen adoption pick up in sectors that are more exposed to AI. There's actually a very positive correlation between the exposure scores we constructed over two years ago and who's using AI today. But if we relate those to labor market slack indicators, things like unemployment rates, job finding rates, layoff rates, average hourly earnings, hours worked per week, there's really no sort of meaningful correlation.

3:00And so when I look at the impact that AI has had on the overall labor market data so far, it looks pretty small to me. Fascinating.

3:08Allison Nathan:I was struck by the 9 % number. It feels lower than my intuition. I believe actually that number though, if you weight it by employment, it's slightly higher. Is that right? Yes, absolutely. If we look at large companies, you know, those with more than 250 workers, these are those that have the in-house technological expertise that can really develop the AI tools on their own. Adoption rates there have gotten up, into the mid to high teens. It's small companies that are really waiting for the plug and play solutions that we haven't seen adoption pick up in any meaningful way yet. That's great.

3:42And if you think about industries, what industries have been most affected thus far? Definitely the tech sector. One of the things that we flagged in our most recent report is that if you look at the tech sector's employment trends, they'd been basically growing as a share of overall employment in a remarkably linear manner for the last 20 years. Over the last three years, we've actually seen a pullback in tech hiring that has led it to undershoot its trend. And so, you know, this is telling us that in the tech sector, which I think is the one that has gotten the most attention in terms of leading the way for AI adoption, there has been, you know, some meaningful headwinds to hiring and job growth.

4:19There's other sectors as well. You know, finance is an area that is showing pickup and adoption rates. Education is showing pickup and adoption rates, business services more broadly, anything exposed to content generation. But even within these sectors, the adoption rates are still relatively low. On the tech side in particular, while you're seeing some impact in terms of efficiency gains, are you not hiring a lot of AI engineers? I mean, does that show up at all in the numbers? Yeah. So we've definitely seen a pickup in job postings that are mentioning AI. I think they'd increased by 25 to 50 percent in our latest AI adoption tracker.

4:56This is relative to other job postings. And so companies are trying to hire workers that have the expertise to build out the capabilities and the tools to unlock the productivity gains that we think are possible. It's just that this is a very small share of the overall economy, a very small share of the overall labor market. We're still in the very early days and seeing it being distributed and the productivity and employment benefits of AI being distributed more broadly. And there's also been a lot of discussion about the actual type of role that's being impacted here. So are you seeing some evidence in the data?

5:29The more junior roles, I would imagine, being impacted? What are you seeing? Yeah, there's been a lot of questions around the lag in hiring rates or the difficulties facing recent college graduates. I'm sure that we all know people who have had trouble finding jobs or a harder time than they would have normally following their recent graduations. This is validated in the data. We're definitely seeing a lower hiring rate for recent college grads. A lot of this is just related to the fact that the labor market has shifted back to a low hiring, low firing labor market. We're seeing this very broadly across sectors, across different industries.

6:06And so, you know, I think that the anecdotes and the relationship that the anecdotes have to AI is often a little bit overstated. That being said, if we do look at unemployment rates in the tech sector for young workers, and so those between ages 20 and 30, they have increased by about three percentage points. And this is over since the start of the year. And this is a much larger increase than we've seen in the tech sector more broadly or a larger increase than we've seen for other young workers. And so, again, the story is one where the overall impacts on young workers in the labor market, speaking from an aggregate perspective, is small.

6:39But if we start zeroing in and zooming in on these specific industries where we are seeing AI be used to drive efficiency gains, there are signs that headwinds are emerging there.

6:50Allison Nathan:Yeah. So I would say I think it's a real reflection of CEO uncertainty around this entire phenomenon, which is to say the intuition is that these tools ought to create enormous productivity and efficiency in the enterprise. Yet, as Joseph's statistics suggest, the macro effects aren't fully being seen. And so what can you control as a CEO that seems lower risk? You can lower your intake hiring and kind of adopt a little bit of a flat as the new up perspective as it relates to your headcount. That feels like a more prudent move than beginning to aggressively harvest more senior professionals, et cetera.

