Ernie Tedeschi: What the Productivity Numbers Say About AI

Ben Zweig: We can see a productivity surge in the data. And my guest today goes through all the potential drivers of which there are a lot. Ernie Tedeschi is the Chief Economist at Stripe and former Chief Economist at the Council of Economic Advisors. He's in a unique position to speak on the recent changes in productivity and why that's related to every facet of our lives, not just new technologies. Here's our talk. Okay, welcome to the Economics of Work. I'm excited to welcome Ernie Tedeschi, Chief Economist at Stripe, former Chief Economist at the Council of Economic Advisors and former Head of Economics at Yale Budget Lab. And generally someone who brings a lot of analytical rigor to big important macro questions. Ernie, welcome to the show.

Ernie Tedeschi: Thanks for having me, Ben.

00:38 Can we see AI in the productivity statistics?

Ben Zweig: Okay, so big question. We can see AI everywhere. Can we see it in the productivity statistics?

Ernie Tedeschi: Yeah, it's a great question. So we've seen a surge in productivity in the United States. I say surge, like if before 2023, we were running at maybe one and a half percent annual growth, now it seems like we're running more like two and a half percent annual growth. So an extra point or so in productivity growth, which like, if you're a macroeconomist listening, you know that's a major upside to productivity.

Ben Zweig: I mean, if we think about the macro implications of that for deficits, for whatever for like everything we wanna do, that gets us very far.

Ernie Tedeschi: That's right. Even if it doesn't solve the problem entirely, look, an extra percentage point of productivity growth, if it were sustained over time, would help so much with the challenges that we have on the fiscal side, it would boost incomes over time, it would be great. And that's not implausible, by the way, that we get that for some amount of time. Here's what's interesting though, is that the story that we tell ourselves about how AI is going to affect productivity is a bottoms up story. AI is going to boost task level productivity. I'm gonna be able to write this memo, generate this image, write the code 20, 30, 40% faster. And then you, I mean, think about the aggregation of tasks like put together all of those tasks that are now much more productive, put together all the workers, right? That should sum up to stronger productivity growth. That's the part that we don't see in the statistics. And here I wanna emphasize, Ben, we need to put a little bit more of a grain of salt with what I'm about to say than the earlier part about labor productivity. But measures that try to get at like the pure technological part of productivity growth, what economists call total factor productivity, are actually quite weak right now. In fact, one of the most widely followed and cited ones from the San Francisco Federal Reserve is growing at basically 0%. And that's completely inconsistent with a like task level, worker level, AI productivity boom happening right now. Again, we have to, you can't directly observe that part. You have to model it. So like we need to keep an open mind about where things might evolve. But that's what we're seeing in the data right now. So the way to resolve those two things, like strong vanilla productivity growth, but really weak total factor productivity growth is I think two things. One is that the challenge with AI productivity right now is not at the task level. I think it's been pretty well proven that way.

03:36 How total factor productivity is measured

Ben Zweig: Can I actually just, can we actually take a second, just like break down how we would go about those measurements just to see if we buy the disconnect? Because TFP is a residual term, right? We see growing TFP if we have productivity that is not explained by increases in the labor stock and capital stock. If we see like increased returns to those that we can't explain by the stock of labor and capital. And it's weird to think that TFP is not moving while labor productivity is moving because labor supply, we don't see labor deepening even though it's not really a word, but we don't see increases in labor supply. Like demographics are really bad. We have an aging population, there are reductions to immigration, like no labor deepening. Do we see a lot of capital deepening? Like, I don't know, like, is that really part of the story? Like, I don't know, it doesn't sound so consistent as a narrative. So like, I guess I wonder how we're getting at these TFP numbers. And I know like there are different models and people get at them in different ways, but like, do we think that this might just be a measurement issue?

