Related Links

• The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks
• Firm Data on AI
• How Might AI Change the Workplace? Evidence From Corporate Executives
• Mind the Gap: AI Adoption in Europe and the US

Transcript

Tim Sablik: My guest today is John Haltiwanger, the Dudley and Louisa Dillard Professor of Economics at the University of Maryland. John, welcome to the show.

John Haltiwanger: Thank you for having me.

Sablik: We've talked on the show before about CORE Week, which are weeklong collaborations between Richmond Fed and academic economists, hosted in Richmond about seven times a year. You're here attending and presenting at the August CORE Week, which is happening now as we record this. I'm grateful that you could take some time out of the schedule of events to join me in the studio and talk a bit about the paper and research you presented here.

The paper that you presented is co-authored with several researchers at the U.S. Census Bureau and the University of Maryland. You attempt to measure how firms and workers are using AI, which I think is the billion- or trillion-dollar question in the economy right now.

To start, could you talk a bit about the existing measures of AI adoption that we have before you started this project? What can those measures tell us and in what ways are they, maybe, falling short of providing the complete picture that we would like?

Haltiwanger: It is the big question as to what the impact of AI is in terms of both firm and worker use, and what the impact is going to be in terms of firm performance and labor market outcomes.

The existing information is not great. Some of the best information is out of the Federal Reserve System. There have been some very nice surveys done by both the Atlanta Fed and the Richmond Fed. I think one of the very nice things both the Atlanta Fed and Richmond Fed have done is, in terms of their surveys, is add supplements and targeted information about changes. And so, not surprisingly, in both Atlanta Fed and Richmond Fed surveys, there have been questions about AI.

We've learned a lot from those. One paper that came out this year that got lots of attention out of the Atlanta Fed had about 1,000 U.S. firms. It suggested an incredibly high adoption rate, close to 80 percent of firms. If you actually check carefully, that number was on an employment-weighted basis.

Sablik: So, they're weighting the firms that have more employees.

Haltiwanger: That's right.

That number fed into the sense that this AI wave is coming really rapidly at us. If we think about past periods of innovation, it was more of a slow diffusion process. Some have speculated AI is different; that this is going to hit us really hard. I don't think we even know the answer fully to that question yet. The platforms themselves are releasing statistics about use rates and so on: the number of users is phenomenal, in the hundreds of millions.

Both from the Atlanta Fed and the Richmond Fed, another nice survey of CFOs asked about AI investment. The number that pops out, the comparable number, isn't quite as high as the Atlanta Fed number but it's still pretty high. It was about 58 percent.

By the way, that particular survey had about 750 respondents. That's a healthy number, but there are six million firms in the United States. So, how well do the either the Atlanta Fed or this other survey capture what's actually going on in the full U.S. universe of firms?

Alexander Bick and co-authors out of the St. Louis Fed have something called the Real-Time Population Survey. Bick's approach is different. It's a household survey, but he was particularly interested in asking about workers and their [AI] use. He was particularly interested in generative AI implementation — there's a sense everybody seems to be using it. Alexander Bick's [AI adoption] numbers are in the low 40s. It's about 43 percent.

There was a very nice article by Ben Casselman — one of the very fine columnists of the New York Times — who wrote in the last month, based upon all this, if you want to know how many firms and workers are using AI, good luck because the numbers are all over the place. I think the Census Bureau's new Business Trends and Outlook Survey AI questions, and particularly its AI supplement, is beginning to give us more reliable numbers.

Sablik: That's the dataset that you were alluding to that you and your co-authors designed and worked with on this study. It sounds like you took up Casselman's challenge.

Maybe you can talk a little bit more about that dataset, how it developed, what kind of things you had in mind when you were designing it, and the questions you were looking for.

Haltiwanger: We need high frequency indicators tracking changing patterns in the U.S. economy. There were some efforts at Census — in particular, what was called the Business Pulse and the Household Pulse — early in the pandemic to track what are people doing here in this crazy time. What's emerged out of that is what's called the Business Trends and Outlook Survey [BTOS]. They're in the field for two weeks with a sample of 200,000 firms and they're trying to track changing patterns of economic activity on all kinds of dimensions. Now, as AI has come on, it was obvious this was an area of interest.

I'm advisory to a group at the Census Bureau to track the impact of AI on the economy. We convinced Census' upper-level management to add a supplement to the BTOS. The supplement that I presented about here at this conference is a supplement that went into the field in late November 2025 over a 12-week period of time. So, 200,000 distinct firms every two weeks were asked about the supplement.

Given both its scale of the survey, but also it's truly national representation, it allows you to cut the data in very rich ways. You can look to see what's happening about very small firms versus more large firms. You're able to look at sectors in great detail. Once you do that, then you actually are starting to be able to reconcile the widely disparate numbers that are out there.

Sablik: Let's dig into that a little bit. You have this much larger sample and you can look at different types of firms. What are some things that you and your co-authors learned about AI use when we look at the firm level?

Haltiwanger: What has popped out at us was just enormous variation across sectors in terms of AI adoption. There are some AI-intensive sectors. It's the information sector where the AI developers are, but also applications to information technology. Professional, scientific, and technical is another; it's what's called NAICS 54. Other heavy AI-intensive sectors are education, finance, and health.

In contrast, there are clearly some AI laggards at this point. The retail trade sector is below the national average. Go to your local restaurant or hotel. Are they using AI intensively? The answer is no. Transportation and warehousing, no. Agricultural services, no. And so, there is a large number of sectors that account for an important fraction of U.S. economic activity where they're well below the national average. So, sector matters a lot.

It turns out, quite interestingly in this case, so does firm size. We find that the adoption rates are much higher at larger firms than at small firms. This turns out to be, in many ways, the critical factor [of AI adoption]. In the United States, 90 percent of all U.S. firms have less than 20 employees. Those are small firms. Literally, the median firm has less than five employees.

You might say, should we care about those firms? If nothing else, for an accurate measure [of AI adoption] we want to know who we're asking and who we're not. But that 90 percent accounts for about 20 percent of U.S. employment. That's a big enough share that you ought to pay attention.

If I take AI-intensive sectors and I weight the firms that are larger firms in those sectors, then I'm getting estimates in the range of 50, 60, and 70 percent adoption rates, depending upon exactly how I cut the data or what particular sector I'm looking at. Those were exactly the kind of numbers that are consistent with numbers that came out of the Atlanta Fed and the Richmond Fed. We had a very similar size distribution and a similar industry distribution as these surveys, and we got very similar numbers.

So, I think at the end of the day, has AI adoption really taken off? In some really important parts of the economy, the answer is yes. Is it really a general-purpose technology in the sense that it's just ubiquitous and everywhere, all sectors and all firm sizes. The answer is no.

Sablik: Or, at least, not yet. [Laughter]

Haltiwanger: Not yet.

Sablik: Sticking with this difference around firm size, another big focus of your research is entrepreneurship and business formation. As you well know, since you've done a lot of research and written papers on this, since the pandemic there's been a big surge in new business formation. Maybe you could talk a little bit about what seems to be behind that surge and what was driving it.

Haltiwanger: In the pandemic, we've seen this surge in new business formation in the United States. This is after a couple of decades-long decline in entrepreneurship, particularly for new employer businesses. When that first surge happened in the pandemic, everybody was going, "What's going on? Why suddenly is the United States becoming more entrepreneurial in the midst of this chaos of the pandemic?"

Slowly but surely, we saw there was a whole bunch of factors that were going on. Some of it was the relocation of activity that was going on. Because people were spending more time working from home, you were seeing a surge of support businesses in the surrounding suburbs of downtown areas. Some of it was related to the working-from-home phenomenon that we could do a lot of business activity remotely, that we could interact with each other remotely in ways I don't think we've quite understood before. Individuals figured out, wait a second, if that's true, I can go out on my own.

Sablik: More recently, there's been some commentators who have suggested that maybe AI might be helping more people start their own business. Do you find any explanations for the startup surge that might be connected to AI?

Haltiwanger: Absolutely. Slowly but surely, we've emerged from the pandemic. Entrepreneurship has stayed equally high. As you start to unravel it, you actually realize an important part of this is AI.

In some respects, AI is not new. Gen AI hit like a tidal wave, but machine learning and large language models and the like have been under development for the last couple of decades. So, one question is whether the surge in AI and new business formation would have happened in the absence of the pandemic. But if you combine that with everything else I said in terms of people [thinking] they can go out on their own, these things reinforce each other.

So, we had this surge in the pandemic, it stayed high, and then we've had a new surge in 2025 and '26. We're even at higher points than any time in the pandemic itself.

