Podcast
Important Information:
AI and Automation Adoption in the Fifth District and Beyond
Important Information:
Pierre-Daniel Sarte and Sonya Waddell share survey results on businesses' use of AI tools and other forms of automation, and research on the potential effects of automation on the nation's overall productivity growth. Sarte is a senior advisor and Waddell is a vice president at the Federal Reserve Bank of Richmond.
Related Links
Transcript
Tim Sablik: My guests today are Pierre-Daniel Sarte and Sonya Waddell. Pierre is a senior advisor and Sonya a vice president, both in the Research department of the Richmond Fed. Pierre and Sonya, thanks so much for joining me today.
Pierre-Daniel Sarte: Thank you.
Sonya Waddell: Thanks for having us, Tim.
Sablik: Our topic for discussion today is AI and its potential implications for the economy. There have been plenty of reports about businesses integrating AI tools into their workflows, but it's still challenging to get a clear picture of how widespread this adoption is.
Sonya, your team recently surveyed businesses in the Richmond Fed's Fifth District about their use of AI and automation. To start, I wanted to point out that distinction that you mentioned in that report between AI and automation, which you separate in the survey. Those terms are sometimes conflated in discussions of this topic, so how did you define them in your survey when you asked companies?
Waddell: I'm no expert on AI or generative AI. But one of the things that's concerned us over the years in our business surveys is that when firms respond to questions about using AI or GenAI, they're really talking about automating tasks, which is important and can definitely increase productivity but isn't necessarily the same as artificial intelligence.
In our Fifth District survey, we specifically defined AI as computer systems capable of performing tasks that historically only a human could do, such as reasoning, summarizing or solving problems. Generative AI is an even different beast. It uses algorithms to learn and thus can be used to create new content over time without a ton of human intervention.
We did not distinguish in our surveys between AI and GenAI, but like you mentioned, we asked both about automation and AI. First, we asked if firms implemented software, equipment or other technologies to automate tasks previously completed by employees. Then, we followed up by asking which of those automations involved the use of AI tools, and then defined AI.
We asked a similar set of questions in June to respondents of the CFO Survey, which is a national survey of CFOs that we've talked about on this show before. Now, if you remember, it's jointly run by three institutions: the Federal Reserve Bank of Richmond, the Federal Reserve Bank of Atlanta, and Duke's Fuqua School.
The biggest difference between those two sets of questions is that in the CFO Survey we asked for automation since the beginning of last year, but in the Fifth District surveys we wanted to go back further and ask for automation since 2022. In the CFO Survey, we also followed up with a question on examples of how firms have used or intend to use AI, in part to get a sense of how firms are using the tools and in part to make sure that we interpreted their responses reasonably correctly.
Sablik: We'll get to some of the different findings from those two surveys. But first I wanted to ask about the share of firms in the Fifth District surveys that report automating tasks in the last two years. What is that share, and how does the answer differ for manufacturing versus service sector businesses?
Waddell: When we asked these questions of Fifth District firms in June, just under half of firms had implemented software, equipment or other technologies to automate tasks previously completed by employees, and they did that in the last two years. For manufacturing firms that share was higher — closer to 53 percent — and for service sector firms it was a little lower — closer to 43 percent. Perhaps this is not surprising, as a lot of our service sector or non-manufacturing firms are in more labor dependent industries such as restaurants or hotels. This survey also includes industries such as real estate or construction that still rely heavily on workers.
Sablik: During the recovery from the COVID-19 pandemic, there were a lot of reports of firms automating tasks because they couldn't find workers. What did the businesses you surveyed say about their reasons for automating?
Waddell: The primary reason businesses reported automating was to enhance output or effectiveness of their operation. The second most common response is around increasing the quantity or the quality of employee output. Firms were less likely to cite reducing staffing or labor costs as a motivator, or automating because they were unable to find or hire workers with the necessary skill set.
It seems like, at least right now, more firms are automating for better quality or more quantity per employee. That's not to say it won't lead to lower hiring in the future, but firms don't cite lowering labor costs and reducing hiring as the most common reason to automate today.
Incidentally, a much [broader] array of firms cited the inability to find workers as a challenge one year ago than they do today.
Sablik: Comparing the results from the CFO Survey and the regional business surveys, what are firms saying about their adoption of AI?
Waddell: Let me start with those regional business surveys. In the Fifth District, about 15 percent of all firms engaged AI in at least some of their automation. This was similar for manufacturers and service providers. In the CFO Survey, about 60 percent of firms reported automating. Of those, about 40 percent adopted labor saving AI tools in the last 18 months or so. In other words, while about 15 percent of the Fifth District firms reported adopting AI, that was closer to 20 percent among the CFO Survey firms.
In the CFO Survey, we also found that, perhaps not surprisingly, larger firms were much more likely to adopt AI tools than smaller firms.
Sablik: Turning to some of the bigger economic picture issues, Pierre, there has been a lot of excitement about the potential for AI to boost worker productivity. Some researchers have predicted large increases to productivity growth and GDP over the next decade as AI adoption grows, while others have estimated more modest effects.
You recently wrote an Economic Brief that we'll link to which examines this question. What relationship did you find between the adoption of past technologies such as computers or smartphones and productivity?
Sarte: In 1987, Robert Solow famously quipped, "You can see the computer age everywhere but the productivity statistics." As it turns out, that quote remains true, even in the context of a 21st century world. There does not appear to be a well-defined relationship between the adoption of new technologies and productivity growth.
