Related Links

• Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives
• How Might AI Change the Workplace? Evidence From Corporate Executives

Transcript

Tim Sablik: My guests today are Sonya Waddell and John Graham. Sonya is a vice president and economist at the Richmond Fed, and John is the D. Richard Mead Professor of Finance at Duke University's Fuqua School of Business. Sonya and John, thank you for joining me.

Sonya Waddell: Thanks, Tim.

John Graham: Pleasure to be here.

Sablik: So, you're both involved in running the CFO Survey, which is conducted by Duke's Fuqua School in partnership with the Richmond and Atlanta Feds. Sonya is part of the Richmond team that works on the survey and John is the director of the survey. Along with the other members of that team, you recently published an NBER working paper that used responses from the CFO Survey to study the effects of artificial intelligence on productivity and the workforce. That's going to be our topic of discussion for today.

Certainly, this goal of quantifying the effects of AI is one of the most pressing economic questions today, I think. Many economists have been attempting to estimate the size and impact of AI adoption using various methods. Sonya, what makes surveys like the CFO Survey effective tools for finding answers to these questions?

Waddell: There are many analysts and economists trying to understand how AI technology will affect our economic potential and the kinds of jobs that will be available. The impact of AI on productivity and labor markets depends critically on when and how firms adopt the technology, and no one knows this better than the firms themselves.

Is a financial leader in a private sector business all knowing about the potential for technology in the future? Definitely not. But does the CFO of a firm know more than just about anyone about how and to what end their company will use technology to meet demand, save costs, and maximize profits? Absolutely.

So, by aggregating responses across firms, surveys like the CFO Survey can learn quite a bit about the current state of technology use across industry in the broader economy and how leaders in those industries see the technology being used going forward. I believe that respondents are also likely to have a pretty good sense of what their competitors are planning to use, too.

The other thing that we have in so many of our business surveys, including in the CFO Survey, is data from the past and the promise of future data. If we continue to ask questions about AI, we'll be able to evaluate how their responses change over time and calibrate the aggregated responses across what we know to have happened in the economy. That can be a powerful tool for evaluating both the current and future state.

Sablik: Getting into those responses and timescale, you all surveyed senior financial executives for hundreds of businesses from the end of 2025 to early 2026. John, what did these business leaders have to say about the way their companies are using AI now and plan to use it in the future?

Graham: We asked two questions to try to answer the question you just posed. First, we asked companies what are your motivations for adopting AI. Then, we asked a second question about what have been the effects of your actual usage of AI. It's notable both from what companies said and from what they didn't say or what they emphasized less in their answers.

In this first question about motivation to adopt AI, they said the number one factor motivating them was to improve production efficiency. Number two was to enhance decision making and management. Number three was to improve labor productivity — now, labor productivity, remember the definition is the output divided by the number of workers. The fourth thing they mentioned was reaching and trying to serve their customers better, and/or innovating new products and services. We call those last two the satisfying demand channel.

What they didn't say they were trying to accomplish by investing in AI is to reduce costs, or necessarily to reduce labor very much. Yes, we expect efficiency gains. But that's offset, at least to some extent if not fully, by increased costs of buying AI services.

Then, we asked that second question about what's actually happened since you have invested in AI. The answers line up pretty well with the motivations, but there are some differences. One of the top things they've gotten out of investing in AI is increased productivity. They've also talked about increased decision speed and accuracy, and the ability to spend more time on high value-added tasks.

So, all of these things are about the positive side of creating better and more efficient processes and products from AI. They haven't reduced their cost structure and only with some exceptions have they actually reduced their workforce attributable to AI at this point in time.

Sablik: As part of doing the surveys, you collect details about the businesses that are responding. Are you able to use that to see which types of businesses reported higher AI use and which reported lower adoption? Did these results change when you asked firms to look towards the future?

Waddell: First, it's worth noting that from this survey, AI adoption is already relatively widespread and is expected to expand pretty rapidly. Across the 750 firms that we surveyed, well over half had invested in AI in 2025. When it comes to expectations for investment in 2026, that number jumps to more than 80 percent.

Of course, investment magnitudes vary widely across companies, with larger firms spending substantially more in absolute terms. Although relative to number of employees or total capital expenditures, smaller firms are often investing more.

