EXCITEMENT ABOUT ARTIFICIAL INTELLIGENCE (AI) is on the rise among business leaders and investors, and for good reason: AI has the potential to transform business and the economy. But are businesses embracing it yet? If so, to what extent?
Very little evidence exists about which technologies have actually been adopted thus far, who the adopters are or where they are located — which has made it difficult to understand AI’s true economic and managerial implications. In a recent paper, I and my colleagues Erik Brynjolfsson (Stanford University), J. Frank Li (UBC) and Zachary Kroff, Emin Dinlersoz, Lucia Foster and Nikolas Zolas of the U.S. Census Bureau set out to fill in some of the blanks. Our study sheds light on the rate and direction of AI diffusion and the influences that will shape its ultimate impact on workers, firms and society. In this article I will summarize our key findings.
Why Study AI Adoption?
Throughout the history of big technological shifts, firms have required time to adjust and achieve new productivity gains. Changing processes, restructuring the organization, hiring the right talent and learning how to integrate new technology into products and services is difficult. So difficult, in fact, that adjustment often unfolds unevenly across firms.
AI represents an important new input into the production of novel ideas. If AI provides an antidote to purported declines in innovation or a pathway to new or better jobs, this could be cause for optimism. Understanding the firm context shaping this innovative activity is essential. At the same time, however, concern is growing that many industries are becoming more concentrated — and changing technology is increasingly taking the blame. In recent years we have seen the rise of ‘superstar firms’ like Amazon and Tesla that may benefit even more from technologies like AI.
One key issue for leaders is this: If AI goes hand-in-hand with rising inequality between firms and workers, we must think carefully about its broader competitive and socio-economic implications. Yet little guidance exists for decision-makers at this critical time.
A new nationally representative survey, the Annual Business Survey (ABS), conducted jointly between the U.S. Census Bureau and the National Center for Science and Engineering Statistics (NCSES) addresses this gap. The ABS began collecting information on the adoption and use of several advanced technologies — including AI — as of 2017. We sent our survey to 850,000 firms, and obtained a representative sample of 573,000.
In our work, we matched firms in the first year of the ABS to the Census Bureau’s Longitudinal Business Database (LBD), which provides administrative data on outcomes such as employment and revenue across the firm lifecycle for a nearuniverse of firms in the economy.
Uncovering the current state of geographic disparity in AI adoption was a key objective of our analysis.
Our study progressed in stages. First, we uncovered broad patterns of AI use across the entire U.S. economy. AI use showed up primarily in firms that also had high reliance on digital information and cloud computing, suggesting key interdependencies among recent technology trends. We found that firms that rely on robotics also typically use AI in production, pointing to the importance of embodied technology that can spread through tangible inputs. In other words, the ‘recipe’ for using AI effectively is not simple and requires attention to other types of investments and practices. Thus, we next drilled down into a large sample of 75,000 younger and arguably more dynamic firms for which we had unusually detailed organizational data. Following are the firm characteristics we analyzed in this subset:
OWNER CHARACTERISTICS. In previous research, high-growth entrepreneurship has been linked to a number of founder characteristics and motivations. The founder’s experience and identity are influential in early decision-making, when objective data is sparse, and often entail difficult-to-reverse commitments that affect the trajectory of the firm. Owner characteristics matter for performance throughout the lifecycle, as well.
We expected that several features of the founder would play a role in AI adoption and/or firm growth. Intuitively, founders with advanced degrees are likely to lead firms that are more technology-focused, equipping them with the objective data inputs required for advanced analytics, including AI. They are also more likely to be familiar with advanced technologies and their applications as a result of their advanced formal training.
Informal training is often important, as well. Prior entrepreneurial experience has been identified as one of the key predictors of firm survival and performance. In particular, ‘serial entrepreneurship’ may be correlated with a better assessment and exploitation of the opportunities advanced technologies offer. We also expected that the goals and motivations of entrepreneurs would be critical to the trajectory of the firm. For instance, smaller, older firms lagging in advanced technology may have been set up to pursue objectives other than growth, potentially lacking the inputs and scale that would justify AI adoption. In this vein, the entrepreneurship literature has noted a clear delineation between ‘lifestyle’ and ‘high-growth’ entrepreneurship.
We anticipated that lifestyle-focused business owners would be less likely to make the technology and organizational investments required to deploy AI in a meaningful way, if at all. We further anticipated that founder motivations may have farther-reaching influence on how AI is deployed, in practice. For instance, recent work points to the important distinction between founders motivated by pro-social aspirations and those motivated by earning potential. The connection between founder motivations and the use of advanced technology remains poorly understood.
START-UP CONDITIONS. A firm’s growth potential is not exclusively determined by its owners. ‘Insiders’ such as venture capitalists may have an information advantage about firm quality. Indeed, firms funded via venture capital (VC) or other outside investors have been associated with greater growth potential than firms funded through the owners’ capital. The funding amount has also been interpreted as an early signal of venture quality.