It's two key sectors playing a role. One is the AI-intensive sectors I talked about earlier — it's the information sector and professional, scientific, and technical. More than that, clearly it's AI related. When you fill out this business application, they write in "my business is going to do X." Also, you have a business name. There's been some very nice work that extracted the text information from the write-in information. Using AI-related terms just skyrocketed, particularly starting in 2023 but it's continued to today.

Interestingly, the other sector that's also taken off — that also, by the way, took off in the pandemic — is non-store retailers. Basically, it's e-commerce. Online platforms, and particularly third-party sellers on online platforms, has really taken off. TikTok Shop is growing dramatically in the United States.

AI enables individuals to set up businesses particularly quickly, both set up a website [and] customer service, the ability to interact with customers. In this new surge, we are seeing an increase in new likely employer businesses.

We're especially seeing a dramatic increase in likely new non-employers. The business press likes to talk about these as "solopreneurs." These are individuals that are going into business for themselves and not hiring workers. The working hypothesis that we're still very much investigating is you don't even need much of a support staff because a lot of the support functions you can create through AI.

Sablik: Right. You have these two channels: the folks that are using AI to be solopreneurs, and then the businesses that are responding to the growing AI ecosystem and trying to meet some demand there.

Switching back to the bigger employer firms that we were talking about earlier, from the BTOS — the survey that you put out — did you learn anything about how workers at those firms are using AI and whether it is complementing their workers or substituting for workers?

Haltiwanger: There was some sense that maybe there's lots of workers are using it, but firms haven't formally adopted [AI]. So, there's a very nice set of questions that we asked to try to draw out how workers are using [AI] and how they've been impacted.

We asked the firms, "Are you using it for these different kinds of business functions?" And then we asked, "How are your workers using it, even if it's not part of the formal firm production process?" Turns out the gap isn't as large as we expected.

It is true if we just start to take an employment-weighted basis, the fraction of workers using AI is higher than the fraction of firms [using it]. Well, how can that be? Our interpretation of that relatively small gap is the firm is aware that some of their workers are using [AI] for email or other kinds of purposes, but you're not required to.

Is AI going to displace workers? I think the evidence is, so far, no. It's just not. There's two pieces of evidence that came out of this new BTOS supplement. One is we literally asked if workers are using AI. Is it being used to substitute for tasks that we're previously doing or to, as you said, enhance or augment [tasks]? It's overwhelmingly augmentation at this point. There's relatively little substitution.

Another hypothesis is, oh, you're going to create new tasks. We see some of that. If we ask, conditional on workers using it, what fraction of firms say it's augmentation only, it's around 70 percent. It's less than 10 percent where we're getting numbers in terms of substitution or creation of new tasks.

The interpretation we have of it is a large extent AI at this point is enabling people to do what they were doing previously, but hopefully better. And so there could be productivity gains, maybe even big-time productivity gains out of this. But it doesn't look like it's displacing.

Then we actually asked more directly, "Did you cut jobs because of AI?" The national number that we get is tiny. Only 2 percent of firms say they're cutting jobs due to AI. That 2 percent number is also true at the large firms as well. There's just not much evidence across the size distribution that there are currently big-time employment cuts.

Now, I'm going to be very cautious about that. That's as of November 2025 through February 2026. With rapid diffusion and firms trying to figure out what's going on, could there be an acceleration of this? Yes and, indeed, I think this is exactly why BTOS wants to do what it's doing.

Sablik: Yeah, for sure. Seems like there is great value to continue surveying these firms and tracking this information.

Have you or your colleagues thought about other questions that you might want to ask, or other outstanding things that you'd like to get information on?

Haltiwanger: There's a set of questions that ask if you are using AI, what organizational changes are you making? Are you changing your business practices? Are you investing a lot more in computing power because you're doing all this AI? Are you doing a lot more cloud services because AI requires a lot of processing on the cloud?

We are very interested in being able to try to, in this current process, track what those organizational changes are. Literally, are businesses changing their business practices in key ways and the way they just organize themselves?

A challenge for us is what questions should we be asking about those organizational changes. We asked a set of them in this most recent supplement. Even then, it turned out it was informative in the following sense. Even though I said most businesses aren't seeing a decline in employment, the ones that [have been] are actually the ones that were doing big-time organizational changes. Interestingly, of those businesses that seem to be all in, there's something going on more on employment there, enough so that we may be seeing the displacement that people are concerned about.

Sablik: Yeah, watch this space, right?

Haltiwanger: Exactly, exactly.

Sablik: John, thank you so much for joining me to talk about your research. It's super interesting.