Ernie Tedeschi: It's very possible it's a measurement issue. I wanna be very open to that possibility. Look, productivity just at a basic level, right? Is a ratio of two things that are very noisy and very hard to measure. You have overall output for the economy as a whole or by an industry. That's actually very challenging to measure, more challenging than I think I appreciated before I became a macroeconomist. And then you have hours of work, which sounds straightforward, but like you start thinking about it. And it's actually, that's also really hard to measure. How many hours did you actually work this week? Like, are you just saying 40 because you're on salary? What if you work at home? Are you counting commute time and all of that? It's actually a really hard thing to get at in a consistent way, even within a country. And then you start getting into these other conversations about like global productivity comparisons and all bets are off at that point. Like everybody measures it differently. So you start with like noise over noise.

Ben Zweig: Right, and it's a rate of change in those ratios.

Ernie Tedeschi: And it's a rate of change. 100%. And that's just the basic measurement. And then we're getting into like the theoretical part that you touched on, right? Which is that we know that as we add more workers or we have more skilled workers, more educated workers, higher quality of workers, right? That's just mechanically going to raise economic growth and therefore productivity growth over time. And that's a real thing, that's good. But we want to get out a measure, right? That filters through that part as we invest, as we add capital, capital deepening, which you said. That's also a legitimate way to grow an economy and to reap productivity dividends. But like with TFP, we really want to filter that out too. We want to get at like the nub of like the technological impact of it. And those things are, as I said, very, very challenging to filter out. I will say though that we're clearly in a world with very strong capital expenditures and investment in AI. So to me, the idea that once you filter out that part, the technological residual is going to be much smaller than may be indicated by overall productivity growth is not completely implausible to me. But even there, like even something as basic as, well, how much is a data center worth, right? Like how do we adjust the price of a data center? Because remember, we're trying to do all of this in inflation adjusted terms so that it's comparable over time. Well, like trying to inflation adjust a data center is crazy given how quickly prices are changing and like what's the right way to capture technology. We want to quality adjust for a data center over time. How the hell do you do that? You know what it reminds me of Ben? It reminds me of like a completely different problem. But when I worked for a financial firm years ago, before I joined the White House, somebody asked me, oh, we need data on like US defense spending going back to 1790 or something. Which you could actually get. And then they were like, oh, and it needs to be inflation adjusted. And I talked to the guy and I'm like, what does it mean to inflation adjust a like hundred gun ship of the line from 1790 versus like an Aegis cruiser today? Like how do you even go about doing that? And now think about that same basic challenge, but like over the course of like weeks or months rather than over the course of centuries. Like how do you compare the value or the price of a data center or a piece of software or a computer component today versus a year ago? Like it's so fast moving and so hard to pin down.

Ben Zweig: Yeah, it's a great point. Yeah, I really, I really haven't really thought about like how people adjust for quality of like different like types of hardware and whatever and like it's, yeah, it sounds like a beast. I mean, I'm sure people are doing it and like doing their best and probably doing a reasonably good job. But like, it sounds very daunting.

Ernie Tedeschi: It's a, so it is a very hard problem. Luckily we have very smart people thinking about it at places like census, the Bureau of Labor Statistics, Bureau of Economic Analysis. And so like, and just shout out to our civil service economists who think about hard problems like this all the time.

09:49 Capital utilization and the AI build-out

Ben Zweig: Okay, so then from this breakdown, we see that it's primarily the aggregate productivity increases are primarily driven by what? By the data center build out, by capital deepening or by, yeah, how do you make sense of it?

Ernie Tedeschi: So I think that there's like a mechanical piece and then there's the like, let's take a step back and let's explain part of what's going on here. When you look at the San Francisco feds measure of total factor productivity, they have a component that they call utilization. And what that really means is existing businesses are using their existing capital stock more than they did before. Like they're utilizing it more. What that really means is I have a GPU. I'm pushing my GPU cycles greater than I did before. I'm using my existing factory more than I did before. I'm not that part of it is not hiring more workers. It's not building more factories or GPUs. It's using what you have more. Again, that is a legitimate sort of driver of productivity growth. And normally it would show up in total factor productivity because it is not adding more capital, adding more labor, but it's still a little bit different than what we're really trying to get at with TFP, which is pure technological change. And so what the San Francisco fed measure suggests is that what we've seen over the last couple of years is greater utilization, which I think just as a bedtime story we tell ourselves makes sense. If we're in a world of greater AI demand and we're falling over ourselves to build more data centers secure supply chains overseas to get sort of the memory that we need, et cetera, it makes sense that a sort of consequence of that or something that would be correlated with that is, okay, to meet demand, we need to use what we have more, right, as a way to try to like address all of this additional demand that we have. So we can try to at least expand capacity. And so I think that's entirely consistent with my sense of what's going on in the economy right now.

