Article

Journal of Service Research

2022, Vol. 25(4) 505–520
Artificial Intelligence as a Service, Economic © The Author(s) 2022
Article reuse guidelines:
Growth, and Well-Being sagepub.com/journals-permissions
DOI: 10.1177/10946705221120218
journals.sagepub.com/home/jsr

Christos A. Makridis1,2 and Saurabh Mishra3 

Abstract
The share of artificial intelligence (AI) jobs in total job postings has increased from 0.20% to nearly 1% between 2010 and 2019, but
there is significant heterogeneity across cities in the United States (US). Using new data on AI job postings across 343 US cities,
combined with data on subjective well-being and economic activity, we uncover the central role that service-based cities play to
translate the benefits of AI job growth to subjective well-being. We find that cities with higher growth in AI job postings witnessed
higher economic growth. The relationship between AI job growth and economic growth is driven by cities that had a higher
concentration of modern (or professional) services. AI job growth also leads to an increase in the state of well-being. The
transmission channel of AI job growth to increased subjective well-being is explained by the positive relationship between AI jobs
and economic growth. These results are consistent with models of structural transformation where technological change leads to
improvements in well-being through improvements in economic activity. Our results suggest that AI-driven economic growth,
while still in the early days, could also raise overall well-being and social welfare, especially when the pre-existing industrial
structure had a higher concentration of modern (or professional) services.

Keywords
Artificial intelligence, digital technologies, gross domestic product, productivity, subjective well-being

Introduction high-tech cities like San Francisco and New York take all the
gains from growth in the technology revolution, leaving other
Artificial intelligence (AI) jobs as a share of the total have cities behind (Agrawal, Gans, and Goldfarb 2018; Fernald and
grown from 0.2% in 2010 to roughly 1% in 2019 with sig- Jones 2014; Perrault et al. 2019). But the Covid-19 pandemic
nificant potential for continued growth (Perrault et al. 2019). has called this phenomenon into question as working-from-
Due to this growth, there are differences in opinion about the home continues to expand, spreading the opportunity for AI
consequences of AI on the economy and social welfare jobs beyond high-tech cities (Brueckner, Kahn, and Lin 2021).
(Brynjolfsson and Mcafee 2014; Frey and Osborne 2017). In this sense, understanding the relationship between AI,
optimists believe the diffusion of AI technology will provide productivity, and subjective well-being (SWB) is important as
opportunities for economic growth and also improve the state of federal, state, and local policymakers deliberate on whether and
human well-being. Pessimists share a different view. They how to regulate and support innovation over a potential fourth
believe a small minority of skilled workers (special cities or industrial revolution. Since enlarging the size of the economy
large corporations) will capture the gains from AI technology and improving well-being are important objectives, this paper
leaving other communities more distressed.1 integrates the two strands of literature to study the relationship
Like the revolution in digital skills that were observed over between growth in AI labor demand, economic growth, and
the past three decades (Gallipoli and Makridis 2018), the ex- well-being. This is because economic growth and well-being are
pansion of the demand for AI labor could lead to an expansion
of higher-skilled and knowledge-oriented sectors, leading to a
new wave of structural transformation (Aghion and Jones,
2018). Although AI jobs are growing at a rapid pace 1
Stanford University and Columbia Business School, Stanford, CA, USA
(Perrault et al. 2019), there is significant heterogeneity in AI job 2
National Artificial Intelligence Research, Department of Veterans Affairs,
growth across cities and sectors. There are also increasing Washington, DC, USA
3
concerns about polarization across cities and the incidence of Stanford University, Stanford, CA, USA
productivity gains (Autor 2019; Autor and Dorn 2013;
Corresponding Author:
Hornbeck and Moretti 2020), at least up until the COVID-19 Christos A. Makridis, Stanford University, Stanford, CA 94305-6104, USA.
pandemic.2 Due to this, there is growing public attention that Email: [email protected]
506 Journal of Service Research 25(4)

both efficiency and equity consequences of technological prices respond to scarcity and capital is mobile, labor will retain
progress (Berg and Ostry 2011; Okun and Summers 2015). a comparative advantage (CEA 2019a).
After defining AI and laying out a conceptual framework for Lastly, the paper offers an interdisciplinary and unique data-
why we would expect it to affect economic and social activity, driven perspective to formalize the links between AI, service
we then investigate these claims using empirical data. Our paper revolution, structural transformation, and human well-being
contains two important innovations. First, we have both cross- (Ghani and Kharas 2010; Herrendorf, Rogerson, and
sectional and longitudinal variation obtained by examining AI Valentinyi 2014; Mishra, Lundström, and Anand 2011;
job growth evolution across varied local labor markets, spe- Mishra, Tewari, and Toosi 2020). In classical economic theory,
cifically cities, over time. This allows us to control for national an economy’s growth potential lies in its innate capabilities that
changes and isolate the co-movement of variables at a more are rooted in its production structure. This can be attributed to
granular geographic level. Second, we access data on SWB, the early ideas that viewed growth and development as a process
which transcends the typical measures of GDP or income per of structural transformation in the productive structure of the
capita. Given the concerns about AI and its potential adverse economy (Hirschman 1959; Kaldor 1967; Kuznets and Murphy
effects on social welfare, having direct measures of individual 1966; Lewis 2013), that is, the reallocation of resources from
well-being offers a more complete view of the distributional low productivity to high productivity activities. Services are
consequences of digital transformation. In particular, we pro- typically viewed as a low-productivity sector and a passive
vide the first empirical investigation into the role of AI as a input into the production of physical goods. Yet service ac-
leading driver behind transformative human-center services by tivities can be unbundled, disembodied, and splintered in a
quantifying the increases in SWB through productivity growth. value chain like goods (Bhagwati 1984). Emerging technologies
The paper provides the first quantitative investigation of the like automation, artificial intelligence, and big data are changing
relationship between AI, economic performance, and SWB the characteristics of service. These ideas were originally
using new data on AI job postings from Burning Glass Tech- pioneered at the onset of the modern services revolution as
nologies and Gallup’s U.S. Daily Poll across 343 metropolitan manufacturing began to decline in developed countries, al-
areas (Gallup U.S. Poll 2017; Perrault et al. 2019).3 The fol- though many questions remain (Karmarkar 2004, 2015;
lowing observations were made. First, metropolitan areas with Karmarkar, Kim, and Rhim 2015; Karmarkar and Apte 2007).
greater increases in the share of AI job postings between 2014 Furthermore, rapid urbanization and growth in city clusters,
and 2018 also exhibit greater economic growth. Second, the characteristic of emerging cities, reduce demand for physical
positive relationship between AI job growth and economic goods and increase demand for services. Agglomeration and
growth is explained by the initial concentration of technology- economics of scale in service cities are incentivizing service
based services (or modern service) structure in US cities. Third, hubs of exports even across US cities (Berube 2007), similar to
and most novel, increases in AI jobs are associated with in- the factory-driven industrialization in Asian economies.
creases in SWB (especially physical, financial, and social at- However, the process of structural change can lead to inequality.
tributes of well-being) where the transmission channel of AI job Building on empirical literature in economics that uses data on
growth to increased subjective well-being is explained by the SWB to understand social welfare (Deaton 2012; Kahneman and
positive relationship between AI jobs and economic growth. Deaton 2010), and responding to a call for more research on the
These results counter the prevailing concern that AI and au- relationship between AI and SWB (Musikanski et al. 2020), our
tomation will have net-negative effects on social welfare. Our paper contributes to the ongoing debate about the effects of AI on
explicit measurement of SWB, rather than economic perfor- social welfare. For example, Amartya Sen remarked that “… quite
mance, is important since standard measures of GDP may clearly the notion of living standard and that of utility or happiness
overlook unobserved changes in social welfare.4 are not identical. The former may be typically very important to the
The paper also contributes to the emerging literature on the latter, but they are not the same thing.” (Sen 1988, 1993). Ac-
role of AI in the service economy with broader implications for counting for these dynamic facets of the economic system, our
macro-economic growth and well-being (Huang and Rust 2018; results quantitatively characterize the linkages between AI jobs,
Rust and Huang 2021). Since AI is a service, the paper is also service revolution, economic growth, and well-being.
related to the discussion on the implications of AI and its impact
on the broader economy (Brynjolfsson, Rock, and Syverson
2019). Although AI can have many positive effects on pro-
Definitions and Conceptual Framework
ductivity and the reallocation of tasks (Brynjolfsson et al., 2018; Although we recognize that there is a lot of debate about what
Brynjolfsson & McElheran, 2016), AI and automation can also AI is and is not, we defer to the definition from the Future of
have adverse effects on employment (Acemoglu and Restrepo Artificial Intelligence Act of 2017 (H.R. 4625) that AI is “any
2020). Moreover, if automation displaces tasks at a faster rate artificial system that performs tasks under varying and un-
than new tasks are created, then the economy can converge to a predictable circumstances, without significant human over-
labor share of zero (Acemoglu and Restrepo 2018a, 2019). As sight, or that can learn from their experience and improve their
AI becomes capable of performing intuitive and empathetic performance… They may solve tasks requiring human-like
tasks, some are concerned that it could pose a threat to human perception, cognition, planning, learning, communication, or
employment (Huang and Rust 2018). Nonetheless, as long as physical action.” Artificial intelligence often uses machine
Makridis and Mishra 507