In the early period of the technology adoption boom in the postwar US economy, we saw diverging movements between productivity growth and new emerging technologies. For example, from 1953 to 1987, the share of capital attributed to software and R&D increased by 40 percent. Likewise, the share of IT equipment grew by 200 percent. However, over the same period, trend productivity growth fell from 3.4 to 1.7 percent.
When we examine what I call the "post Solow quote" period from 1987 to 2022, we find a similar picture. The share of intellectual property products increased by 100 percent and the share of IT equipment falls by 11 percent, yet the trend in labor productivity growth is essentially unchanged — around 1.7 percent in 1987 and 1.8 percent in 2022.
Sablik: The US experienced faster productivity growth in the '90s and early 2000s, which many people attribute to the adoption of computers and the internet. Based on your research, what seems to have played a bigger role, or the biggest role, in driving that growth?
Sarte: Undoubtedly, the widespread adoption of computers and the internet in the 1990s contributed to a temporary boom in productivity. At the same time, it's important to remember that it is not the only thing taking place in the 1990s. For example, the wave of deregulation and merging of different industries, including the telecommunication industry, in the '90s would likely have had a direct effect on productivity.
Aside from this, it's also important to recognize that when we talk about computers, we also have to keep in mind who is using those computers and in what way. It isn't enough to have the latest and fastest technologies. You also have to consider how those technologies are being applied.
It turns out that while labor productivity statistics appear somewhat disconnected from the adoption of new technologies, they show a noticeable correlation with labor demographics. In particular, the share of experienced workers in the labor force — those between the ages of 35 and 54 — appears to co-move strongly with productivity. The correlations are much weaker for the young and older workers.
To put it another way, the well-documented decline in productivity growth in the 1970s starts when the first baby boomers — those born in 1946 — enter the labor force as relatively inexperienced workers at around 25 years of age. By the time these same baby boomers become more experienced in their jobs and reach 35 years of age, productivity starts to rebound and continues to increase as the subsequent waves of baby boomers turning 35 and older enter [the workforce] during the 1980s and 1990s.
The effective application of new technologies to different tasks in the economy takes time. These observations suggest to me that this process is likely enhanced by experienced workers. There's a limit, of course. The data suggests that as workers get closer to retirement age, productivity starts declining again.
By the way, I hate to say it, but I'm past the 35-to-54 age group.
[Laughter]
Sablik: Bearing these things in mind that it takes time for workers to figure out the best uses for technology — it depends on the demographics of the workforce — when trying to forecast future productivity growth, how much does the adoption of automation and AI matter? Could AI be different than past technologies, or should we still be focused more on worker demographics?
Sarte: For now, I do not view AI as being so different than past technologies that, in their time, also appeared revolutionary: the invention of flight, electricity, penicillin, the internet and so on. It's important to consider not only the state of the technology being used, but also the experience of workers using that technology. I don't know that we should disassociate the two. It remains important to consider both current and forecasted demographic trends.
Sablik: So, would either of you like to give your prediction of where you think productivity is headed in the near term? Sonya, we can start with you.
Waddell: I'm going to leave the predictions to Pierre, but I'll say this. Although a surprisingly large share of firms in both the Fifth District business surveys and the national CFO Survey intend to use AI in the automation of tasks, the use of AI remains relatively limited in scope. In the Fifth District survey, almost 75 percent of firms who plan to automate in the next two years intend to use AI and well over half of firms are planning to automate. In the CFO Survey, well over 50 percent of firms who plan to automate plan to use AI tools.
However, so far, the ways that firms are currently using AI tools focus their use on tasks like relatively standardized accounting tasks, content creation, data analysis and HR activities. Customer support such as writing emails, creating presentations and writing HR policy were commonly reported, as was software development and fraud detection.
In other words, firms were using AI but using it for specific tasks. Many fewer firms referred to organization-level process or system enhancements. AI is helping to automate, but so far, our survey respondents don't report using AI in a way that transforms their organization.
On the other hand, the exploration continues. As one firm writes, they are currently undertaking a pilot to assess the options for where AI could have an impact.
I'll leave it to Pierre to tell us what the future will hold.
Sarte: The recent boon in interest regarding AI and its impact on labor productivity growth represents an updated version of what we examined in the Economic Brief and what Solow initially observed. To harness the power of generative AI and large language models takes skill and knowledge, much of which comes with experience. As much as being able to access information and answers easily surely matters, so does asking the right questions.
We've already had several years of data regarding the adoption of AI. OpenAI was founded in 2015 and ChatGPT 2 was released in 2022. For now, there's little evidence that we have suddenly become unfathomably productive.
That said, at least some temporary boom in productivity is to be expected going forward. In considering this exact question, Daron Acemoglu at MIT concluded that in the next 10 years, AI will lead to an increase of 0.07 percent per year in productivity growth. I'm inclined to agree with the general spirit of that estimate. In fact, given the current demographic trends, I wouldn't be surprised to see further declines in labor productivity growth as more experienced workers age out and leave the labor force.
In the very near term, I don't expect to see much in the way of productivity change. In using any form of AI to help complete a supplemental task, I find that it takes time to frame the right question, to consider the answer given and to potentially re-ask the question in different ways, not to mention that answers are not always correct. I cannot imagine that experience being too different for a plant worker or an engineer. My colleague Andreas Hornstein wrote an EB recently that examined these questions from a different perspective and a different time horizon.
Sablik: We'll definitely include a link to that as well.
Pierre and Sonya, thanks so much for coming on the show to talk to me about this.
Waddell: Thank you, Tim.
Sarte: Thank you, Tim.