While higher initial capitalization may reflect firm quality, care in interpretation is in order. Higher capital intensity within industries may reflect production strategy as much as firm quality and is positively associated with the presence of advanced technologies that tend to increase a firm’s capital-to-labour ratio. Detailed industry controls only partially address this; thus, we leveraged additional information on firm strategy and positioning that is rarely observable at this scale.
INNOVATION AND BUSINESS STRATEGIES. Innovation and related activities geared to exploiting innovation outcomes — in particular, formal intellectual property use — have been correlated with a firm’s growth potential. Recent research suggests that firms that apply for trademarks exhibit higher growth after application. But harder-to-measure complementarities between AI and innovation may also exist, particularly if new applications are difficult to patent or protect by other means.
The survey we leveraged directly queries firms about their recent innovations in new products or services, as well as processes. These innovations, in turn, can be linked to different business strategies. In particular, we can distinguish ‘cost-reducing objectives’ from ‘growth-oriented goals.’ These different strategies may lead, in turn, to distinct AI adoption patterns and outcomes.
GEOGRAPHY. AI is still in a relatively early stage of its geographic diffusion, and its future diffusion path will likely depend on where it is concentrated now, as geographic spillovers in AI use may determine the spread of utilization from current hubs to nearby locations. Therefore, uncovering the current state of geographic disparity in AI adoption was a key objective of our analysis.
The 2018 ABS was sent to 850,000 firms nation-wide based on a sampling method that classifies firms by state, industry and a range of ownership characteristics. Approximately 590,000 firms responded to the survey — a relatively high response rate (over 69 per cent). Roughly 573,000 were linked to the Longitudinal Business Database (LBD).
The vast majority of firms in the U.S. are very small, and this was reflected in our sample. In fact, the ABS sample differs dramatically from standard data sets (like Compustat) that primarily represent large, public firms. While the average firm in the raw sample had roughly 90 employees, this fell to 26 with weighting, and nearly 70 per cent had fewer than 10 employees. Thus, care is required to disentangle young, high-potential ‘gazelles’ from mom-and-pop firms that contribute to the economy in broader, less eye-catching ways.
The technology module in the 2018 ABS contained three novel and connected questions. The first queried firms’ reliance on digital information (data), widely regarded as a key input and even a prerequisite to more advanced uses of digital technologies. A second enabler is sufficient computing power to manage and exploit massive quantities of data. Thus, the second question explored the extent to which firms rely on cloud services for their IT functions, as this has shifted the cost structure and speed of access to IT resources for most firms.
The third question — and the primary focus of our research — asked directly about the use of advanced business technologies in producing goods or services. The specific technology definitions include five commonly associated with AI: machine learning, machine vision, natural language processing, voice recognition software and automated guided vehicles.
Our key takeaways can be summarized as follows.
DIFFUSION OF AI IS LOW, YET IT IS CONCENTRATED IN IMPORTANT SEGMENTS OF THE ECONOMY. Overall, we found AI adoption to be low and concentrated in a smaller number of relatively larger firms. We found that under six per cent of firms actually used an AIrelated technology in their production operations. However, importantly, the majority of very large firms (>5,000 employees) used at least some AI, and the intensity of AI use was higher in these firms.
Readers may note that the level of AI diffusion we found is significantly lower than that reported in the European Commission’s survey of AI and other surveys by McKinsey, Deloitte and PwC. However, those surveys do not claim to be representative of the underlying economy; instead, they focus on larger, publicly traded companies. In contrast, the ABS sample included many small firms for which AI adoption is quite limited, but whose inclusion allows us to estimate the correct ‘denominator’ for country-wide use rates.
The low adoption rate does not make AI irrelevant for the aggregate economy. In fact, the potential exposure to AI by workers was much higher than its usage rate among firms, a difference we estimated using detailed employment data at the firm level. AI use showed up primarily in firms that also had high reliance on digital information and cloud computing, suggesting key interdependencies among recent technology trends. Firms that relied on robotics also typically used AI in production, a relationship that is often missed when one or the other is studied in isolation.
Several impediments to AI adoption appear to limit its diffusion. These include scale effects (many businesses may not possess the vast amounts of data to utilize AI applications), as well as a potential hierarchy of technology adoption needed for proper deployment of AI applications. A possible interpretation of these relationships is that information digitization is a key prerequisite for AI adoption, as is infrastructure such as the cloud or other advanced IT systems.
To the extent that it takes time, resources and potentially economies of scale to accumulate technology complements, considerable frictions to AI adoption and use likely remain. Moreover, a large fraction of firms may also not yet benefit from AI until new applications arise.
The majority of very large firms (>5,000 employees) used at least some AI.