7:27Allison Nathan:And so I think, again, it's a little bit of a temporal phenomenon, which is to say, I think this is going to be really meaningful. How do I begin to streamline my enterprise so I can be more flexible, more adaptive? and do it yet without harming our competitive edge. And unfortunately, I think young employees for this period of time are a little bit the casualty of that. And George, more broadly, you speak to a lot of companies. And does what Joseph is basically observing in the data really reflect in practice in the conversations you're having? Very much so. And, you know, I think I would echo one thing that Joseph said, which is technology industry is, I think, the most profoundly affected in the early days.

8:06Allison Nathan:And there are sort of two theories of the case there. One would be, well, that's natural because the place where the models have the most capability and utility are in software development. And that's a huge part of what technology companies do every day. So no wonder displacement's beginning to hit there first. The other theory of the case is that because these companies are on the leading edge of developing and deploying these tools, they're the canary in the coal mine for what's going to occur in other industries. And I think the evidence is, Joseph, is so great about presenting evidence, data, and being balanced about these things.

8:39Allison Nathan:I don't know that there's a determinative answer to that question, but I do think those are the two forks in the road. And I think the biggest question that we have is ultimately, will this lead to a net decline in employment, in job availability? And just one of the striking stats that you came up with was that we will see, you call it transitional displacement on the order of 6 % to 7 % off the, which is, I think, a big number. But I hear you're getting pushback that that's too small. It's interesting. Our AI productivity growth forecasts have always assumed a 6 % to 7 % displacement rate. And after having gone through the exercise over the last couple weeks of revisiting that and cranking all the numbers to try to see, is that still the right estimate?

9:22That's broadly in the ballpark of where we came out in terms of the overall displacement rate that will happen following full adoption of AI. When thinking about how much AI is going to translate to an increase in unemployment, which is kind of the flip side of a decline in employment, it's useful to break it down into two types of unemployment. The first is the more concerning long-run technological persistent unemployment. I'm much less concerned about this. If we look back historically, technology has always added new positions. 85 % of job growth over the last 85 years has been driven by technology.

9:56I think that this is a trend that will reassert itself once we've seen the rise in aggregate incomes and all the new opportunities that AI creates. What I think is more realistic is that we are going to see a period of frictional or transitional unemployment. where it does take time for these 6 % to 7 % of workers that do lose their jobs because of automation have to find new positions in potentially new occupations. And this is a dynamic that we've seen play out historically, that any time we've seen, say, a one percentage point boost in labor productivity due to technology, the unemployment rate tends to rise by around 30 basis points, three-tenths over the next year.

10:36After two years, there's no effect. If we try to translate that 6 % a 7 % number to an increase in the unemployment rate in any given year, I go back to the adoption speed as a key variable to watch. And the reason that I say that is that if we're wrong and that AI adoption and all the displacement takes place over a one to three year period, then all of a sudden that 7 % displacement rate translates to a two to two and a half percent boost to the unemployment rate. That's a pretty big macroeconomic shock. It has significant impacts on spending, on GDP. On the other hand, if we're right, and the AI transition takes 10 to 15 years, then that 7 % displacement rate translates to something like a half point, maybe a little bit less boost to the unemployment rate.

11:24That seems very manageable and less disruptive from a labor market perspective.

11:28Allison Nathan:Yeah, I think at the heart of that question is your belief as to whether this is going to be a continuous function of adoption of technology and will take that longer period of time, or it has the dynamic of a tipping point where we reach some salient interval where these tools are mature enough and we have a very sharp increase. Any reflections on that question? I agree that if we see the application build out happen very quickly, and again, the application build out for a lot of companies is a necessary step to start using the technology, then we could be wrong and the unemployment rate increases could be larger.