Ben Zweig: Yeah, that's so interesting. I mean, when you say like capital utilization, I'll tell you where my mind immediately went to. Yeah. My first instinct was like, oh, return to office. That like you have all these buildings that were like unutilized during COVID and now like return to office and you have like this very expensive capital that is being utilized, but that would show up earlier than like GPU utilization. Right. What does the timing look like?

Ernie Tedeschi: Yeah, so the timing looks, so utilization wasn't as much of a factor in say, 2023, 2024. But I think beginning in late 2024, 2025, that's when again, using this measure, we start seeing that become much more of a factor. So I don't, so if you think of the AI arc as starting with say, ChatGPT, late 2022.

Ben Zweig: Right, November, 2022, the big--

Ernie Tedeschi: There you go, that's the date that everyone throws out.

Ben Zweig: Yeah.

Ernie Tedeschi: Utilization is not a consistent story across that entire arc. But if you're like me and you think of really the AI revolution as one that's really about agentic and coding tools that really, that are more, like ChatGPT in 2022 was incredible, but it wasn't, in my opinion, it wasn't displacing anyone's job. AI agents over the last year really are in a position where they are substituting for some real human work with enough reliability, right? That at least now that conversation is plausible. And that's really been only in the last year or so. I would say it's a 2025 story. And that's when you really start seeing utilization take off as a major factor in all of this.

Ben Zweig: Yeah, interesting. I mean, the framing around utilization is like, I mean, I think that sounds very plausible. I also think that like one thing I think of with like high utilization rates is that, that tells us a little bit of like a supply chain constraint story where like, we just don't have enough of what we have. And maybe that is like when utilization rates are really high, that's an invitation to invest more in capital because you're putting your capital to good use. So maybe that is like the precursor to capital deepening. Yeah, no, that's a very, it's a very interesting framing. Love it.

14:28 The bottleneck between tasks and revenue

Ernie Tedeschi: Yeah, so I think like mechanically, that's part of what's going on right now. I think the other thing, and this is harder to get at in the data, but there is micro rather than macro oriented research that is beginning to show this. I think the other thing going on is like AI has already, like AI 1.0 is very focused on task level productivity gains. But you have this like outer ring of frictions that still remain. Like how do you convert 40% faster writing or image generation or coding into revenues? Well all of those products still have to be reviewed by human managers. You have to strategize, you have to you have to align it with your business plan in terms of what you offer to users. And I think that's where a lot of the like inner ring productivity gains are getting held up and where the bottleneck is right now. And that's okay, that's true of every technology that's ever come about in human history. Like look at the automobile, like you can invent the automobile, but like there are bottlenecks to how it enhances economic activity and productivity growth if you don't have the roads, the gas stations, if you haven't built like trucks rather than rail cars as a way to transport, right? And truly unlocked all of the productivity gains that you can get from the technology. That just takes time. And so my hope is, my thinking is that AI 2.0 will be, will start being more focused on that outer ring of frictions and bottlenecks. And that's when we're gonna start seeing like truly task level productivity gains being converted into something that we can see at an aggregate level on a more sustained basis.

Ben Zweig: Right, I mean, in some way, like I think, I think of this kind of mantra that like technological innovation has to be, has to be paired with business process innovation to get productivity. And the business process innovation is about like this sort of task chaining process building workflows around some technological deployments that get paired with human deployments. But also, it makes me a little skeptical of this like, bitter lesson idea from Richard Sutton, that basically like the gains from AI come from just whoever has the most advanced general intelligence, rather than like domain specific applications of AI, because the domain specific applications are going to be the ones that like can address the nuts and bolts of these bottlenecks.