The disparity in adoption rates across firm types suggests an early ‘AI divide’ among businesses in the U.S.: A small set of firms with the special characteristics we uncovered lead the diffusion curve, with a large number of laggards that are yet to overcome the impediments to AI adoption or to benefit from AI. Yet, the prevalence of AI among innovating start-ups with high growth potential indicates that we might expect significant dynamism in these patterns in the years to come. Delving into our subsample of start-up firms is instructive:
FIRM-LEADERSHIP CHARACTERISTICS PREDICT AI USE. More educated, more experienced and younger founders are more likely to adopt AI. Education may be a proxy for the complexity of products and services offered by the firm, which may require a higher degree of education — and, at the same time, advanced technology like AI — to be produced. Serial entrepreneurship, which has been linked to firm performance in prior work, also predicts AI use, suggesting a novel explanation for performance gains in the digital age. Likewise, our finding that firms with owners younger than 35 are more likely to adopt AI is consistent with research that finds younger cohorts of entrepreneurs tend to be more sophisticated users of advanced technology and may be more open to adopting these technologies in their businesses.
High-growth and innovative firms are not only led by individuals motivated by bringing new ideas to market, but also by those who report prosocial values such as helping society. In turn, such motivations are highly correlated with AI use. Other motivations for founding or owning firms matter, but there is a clear distinction between markers of growth-oriented and socalled ‘lifestyle entrepreneurship’ when it comes to AI use.
FIRMS WITH MARKERS OF ‘HIGH-GROWTH’ ENTREPRENEURSHIP ADOPT AI. Higher initial capital levels at start-up and VC funding are associated with higher likelihood of AI use. For instance, the presence of venture capital was associated with a 3.6 per cent increase in AI adoption likelihood, holding everything else constant. Recall that such markers often provide early signals of start-up quality and growth potential.
Being a high-growth firm in the first few years of life is positively and significantly associated with the adoption of AI at some point in the firm’s lifecycle — a firm trait that is directly observable in the rich Census of Bureau. Yet, when all of these observable factors were included together, revenue growth had a far weaker relationship with AI adoption. Firms with patents were more likely to adopt AI (by 3.5 per cent), and firms that identified IP protection as ‘very important’ had a 6.1 per cent higher likelihood of being AI users.
Importantly, once we controlled for early growth, we found that late growth is still robustly correlated with AI use. While our study lacked the ability to establish a causal relationship between AI use and growth, this finding establishes that not only is AI use prevalent among fast-growing firms early in their lifecycle, but that higher revenues later in life are also linked to AI, even when many of the other well-known drivers of growth are removed from the picture.
FIRM STRATEGY MATTERS FOR AI ADOPTION. Our results suggest that a firm’s innovative activity and innovation strategy matter significantly for AI adoption. Both product and process innovations are positively correlated with AI adoption, but the latter has a much higher correlation. Firms with process innovation are 5.4 per cent more likely to adopt AI — a magnitude that is equivalent to the overall adoption rate of AI for the firms. Similarly, firms that possess patents and those that identify IP protection as ‘very important’ are more likely to adopt AI.
GEOGRAPHIC DISPARITY IN THE ADOPTION OF AI IS PRONOUNCED. We found substantial geographic disparity in the adoption and use of AI. Our data are most informative for start-ups, which tend to be single-unit firms and hence easy to situate geographically. In particular, the rate of AI adoption tends to be higher in the southern and western parts of the country. Technology hubs such as Silicon Valley, the Research Triangle and Provo-Orem, as well as large metro areas such as the Capital Region, Los Angeles, Houston and San Francisco stand out in AI adoption. Once we weighted by share of local start-up employment, Riverside, San Francisco and San Jose stood out in California; while Louisville, Columbus, Austin, Atlanta and Nashville also showed concentrated AI use. This has the potential to fuel an AI-based ‘digital divide’ if initial patterns persist.
The complementarities between AI use and measures of firm quality and growth potential suggest that while the diffusion of AI is low — and may remain low for some time — AI’s impact on the economy will be disproportionately large because it is concentrated both among large firms and in a niche of start-ups that are innovative, highly capitalized and fast-growing — the type of firms that contribute significantly to aggregate growth and productivity. While AI adoption is still in its very early stages, its concentration among innovative, high-growth firms and in certain geographies portends not only an opportunity for transformative economic effects, but also ongoing challenges for creating broadly shared prosperity.
Kristina McElheran is an Assistant Professor of Strategy at the University of Toronto Scarborough and the Rotman School of Management. She is also a Faculty Affiliate at the Schwartz Reisman Institute for Technology and Society at UofT; a Digital Fellow in the Digital Economy Lab at Stanford; a Digital Fellow of the Initiative on the Digital Economy at MIT and a Lab Economist at the Rotman School-based Creative Destruction Lab.
Share this article: