Many future-of-work researchers classify artificial intelligence as a general-purpose technology—a category of technologies that are defined by their ubiquity, their potential to be improved upon over time, and their ability to give rise to other innovations. General-purpose technologies aren’t invented every day. Other examples include inventions like the electricity and the internet.
In an essay in Academy of Management Discoveries, a team of researchers argues that, similar to other general-purpose technologies, artificial intelligence will only provide value to companies that create the right infrastructure to support it—or as they explain, the ones that invest in “complementary assets,” such as talent and data.
We spoke with Robert Seamans, one of the essay’s authors and an associate professor of management and organizations at New York University’s Stern School of Business, about the importance of complementary assets to AI adoption. Here are excerpts from our conversation, lightly edited for clarity:
Can you tell me a little bit about yourself and your research interests?
My name is Robert Seamans. I'm a professor at New York University's Stern School of Business. I've been there 14 years, and I'm in the department of management and organizations. I'm an economist by training. Most of the research work that I do is around the intersection of technology and strategy. So thinking about how firms adopt and use new technologies, how it affects their strategic interactions with other firms, how it affects their interactions with their customers, with their employees, and other stakeholders. I've looked at a range of technologies over the past decade and a half or so, but over the last four to five years, I've been focused on robots and AI.
In your essay “Capturing Value from Artificial Intelligence,” you argue that the companies that reap the greatest reward from general-purpose technologies like AI are the ones that invest in complementary assets. Can you share an example of how that’s played out in the past?
One of the best examples I can provide would be a historical case study by Paul David, who was an economist at Stanford University, about manufacturing firms that relied on steam power to power everything that they did and then switched over to electricity. He had two findings. The first is that it took firms years before they saw any productivity gains from that switch.
Which begs the question, why? And that's the second interesting finding of his paper. He digs into that and finds out that the way that these production processes are set up—keep in mind that they were all running on steam, and so they were optimized to run on steam—so if you switch out the steam, the performance is going to drop. And so each of the manufacturing firms has to tinker around with their production process and figure out a new way to set everything up to take advantage of what the electricity can do. Now, what's interesting is that it's not as if there's a single right answer that every manufacturing firm should then put in place. It turns out that the right answer for firm A is a little different for firm B and maybe a lot different for firm C, just depending on the idiosyncrasies of their production process.
So the idea behind that last point is that there wasn’t really a best practice for this? It was that every firm just kind of had to figure out what worked best for them and their workforces?
How should organizations identify the types of complementary assets they need to invest in?
What makes investments in complementary assets hard is that you don't necessarily know what they are ahead of time. It's going to involve incurring additional expenses through trial and error. I've visited manufacturing firms that are part of the automotive supply chain, and I've observed firms that are starting to put robots in place. And it's a similar type of thing where you can't just take a human out and stick a robot in. There are things a human does well. There are things a robot does well, but they're not exactly the same.
What you have to do is rearrange your production line in a way that takes advantage of the things that the robot can do. For example, if it's a stamper, where the robot will be feeding metal into the stamp and pulling away the stamp, in the past, a human would visually inspect it, but the robot can't do that. In order for that inspection to happen, you have to purchase sensors, and you have to wire that stuff up to a computer system. You have to purchase a new bunch of software that'll analyze it. Typically, this also involves hiring people that have very specific human capital around how the robot works and how the production process works, but also interpreting everything that is being fed to them via that software.
I feel like that example paints a picture of all of the different types of complementary assets that can matter, and the fact that these are going to really differ on a production-line basis. So for sure, firm by firm, but even within a firm from one production line to the next, it could differ. I've been talking about manufacturing, but you could imagine similar types of things happening within any given company, where there's just very different processes.
What are some of the most important complementary assets for AI that's being used in the context of knowledge work?
Just stepping back a little bit, I like to talk about the following categories of complementary assets. There's human capital, physical capital, and then digital capital. I don't think the physical capital matters as much here, but for sure the human capital does, and the digital capital does as well. It’s rarely the case that a firm is just adopting AI. They’re typically also starting to digitize a lot of the records that they have and things like that. So you need some process. At a minimum, you have to purchase some software to help with that digitization. You have to think about how you're going to store it, how you're going to retrieve it. So I would think of that as all the digital capital that one invests in.
In terms of specialized human capital, some of your existing managers might be able to handle some of that, but you're going to need to train up people in terms of the new systems that you're putting in place. In some cases, you might want to, for security reasons, do some of this on-site or something like that. And so maybe you also have to start hiring folks that know how the AI itself works and know what other types of assets need to be invested in terms of, again, our sort of hypothetical of this records digitization, storage, retrieval. So someone who has some knowledge about whatever that industry is and how records get created, which records matter, and different things like that.
How should organizations incorporate investments in complementary assets into the way they think about hiring and upskilling?
There's the short term, and then there’s thinking a little bit longer term. What makes the most sense longer-term is having a strategy that’s continuously updated. It's not trying to cycle through different workers that have different skills. It's a company trying to find the workers it's most comfortable with, that share some values that the firm thinks are important. And then over time, it’s making sure the workers have the skills that are needed to take advantage of whatever the new technologies are. You mentioned generative AI, but there might be something different a year from now. Moreover, what matters now in terms of generative AI is almost certainly going to be very different a few years from now.