Ernie Tedeschi: So-- I agree with that. Yeah, and I think that there are some areas with high return proprietary research, pharmaceuticals is an obvious example, where I think just like the business model has, it will have to converge to domain specific AI models, right, that the company themselves will own and train, rather than the ChatGPT. And then it's not just that like pharmaceuticals are a high value industry, it's that like, they are, it's an industry that is based on human care, right? And so they're just, they're ethical and legal and regulatory constraints involving that industry for good reasons, that mean that like you need to, you need a tool that is like slight, that is more bespoke to those particular constraints, rather than just a general purpose using Claude or Codex to run your clinical trials. And that's good. So I think that like that, there are domains where that's constantly going to exist. I think it'll be interesting to see, I think the next interesting challenge with general purpose models, right, is what happens with the competition and the cost of tokens, the price of tokens over time.

Ben Zweig: Open source Chinese models will just eat everyone's lunch.

Ernie Tedeschi: Yeah, that's definitely a possibility. I feel like again, I think the, I think what people miss by focusing so much on November, 2022, as the beginning of all of this is that really adoption took off at least in the private sector. And this is partly from the experience of looking from the inside beginning mid 2025. And there was just, there was like this surge in adoption and usage of agentic AI tools for all of these serious high value tasks. And now suddenly businesses are having to grapple with the operating expenditure consequences of that. And they will, like everybody's having that challenge, but it was that was a stark shift in things, right? And I feel like there has to be then a dynamic, there needs to be another shoe dropping from that. Or there needs to be a dynamic where there is competition because like, I guarantee you, there's a lot of, there's a lot of demand on the business side for more competitors in that space.

19:55 Ernie's forecast for AI and TFP

Ben Zweig: Yeah. So do you have a prediction of when, and by how much we'll start seeing AI affect TFP?

Ernie Tedeschi: Yeah. My prediction is, so my prediction is that if you look over the next 10 years, you would see TFP gains that will average between a quarter point and a half a point per year over that decade. That doesn't sound like a lot if you're not a macroeconomist, I guarantee you as a macroeconomist, that is a really good return from all of this. That will that's again, as I was saying earlier, that's not gonna like solve all of our social and fiscal and economic problems alone, but I think that will help immensely. I think if you look on a year on year basis, I would expect like the earlier years, 2027, 2028 to be much higher than that, maybe we might get like an extra point to TFP, point and a half, some of those years. And then I think it'll I think it'll kind of taper off and slow down as we get to the second half of the decade. I think that's a really good outcome. And I think it would be on par with what we saw over the 10 years of the dot com boom, which happened under a different fiscal context, but was such a big part of the balanced budgets and the sustainable debt trajectory that we had coming into 2001. So so I'm hopeful and I'm hopeful that I'm wrong on the downside and that we get something even higher than that. If I'm too optimistic, I think it will be because, and this is not like my prediction, but I think where I could be wrong is that AI ends up exacerbating the next recession more than we think that like there are separations. We have a downturn, which is not necessarily going to be an AI downturn, it could just be a run of the mill downturn. Workers get separated and then as companies rebuild, they end up substituting toward AI more than we were anticipating. I don't think that's what's going to happen, but that's certainly been a story that we've seen with individual occupations in the wake of other technological shocks. And so where I have anxieties is that could be a more aggregate general story that we see coming out of the next.

22:31 How recessions accelerate displacement

Ben Zweig: Interesting, so let me see if I understand that. So let's say you have some technology that has some usefulness within firms and in an economic boom, let's say you have just like a good regular good economy, you don't have displacement of roles, you have, let's say transformation within jobs. But in a downturn, you have unemployment, you have displacement of entire roles. So you have less like within occupation transformation and more between occupation transformation. Is that kind of the way you'd understand it?

Ernie Tedeschi: Yeah, we wrote a case study about this for Stripe Economics that you can read where we looked at travel agents in the wake of the dot com boom. And it's really interesting because like Travelocity comes out, all of these internet travel sites come out in the late 90s and you don't see an immediate reduction in the number of travel agents out there. Travel agent employment was actually quite constant in the late 90s, early 2000s. And then the bust hits in 2001. And that's when you start seeing like employment in travel agents decline. First of all, what's interesting is that post dot com boom, travel agent employment became much more cyclically sensitive, much more reactive to recessions than it had before the dot com boom.

Ben Zweig: Right, like it's a luxury. It's a luxury. Yeah. I would think like air travel or whoever uses travel agents is always like kind of a luxury. But the more premium of an experience it is, the more cyclical, like that makes sense that it would be more and more cyclical. And like now, if someone has a travel agent, it's like super bougie.

Ernie Tedeschi: Exactly, and that's exactly what happened over time is that like they upsold their services over time to a wealthier, more upscale clientele. That's kind of an interesting story in all of this, is that like the travel agents that remain are paid more because their services are literally worth more now, right? Like they are servicing a more upscale clientele that is willing to pay for that human touch. But I think, but bringing it back to the cyclical conversation, I think what's interesting about the travel agent case is that it wasn't like displacement with travel agents wasn't this linear smooth thing that happened after 2000. It happened in fits and starts with recessions where they would get laid off during recession and then they wouldn't get hired back when the boom then happened the recovery happened after the recession and then like repeat that over three cycles. We have the dot com boom, the dot com bust, the financial crisis, and then the pandemic. And you see that sort of like level shift happen in every single one such that like we only have about half of the travel agents that we did back in 2000 employed now, which I'm surprised that we even have that.

25:30 Sticky wages and creative destruction

Ben Zweig: Yeah, that still sounds pretty high. It's crazy. Yeah, this sounds almost like a sticky wage story where you have like declines, but then at some point the decreases in demand, there's some moments where those decreases in demand affect quantities more than prices. And those are during economic downturns. Do you think this is like kind of related to a sticky wage story where maybe there's like some secular decline in demand. Let's take out the cyclicality part of it if we can just put that aside. But there's sometimes where that like eats away at earnings and sometimes where it kind of shocks quantity. Do you think that is like during downturns, there's much more likely to be a quantity effect because wages can really only decline, real wages can really only decline at the rate of inflation. And like there's some just cap on like earnings erosion.

Ernie Tedeschi: Yeah, so that's plausible to me that ultimately what's happening here is a combination of a couple of things. One is wage rigidities, like you said, so wages can't adjust in the wake of these downturns. And then also just recessions bring with them creative destruction of firms. So firms will just evaporate entirely. And then I think what's happening is that as they come back after these shocks, these negative shocks and they rebuild as they invariably do, they rebuild themselves using the latest technology and with the business plan that they can do with the business plan that adapts to where consumers are at any given time. There are many things about that model that are good, that we don't wanna give up, right? That is one of the reasons the fact that America allows that to happen to the extent that it does is one of the reasons that we have strong productivity growth in this country because businesses rebuild with the latest technology. So there's a dynamism upside to all of that. The potential downside for the worker, but I'll have an asterisk to the downside in a second. The downside for the worker, right, is that there's not a job there waiting for you on the other side of the downturn. And so if you are a career travel agent, suddenly your firm has gone belly up and the new firm that comes in to replace it, right, just doesn't have an opportunity for you or is can hire younger, less experienced workers for cheaper and that meets their needs. The thing is though, and this is where I come back to being more on the optimistic side, is that if this pattern repeated itself sort of ad nauseam throughout the entire economy, we would expect aggregate employment to be lower over time. In fact, just like extend that logic through all of human history. If every technological shock was a long run displacement event, we would have no employment left, right, in humanity, just after thousands of years. Instead, when you look at employment in the United States and in other advanced economies, especially if you control for the aging of the population, right, which as the population gets older, just organically employment rates are gonna be lower than they were before. Once you control for that, we're probably at an employment rate that is at a non-World War II high in this country, right? Like you are more likely to be employed controlling for age, than probably at any non-War part in American history right now, completely the opposite of what we'd expect. So then it really becomes more of a short run long run story, right? Which is like short run disruption, which I wanna be clear is legitimately painful for workers and that policy makers and businesses really need to invest time in thinking about how to manage that and how to do what's like best for the workers to be able to guide them through that. Versus a long run story, which just to me history keeps telling us over and over again, tends to work tends to be okay in the long run. Yes, the automobile got rid of livery occupations and that was painful for those livery workers aggregate employment in the end was higher because the automobile created more jobs than it destroyed at the end of the day. So I am much less worried about the long run. I think that our policy focus and our economic conversation should be much more focused on the short run.

30:25 AI and new business formation

Ben Zweig: Yeah, yeah, interesting. I mean, I really, I'm very sympathetic to this short run argument that like during downturns, we have more creative destruction, we have destruction of whole occupations in some way we have like diminishment of some whole occupations, but then the new businesses form in a more modern way and they take advantage of those productivity benefits. So to the extent that productivity increases come between firms rather than within firms. So let's say we just have more productive firms and they displace the old legacy firms, then that seems like a very strong productivity story. And it makes me think of some research that you and your team did on new business formation. So I think one part of the story is like, oh, we get new business formation during downturns and we have some freed up labor and capital that can be deployed to more productive ends. But we also see maybe AI is a sort of secular new business formation accelerant that can, where we could get the benefits that we see during these, like during these cycles, but without needing to go through the pain of those cycles. What do you think? How do you square these stories?

Ernie Tedeschi: I agree with that. And I think that's actually one of the more underrated uncertainties about AI, which is what does this mean for the structure of a firm and for business formation over the next several years? And what I mean by that is like the pandemic really kicked this off. And then I think AI exacerbated it, where when you look at business formation at the onset of the pandemic, there was a surge in new businesses that were being formed in 2020. It makes sense for two reasons. One, you had people stuck at home. So now like you have no one to talk to, that business idea that you've been harboring forever, right, like is now like tugging at you and you have nothing better to do. And so you see the surge in new businesses and it's persistent. It wasn't just this like 2020 or 2021 story that then returned back to pre-pandemic levels. Business formation stayed persistently higher beyond that. There was also like policy wise in the United States, at least we had the Paycheck Protection Program loans, the PPP loans, right? And so you had to be a business to engage with those. And so like that was an incentive to actually create a new business. But those went away like within a couple of years after the pandemic and we still saw persistently higher business formation rates. And then AI starts.

Ben Zweig: Well, wait, so what's the theory on why they would persist? Cause it can't just be this backlog of good ideas. It can't be staying at home with PPP. So like, yeah, why do you think they persist? Is it just easier to hire people and grow a team?

Ernie Tedeschi: So I think it's a combination of one, you just had a lot more churn in the middle of the pandemic of people looking for new work, being able to go back and forth. But I really wanna emphasize the work from home aspect of it. Work from home and the shift to work from home and the acceptance as a norm of work from home, even now that many firms are trying to return to office, like work from home rates are just persistently higher of course, than they were before the pandemic is like probably the most underrated social change of the last decade. We are still grappling with the implications that it means for not just how firms are organized, but employment rates, wages, et cetera. It's a massive social transformation. And it's like one of only five, right? Like there were five social transformations over the last five years, right? That we're all happening at the same time, which by the way is one of the reasons why it's so hard to measure AI effects. Cause all of these other things are happening at the same time.

34:38 Five shocks that confound AI measurement

Ben Zweig: Sorry, let's go through them. So there's work from home, which is big. What else is on the list?

Ernie Tedeschi: So you had the, okay, so that's one. So you had the great resignation, the surge in job to job transfers that happened around 2022. And then you have the like bullwhip from that, which was like many companies probably over hired in 2022 and then started hiring and firing less.

Ben Zweig: Right, now we are at like record low hiring and separation rates.

Ernie Tedeschi: And obviously the story there is different for every company, but like I just, I always think about the layoffs from UPS as a helpful example here, which is that they probably hired up in 2022 thinking that the world was gonna be one that was much more goods oriented in terms of the consumer than it had been before. And then what happened after 2022 is that like as things returned to normal, the goods share of consumer spending returned back to, kind of where it had been before the pandemic, which even at CEA, when I was there, we weren't anticipating that it would renormalize. And so, okay, so like UPS probably over hired for that environment. And that's, so that explains that. So that's the third thing. You had the surge in inflation beginning in 2021, 2022, right? That then leads to a surge in interest rates. And that has all sorts of spillover effects in terms of wage expectations, wage demands, in terms of consumer spending, right? So now like capital, more expensive interest rate sensitive goods are more expensive. And then of course on housing, which itself has all sorts of spillover effects, right? Just think about something as simple as mobility. Like I've lived in this house since 2014 and I have a mortgage that I refinanced in 2019. I'm probably never gonna move out of this house now, right? Because I don't want that new mortgage.