By industry, companies in high-skill sectors like professional business services, educational services, and healthcare have invested and plan to continue investing the most in AI. Lower-skilled services firms such as retail trade, hospitality, or transportation invested less in 2025 and have lower investment intentions for 2026. Of all the industries, manufacturing and construction firms are the least likely to have invested thus far, but even their investment plans for 2026 strongly outpace their 2025 investment.

I know you didn't ask this, Tim, but I thought it worth pointing out that most AI spending — especially among smaller firms — was directed towards operational expenses such as software subscriptions, services and training rather than hardware purchases or internal software development. This is different from past technological advancements like the computer revolution of the '90s, which required firms to invest substantially in hardware or equipment.

We also asked firms why they did or did not invest in AI. For those that did invest, they were most likely to report looking for production efficiency or enhanced decision making and improved productivity. The reasons that firms were not investing in AI? The technology is immature, the workforce is not trained to use AI, concerns about privacy, and just a lack of knowledge about AI capabilities were common among firms for reasons why not to invest.

Sablik: On that topic of productivity, there has certainly been a lot of conversation about the productivity potential of AI — the potential to make us all more efficient, more productive at our jobs. What do CFO Survey respondents have to say about the current and expected productivity potential of AI?

Graham: Companies expect productivity effects to be quite large from investing in and using AI. They expect productivity to increase by about three percentage points in 2026. Normally, productivity doesn't go up by even three percent, let alone increase by three percent due to one particular factor. So, this is a very big number.

As Sonya pointed out, one thing that's going on in this current AI wave is most companies are not buying a bunch of hardware. The last technology wave we might have experienced was the computer revolution of the 1990s and you pretty much had to buy a computer to benefit from the productivity. Here, companies aren't spending as much on hardware. The way we would subscribe, say, to Netflix to watch movies, you can subscribe to, say, OpenAI. One good thing about this model, relative to the past, is that there's not this huge upfront cost for most companies. You can just start immediately with a lower subscription, so it means more companies should be able to participate.

Sablik: The upfront cost would be for the AI firms themselves that are providing these services.

Graham: Exactly. Thank you for saying that.

There are some mega firms out there who are doing this massive infrastructure investment — building data centers, developing the software — just like Netflix builds out capacity to produce films and distribute films and the rest of us pay a relatively small monthly fee to stream movies. Likewise, that big upfront cost is being borne by some of those big cash-rich mega firms right now, but your typical firm is just a user of AI and that's really what our survey is more focused on.

Sablik: Sticking with this topic of productivity, in your paper you discussed the difference between implied productivity and reported productivity gains through AI. Can you explain what those two terms mean? And then, are you able to determine how much of the implied productivity is due to AI versus other factors that are going on at the same time?

Graham: We directly ask CFOs, "How much do you expect your productivity or output per employee to go up?" It so happens we're also able to calculate productivity using other items of data from the survey.

For example, we can look at revenue per employee. If the reported or expected productivity will go up by 3 percent, you might expect to see something similar in revenue per employee, but we do not find that. Using the numbers the CFOs tell us, it looks like it'll go up by about 1.5 percent in the year 2026. We call this a productivity paradox.

Now, why might that go up by less? What we think is happening is a delay in the process. Let's say you spend the first half of 2026 building up your AI capacity, training your employees. They're getting really good at using AI and maybe replacing old processes. Then, in the second half of 2026, maybe you're finally producing more products and services. But even that, in many cases, you don't get to sell until maybe 2027. And so, there is just a natural delay in the productivity statistics.

There's a very interesting historical comparison. Back in the 1990s, Nobel Prize winner Robert Solow had a famous quote. He said, "The computer age is everywhere but in the productivity statistics." At that time, everybody was talking about computers [and] productivity, computers [and] productivity. Yet when he measured productivity in the data, it wasn't showing up. That's sort of what we're finding now. Like the 1990s, productivity will eventually go up, but it'll happen with some delay and it'll probably be spread through time.

Sablik: Sonya, as John has been alluding to, the other part of this conversation around AI is about its impact on the workforce. Did businesses in the CFO Survey have any insight into the impact of AI on their workers?

Waddell: We had two ways that we asked firms about how their employment might change due to AI. First, firms reported the expected and realized change in total employment that they attributed to AI. Second, companies were asked to provide open-ended descriptions of roles and responsibilities that have been or are expected to be replaced or complemented by AI tools.

Most firms do not anticipate widespread job replacement from AI. This is consistent with other surveys we've conducted at the Richmond Fed. Although firms look to AI, as John mentioned, for labor productivity, they broadly do not expect to lay off workers.

In this survey, 50 percent explicitly reported that AI will not replace any roles or responsibilities. Less than that indicated some degree of job replacement due to AI. It's also worth noting that some of the role replacement is attributed to those not investing in AI, indicating that labor reallocation might go beyond the firms that are investing.

Second, larger firms are substantially more likely to report employment reductions due to AI, while smaller firms are more likely to report employment gains, in fact. This is already true for 2025 but is expected to be even more pronounced in 2026.

Is this surprising? Well, probably not. Large firms tend to employ a wide variety of occupations and workers that tend to perform a more narrow set of tasks, which increases any given worker's vulnerability to automation. In addition, the large aggregate labor costs incurred by large firms might create incentives to use AI explicitly to reduce employment. It's worth noting, though, that the effect on employment is still expected to be small, even among those larger firms.

Third, CFOs expect any reduction in the share of their workforce doing routine clerical work to grow over the next few years. This reduction will at least be partially offset by an increase in skilled technical workers. We have seen that happen in past technological developments.

Similar to the second point, the effect varies by firm. Large firms are expected to shed routine clerical employees while holding steady creative and technical employment. Small firms are more likely to hold steady routine jobs and increase skilled technical jobs. So, we might see some labor reallocation, both within and among firms.

Sablik: From the survey, do you have any insights into which jobs you would expect to be affected?

Waddell: We found that the largest and most frequent mentions in the [AI] enhancement categories are in areas such as marketing, accounting, finance, and management, in addition to roles such as data analysis and software-related tasks. Under the [AI-driven] labor replacement category, comments were spread over a wider range of tasks that included manual tasks, data entry, customer service, and administrative tasks.

The broader dispersion in replacement compared to enhancement suggests that AI-driven substitution is not concentrated in a single dominant role, but instead affects a variety of activities. This is consistent with measures developed by other researchers.

Graham: Can I throw in one little irony here? We had reams of data to analyze. Some of it was this detailed job information that we were going to connect to the Bureau of Labor Statistics data sources and exact job titles. We tried to use AI to do it and it didn't work very well. [Laughter] We had to hire a research team to eyeball it and use human judgment to make those connections.

Having said that, once they did this, then we were able to create this really cool dashboard. If you go to our paper, you can find a link to it that presents and summarizes all this data that the human beings created and organized, but AI helped to build the dashboard in a really fancy way. So, it looks like our ability to use AI is good in some things and not so good in others at this point in time.

Sablik: That's a great example. Do you have any plans, John, to expand this research? Are there other questions that your team would like to explore around this topic?

Graham: Every quarter, we have what we call the core questions that don't change and then we have these topical or rotating questions. All these AI questions we've been talking about happened as rotating questions. We do plan to, this quarter, ask a clarifying question about AI to learn a little bit more about what categories companies are spending on [with] AI and how they categorize it within their books. We're doing that in part to set us up to ask better, more pinpointed questions in the future about the use of AI and its effects.

Waddell: I'll just go back to what I said in the beginning. One of the most powerful things that we can do with these business surveys is to look over time. We have firms that are on the cutting edge of investing and implementing this technology. We can use these questions over time to understand how the landscape is evolving.

Sablik: Putting in a free plug for ... some of those new AI businesses out there that might be listening and want to join the CFO Survey ...

Waddell: Yes and, of course, also a thank you to our existing businesses and definitely a plug for new businesses. Right, John?

Graham: Absolutely. We're surveying CFOs, so we always need to replenish as certain CFOs retire or move on to a new job. Likewise, new companies we need you in our survey, too.

Sablik: Yeah, we'll include the link for that in the show notes for this episode.

Sonya and John, thank you so much for joining me today to talk about this really interesting research.