12:05The other thing that I'd flag looking at historical data, which I thought was interesting. If we look at the automation of routine occupations, you're right that it hasn't happened in a very smooth manner. What we've actually seen is that during economic slowdowns or recessions, companies that are forward-looking in nature and they're looking to trim labor costs, they often target those routine occupations that they're expecting to automate over the next several years anyway. And these are the areas where you see employment reductions. And so one of the big concerns that I have in my mind when thinking about whether or not AI could have a more disruptive impact on the labor market is that if we do see an economic slowdown in the next one, two, three, four years, then at that point, a lot of the automation and labor displacement that could eventually occur and that we are expecting will occur in a relatively smooth manner, it could happen in a more narrowly concentrated period.

12:57Allison Nathan:So Joseph, you do some great work in terms of identifying jobs that may be more vulnerable to displacement. What about those jobs that are more resilient, less exposed? What are some of those job functions? Yeah. Given that we're in the very early days of the AI transition, it's hard to have a lot of confidence when we're looking across different types of jobs. Where are we going to see more displacement and replacement? And where are we going to see just AI unlocking productivity gains that makes people more efficient? I do think there's a couple of proxies that we looked at that provide a signal or provide an indication that the risk of displacement is lower.

13:37So things like occupations that are more exposed to human interaction. A lot of the commentary that we've heard from corporates flags that back office work is more likely to be automated in the near term, whereas front office work is more likely to sustain. Also jobs that have higher stakes of decisions, and so we're making a mistake, could expose a company to more reputational risk or monetary risk. And then occupations where the type of tasks that people do are less repetitive, more diverse, and the potentially automatable tasks are lower value add than a worker's core function. And so when we run, we constructed risk measures for all these, when we run 800 plus occupations through the different risk filters, things that stood out as being potentially less exposed to automation were medical care providers, pharmacists, door-to-door salespeople, teachers, clergy members, CEOs.

14:30There's a lot of, you know, different drivers for each of these that leads them to be less exposed. But the key things to bear in mind are those jobs that are less repetitive, where decision consequences and stakes are higher, I think are less likely to be replaced in the near term. And it's striking the variability you just get there, you know, teachers and clergy and CEOs. I mean, there's a wide spectrum that should be more resilient. It's hard for AI to go door to door and sell products, just like it's hard for AI to run a company.

14:56Allison Nathan:I think one of the interesting things about this discussion is the differential between micro and macro. When you talk to CEOs, when you're out in the field, this is an assertive mode obsessively focused on topic. And yet, as Joseph's work suggests, it's not finding its way into the macro statistics yet. And perhaps that's always the case where there's some fundamental shift. But I'm struck by that. If you weighted the percentage of time CEOs spend thinking and talking about this issue relative to the discernible effects in the macro, it's a very strong disconnect. So interesting. I have a broader question for you, Joseph, that you touched on, which is you're looking at history when you make a lot of your forecasts and assessments of what lies ahead.

15:41But if we think about just how potentially transformative AI is relative to even past technological innovations, is history likely to be a guide? How confident can you be in that? The biggest caveat, and I should always add this when we're talking about AI, is that our analysis doesn't factor in the potential for emergence of AGI. You know, if we do see not automation or AI not only driving automation, but leading to an acceleration in the pace of innovation and an expansion in the frontier of human capabilities, then in that world, the boost of productivity would be much larger. It's hard to even start thinking about the impact on the labor market, but I would guess there probably and undoubtedly is more room for labor substitution and a more disruptive impact in that world.

16:27There's been a lot of attention on that recently. I don't personally have a strong view on how close we are to AGI. George, you might actually be more connected into those discussions than I am, But I think that is the one potential tail scenario. How realistic of a tail scenario it is, I don't know, that could lead to much more significant impacts on the labor market than we're factoring in. It's a great point.

16:48Allison Nathan:And look, this question, first of all, AGI is a very complex definitional matter. Everyone, I think, sees it a little bit differently. In some ways, I think it's a bit of a canard in the sense that it is being positioned as this momentary shift in the world. I'm a believer this is much more of a continuous function. And that by the time we get to whatever represents AGI or ASI or whatever, it will have seemed, again, a continuous process rather than a moment in time. But only time will tell. And as Jeff has said, there are people smarter than he and I in the world musing on these topics. To the point of history, that's when I've gone back and done a little bit of work on things like very fundamental, emergence of very fundamental technologies like telephony and electricity.

17:33Allison Nathan:And those may be, you know, in the most bullish scenario, might be a more interesting analog than, for instance, the emergence of the Internet or cloud computing. And, you know, the impact of those was extremely vast. And yet, to Joseph's work, took a fair amount of time to really make themselves felt in the macroeconomic picture. So you do think history can serve as a guide, but maybe not the history that's really front and center on most people's minds. Certainly, yes, in our recent, not in our recent memory, perhaps. I don't know, Joseph, any reflections on that? Those kind of fundamental, very deeply historical shifts?

18:07Yeah, I think that the way that I would frame this is the emergence of general purpose technologies. And we haven't seen a lot of general purpose technologies emerge. And I think that electricity and the electrification of manufacturing the U.S. in the early 1900s is probably the best analog. The IT revolution and adoption of software, the Internet, that's the other one that we often benchmark to. But I'm very sympathetic to the idea that given that we only have a few data points of these types of technological shifts historically, we do have to be fairly humble in our ability to use this evidence to extrapolate forward.

18:46I think that is very wise counsel.

18:49Allison Nathan:And you mentioned a thing I think that is also really important to have in this dialogue, which is the general purpose nature of this technology. And that's both a feature and a bug. The feature is the breadth of applications are limited really only by human imagination. The bug is that there's no user manual. There's no trodden path to follow. And it'll take time for all of us to come up with the best ways to leverage this fundamental new capability to drive value. And that's one of the things that extends timelines and makes Joseph's perspective on this so valuable. As always, lots of food for thought.

19:27Thank you so much for joining us, Joseph. Thank you for having me, Alison and George. Great. George, as we sit here and discuss this with Joseph, I am reminded of our recent conversation with Marco Argente, who was really bullish on the prospect of even a hybrid workplace at some point in the near future. In light of the comments we just got from Joseph, do you have any additional thoughts on that?

19:49Allison Nathan:Yeah, I think Marco's perspective, first of all, he's a very keen observer of all this and a very broad thinker about it. And so I thought it was a provocative and interesting perspective. Anchoring to a lot of the commentary we've had here, you know, the rise of agents, their utility in the enterprise, we have to climb a hill of maturity there before we see those really emerge as being effective. And in a way, that's a trailing phenomenon to the rise of generative AI broadly. The impact of that as it matures could very much be in the model that Marco envisioned, which is to say managers have human employees.

20:24Allison Nathan:They're also responsible for a set of agents that are performing work and tasks that humans might have otherwise done. And that hybrid management challenge is going to be new, different, interesting, and possibly an enormous productivity driver in its own right. And I think that a new wrinkle that Joseph's research revealed that we discussed was, you know, if more junior roles are being taken in the AI space, as you said, the striking disparity in the rise in unemployment in junior roles in the tech sector relative to the rest of the economy. What does that mean for managers? You need to have people in junior roles to become senior.

21:05What's that going to mean for management ahead?

21:07Allison Nathan:A fundamental challenge. How will the apprenticeship that creates the next generations of Allison Nathans and Joseph Briggs, how will that emerge in a world where there are potentially fewer junior employees in enterprises? Again, I don't think we should say that that's the certain outcome, but I think it's really a fundamental question. Now, on the other hand, if there are fewer junior people in enterprises, there's the potential that their experiences in those enterprises are more high value, less weighted down by low value tasks, more deeply connected to senior managers, and that the quality of the apprenticeship rather than the quantity may lead us to breed even better senior leaders.

21:51Allison Nathan:Hopeful lens. Interesting point. Thanks very much, George. I always enjoy these conversations. Same. And thank you again, Joseph. This episode of Exchanges was recorded on July 31st, 2025. I'm Alison Nathan.