Ben Zweig: Yeah, you got those golden handcuffs.

Ernie Tedeschi: Yeah, exactly. But like it has a real effect on mobility, which has an effect on dynamism as well. And then like the fifth one is add in all the changes to immigration that have happened. And both the surge in immigration and then the swing in immigration over the last couple of years. And that's had massive effects on the labor force and certain sectors. And like any single one of those shocks that I mentioned would be like worthy of its own chapter in an economics textbook. And we've had five of them before we even get to artificial intelligence and the rise of ChatGPT and then agentic over the last year or so.

Ben Zweig: This is why it's a good time to be a labor economist.

Ernie Tedeschi: I mean, it's a good time to be a labor economist. The joke I make is that like, AI is like the worst timed experiment of all time because like there are just all these confounding effects happening too. It's true.

Ben Zweig: So this is why people should listen to the Economics of Work podcast.

Ernie Tedeschi: Exactly, that's right. Yeah, make sense of it.

Ben Zweig: Okay, so that's all right, fine. So yeah, where were we?

38:02 What Stripe data says about new businesses

Ernie Tedeschi: So work from home, right? So that, yeah. So I think work from home is the context here for the rise in new business formation. Then we start seeing another surge in new business formation around 2024. I am convinced that this is an AI story. What's interesting about this latest surge, unlike the pandemic surge, this surge was among businesses that are unlikely to be hiring employees. So census tracks, based on your application, whether you're likely to end up having workers. Earlier in the pandemic, like we saw a surge in overall business applications and we saw an increase in what census calls high propensity business applications or applications where you're likely to hire an employee.

Ben Zweig: Like high growth startups versus a neighborhood laundromat type stuff?

Ernie Tedeschi: That's exactly right. Yes, that's a great way of thinking about it. We did not see that commensurate surge in high propensity business applications in 2024. They've been flat this entire time. Now I'll tell you that when I worked at CEA, back in 2021 to 2024, we would have observed that pattern and we would have said, oh, those are probably Potemkin businesses or like there's fraud going on or something sketchy is going on here. We should not treat this as real economic activity happening in these businesses. That's where Stripe data helped us out on the Stripe side where we can actually look at the surge in signups on Stripe. And by the way, we see a surge in signups on Stripe that follows the qualitative pattern that we see in the official business application data. And then we can look at things like revenue thresholds for those businesses. And we can ask questions like, of the businesses that started on Stripe in 2025, are they meeting certain revenue thresholds earlier or later than prior cohorts? And the answer turns out to be they're meeting them earlier and in some cases a lot earlier. Like the businesses that were started in 2025 met, reached a million dollars in revenue, 30% faster than just the 2023 cohort did. You can't explain that through things like inflation or other adjustments. And twice as fast as the 2019 cohort did. So that tells me now, okay, like our hypothesis that this was non-business activity driving this increase in business applications, that's not what we see in the Stripe data. We see the exact opposite, that these are meaningful businesses doing meaningful business activity faster than prior businesses did. So then the question is, what is driving this? We have some visibility in the Stripe data for signups that use certain AI tools, things like sandboxes, command line interfaces, et cetera. It's not comprehensive, but we do have some signal. That share has definitely increased substantially over the last couple of years. We also know from other Stripe products, we have a product called Atlas that lets you incorporate as a Delaware corporation specifically. And that collects a lot more information about your business plan. We know that AI native businesses are driving half of the increase in Atlas driven Delaware incorporations over the last couple of years. Okay, I think that this is even more than just AI native businesses. This is not just a story about like, I am a company that is building an AI tool. I think that this is people using AI tools almost as a co-founder, right? To help them start up their business. And so, whereas you might really want a co-founder or a network of people in an industry that you're interested in for your business before you sort of are comfortable enough to take the leap and actually do it. Now there's a built-in resource where you can ask things like, all right, which tax authorities do I register with? Like, what's a good way of approaching a business plan? How should I think about my market? All of those have now fairly competent answers that you can get from AI. And I think that is helping people make the decision to start a business.

Ben Zweig: Yeah, that makes so much sense. I mean, I started Revelio Labs in 2018. And like, at the time, I mean, I'm sure it's still this way, but like starting a business, you have to do a lot of administrative nonsense. Like really just like, doing your like equity stuff and like registering and doing trademarks and all this nonsense that like no founder really has any expertise in. And then you have to be in this startup ecosystem and like make these like silly little business plans and PowerPoint decks and financial models that like you're never gonna use. And then sometimes you wanna like prototype something and do little toy demos that are really not a good use of time. And it would be very nice to just like vibe code your way past the nonsense. So you could focus on like building the real stuff. Yeah, I mean, I'm jealous of companies that start now. Sounds amazing.

43:13 Will solopreneurs hire up?

Ernie Tedeschi: Yeah, it's a fixed cost in terms of time and in terms of money. I think that the uncertainty, Ben, bringing it back to like work and employment is that we politicians always tout small businesses for a very good reason, which is that businesses that start small, right? Then hire up and they are big drivers of aggregate employment growth over time. So there is a substantive reason why from an employment perspective, you should like small businesses. It's unclear whether these new solopreneurs that are coming, we think as a result of AI are going to grow in the same way as prior solopreneurs or partnerships or small businesses in the past. Are they going to hire up and have 5, 10, 25, 50 employees or are they going to stay really small? But on the other hand, is that a problem if what's happening is we're just getting, we're getting smaller firms but more of them, right?

Ben Zweig: Right, at that point, who cares? Then we just see the disintegration of the firm and we're in this like Coasean world where like We're just like yeah, we can like transact at the micro level very efficiently.

Ernie Tedeschi: Yeah, the Coasean optimism, optimistic way of looking at this is we've lowered coordination costs as a result of AI. And so what we would naturally expect is like, people that normally would have been hired in-house to do things are now effectively like all, they're all independent contractors, right? And they're just all one person firms. And as long as coordination costs are now low, that's fine, right? Like they get they're their own boss they get to work from home. That may be a more efficient way of setting things up in the future. And so as long as the volume of those solopreneurs, right, like grows over time, like you said, it shouldn't matter to us that they're not hiring up over time it will, it is uncertain and it is still to be seen whether that pace of growth in solopreneurship will sustain itself.

45:27 The vision for Stripe Economics

Ben Zweig: Yeah, okay, I mean, that is, in my mind, that's like one of the big questions of today's age, like will firms be bigger or smaller? I think like in the last decade or so I feel like we've seen the growth of these like mega firms where there's just like as very small number of companies who account for like a huge amount of market cap and employment and all that. So TBD, thank you so much for your time. This was so interesting. I'm sure we could go on for hours and hours, but I want to be respectful of your time. One final question for you. What is, I mean Stripe just acquired OpenRouter, big things happening at Stripe, you are now you are leading Stripe Economics. What is the vision for Stripe Economics? What can we expect from you and the team?

Ernie Tedeschi: Bigger and more of everything. So we really want Stripe Economics to be, to become like the platform where we talk about what's going on in the internet economy. And so that's going to be a combination of things. It's going to be more content from us that we author more data. So we're going to start putting up Stripe Data that will be publicly accessible and will help shed light on things like how SaaS businesses are doing, e-commerce in the consumer space, eventually knock on wood, assuming that the OpenRouter acquisition goes through more on AI and token costs. So all of those are kind of within our vision right now. We also want to bring on more collaborators as well and just, and be a central hub where outsiders can come in and talk about what they're seeing in the internet economy and how that's affecting things. So yeah, so it's like, I think you started with this. It's a really exciting time to be an economist and to lead this work.

Ben Zweig: All right, well, if you are a potential collaborator or anyone interested in this, follow Ernie and his work. Really interesting Substack and commentary on Twitter and all that. So I encourage everyone to follow Ernie and the Stripe Economics team. Thanks so much for being part of the show.

Ernie Tedeschi: Thanks for having me, Ben.

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.