You don't want to constantly have to chase folks that have these skills. Ideally, you've got a set of employees that are totally bought into what you're doing as a company, and then you're training and retraining them over time to have the skills that are needed to take advantage of the technology and work well for you as a company.
That was the long term. What about a short- and medium-term strategy?
I think any company that wants to be successful should be looking to the long term. And so this sort of strategy that I laid out is the one that they should be focused on. But if for some reason a company doesn’t want to do that, then you can spend a bunch of money right now to hire recent grads that at least know how to put together some of these newer technologies and are very comfortable with it. They lack all of the institutional knowledge that might matter in whatever the industry the firm is in, so they're going to be missing out on that.
You’ve also worked on research that's attempted to map AI capabilities onto existing human capabilities as a way of determining which occupations and industries are going to be most exposed to AI. You found that legal services, securities, commodities, and investment industries are the industries that will likely be most exposed to LLMs. Why those industries?
So together with Ed Felton and Manav Raj, we've been trying to think about ways to study which occupations and industries, and for that matter geographies, are going to be most exposed to advances in artificial intelligence. I appreciate that you also used the word “exposed.” We very intentionally used that word. What we have in mind by exposure is that there are certain elements of the occupation or the industry that look similar to what it is that AI can do. But that doesn't mean to us that there's going to be substitution. We don't know. There could be substitution in some cases. There could be augmentation in many other cases. In some cases, there may be some jobs that are at risk of being substituted by AI. There are also going to be some new jobs that are created that we've never even thought of before. But I think by and large, what's going to happen is that it is just that many of our jobs are going to change as a result of the technology.
The idea behind this measure we came up with, which we call AI occupational exposure, is to try to identify the occupations that will change as a result of AI. We had a paper on that a few years ago. When ChatGPT came out at the end of last year, it struck us that there was this dramatic advance in what language modeling could do. And language modeling was one of the inputs into this earlier approach that we had taken, where we had looked at occupations that are exposed to advances in AI. And so we were able to update that earlier methodology that we had to focus only on language modeling.
So one of the things that came out of that is what you had referenced earlier, which is that some of the top industries that are most exposed to advances in language modeling are things like legal services, securities, insurance, and things like that. If you look instead at occupations, the vast majority of the top 20 occupations are teachers of various sorts.
In terms of why those types of occupations and industries score very high, it is just because a lot of the things that matter in those industries have to do with language. When you think of what a lawyer does, a lot of what a lawyer does every day is reading lots of things as well as writing lots of things. And so facility with language really matters.
How do you think the results of the paper fit within the wider literature on the impact of artificial intelligence on jobs?
There are a few other papers that we're aware of that have tried to do things similar to what we're doing, which is coming up with a systematic way to analyze how different occupations and industries are going to be affected. We all take different approaches. I guess the truth is probably somewhere there in the middle. At some point in the future, we can look back and we can see who was more or less correct on some of these things.
Going back to the first area we talked about, which was capturing value from AI, what kind of advice would you give a business leader who's trying to assess where they are on the maturity curve for AI adoption?
There are definitely tools and reports out there that could be useful. If you look at Accenture, McKinsey, PwC, they all have published various things that I think can be helpful to people as firms want to assess where they are. [Disclosure: Seamans does some work for Accenture.]
In terms of what I’ve seen firms do that they think is helpful—they will implement whatever the new technology is in one business unit, or one part of the firm, or one part of the factory, and work at it until they've got it right. Then from there, they then sort of scale across the rest of the firm. As we talked about before, the scaling part is going to be difficult because what works well in one part of a firm doesn't necessarily work well in other parts of a firm. Some firms have a sort of internal tech transfer office that can help the business unit leaders see what works in one part of the firm and then try to implement it in other parts of the firm. I've heard that firms are doing that. It sounds reasonable to me. I've never seen a systematic study on this though.
You wrote in your essay that people often consider a new technology and its implications as something never seen before, when in fact the underlying dynamics are the same as for earlier technologies. What would you say about AI adoption in that context?
This point about complementary assets has mattered for all the other technologies that have come before AI, and it's going to matter for AI as well. AI can drive productivity improvements in a firm, make the firm more profitable, do whatever it is that they're doing better, but it's not the case that you can just buy AI and sprinkle it on the firm, and you'll get these productivity improvements. It involves a lot of learning by doing, about how you need to change what your firm does to take advantage of what AI can do.
We were talking about robots before—the average cost of a robotics arm in a production setting is about $30,000. The average cost of it with all of those complementary assets that we talked about before is about $90,000. So we're not talking about just purchasing a few other little things here and there. If you're making a decision to invest in robots, it's this huge investment that involves investing in all these other complementary assets. I don’t know exactly what those numbers would be for investing in AI, but firms should have in the back of their head that whatever it costs to in order to purchase AI, at a minimum, you should expect to spend at least that amount in terms of all the other investments that you have to make. There’s no question that that’s going to be a floor. If anything, it’s going to be a multiple of that.