learning to form predictions and make adaptive adjustments Macroeconomic theory suggests that, if AI and automation
based on new information in their environment (Russell and lead to a decline in labor intensity and the overall share of
Peter, 2010). However, it has such a wide array of use cases manufacturing and agriculture, and less productive sectors of
across sectors that many economists view it as a general- the economy begin to expand, then aggregate GDP could be
purpose technology (Agrawal, Gans, and Goldfarb 2019). constrained by what is important and hard to improve and not by
Automation, in contrast, focuses on reducing the labor in- what nations are good at (Aghion, Antonin, and Bunel 2019). If
tensity of a specific process without any stochasticity, whereas the rise of AI and automation is not met with a corresponding
AI/ML involves stochasticity and thus, requires human-like increase in skilled labor, polarization in the labor market could
decision. rise even further (Autor 2014, 2015; Frank et al. 2019). Using
We now lay out the conceptual framework for why we would data on labor demand as a proxy, we examine whether disparate
expect to see a relationship between AI, economic activity, and growth in AI labor demand is yielding broad-based economic
social activity. On one hand, the expansion of AI could raise growth across the sample of metropolitan areas or core-based
productivity growth by increasing the productivity of particularly statistical areas (CBSAs). We also provide empirical evidence
high value-added tasks that constitute a larger portion of aggregate on the relationship between AI growth and economic growth
output than others; such services could include, for example, fi- using the pre-sample share of employment in services to help
nancial and technology services. On the other hand, the expansion explain the linkage.
of AI could lead to stagnation in productivity growth if the requisite Emerging technologies (e.g., automation, artificial intelli-
human capital infrastructure is not built out. That is, if AI tech- gence, and big data) are changing the characteristics of services.
nology expands, without a complementary increase in the required Service activities can be unbundled, disembodied, and splin-
human capital there could be short-term productivity gains that do tered in a value chain like goods (Bhagwati 1984). Such
not translate into the medium to long term. technological changes are driving up demand for complex
CEA (2019b) argues that in theory, the former mechanism will technology-intensive services. The virtual capabilities of this
outweigh, much like the episodes of technological change that we class of services, such as being transported cheaply and swiftly
have seen throughout history. In particular, there are three po- in binary bits, make it more desirable than goods.
tential phases for the introduction of AI: (a) in the short run, We call this class of services modern services or professional
businesses bid wages up as they try to attract high-skilled labor to services, and they include services such as artificial intelligence,
conducive AI-intensive tasks, (bmaking) in the medium run, AI information, or research and development, which are different
substitutes for some of the available demand for labor, and (c) in from traditional services or non-professional services or low-
the long run, investment catches up with AI and real wages tech services, like haircuts and restaurants that require face-to-
continue to grow as the talent pool for high-skilled workers face contact and more manual labor (Baumol 1967; Ghani and
grows. Although the story here is theoretically plausible, em- Kharas 2010; Mishra, Lundström, and Anand 2011).5 In brief,
pirical analysis, especially those related to the modern service modern services are digitally intensive and intangible (Gallipoli
economy, has been limited. One prominent study is Brynjolfsson, and Makridis 2018). Similar to high-value-added manufactur-
Rock, and Syverson (2021). To measure the productivity effects ing, modern services (e.g., AI) may yield greater knowledge
of general-purpose technologies, like AI, it is important to ac- spillovers, have a greater potential for backward and forward
count for complementarity between AI and intangible invest- linkages, or offer an easier pathway toward discovering new
ments, including human capital. In other words, to realize the full specializations with similar characteristics (Loungani et al.
productivity effects of AI, certain investments are needed in the 2017; Spatafora, Anand, and Mishra 2012).
earlier years, but the benefits are not borne out in the data until The old idea of services being non-transportable, non-
later. Brynjolfsson, Rock, and Syverson (2021) find that pro- tradable, and non-scalable no longer holds for a range of
ductivity is 15.9% higher in 2017, relative to reported statistics. modern impersonal services (Ghani and Kharas 2010). Over
These intangibles are especially present in the modern service time, AI will be capable of performing intuitive and empathetic
economy, which exhibits the greatest demand for AI tasks. tasks, enabling innovative ways for human-machine integra-
Developed countries have experienced a significant amount tion that can impact the largest sector of the global economy—
of growth in the post-World War II era. Initially driven by a services (Huang and Rust 2018; Rust and Huang 2021).
move from agriculture to manufacturing (Herrendorf, Scholars have extolled the impact of industrialization of
Rogerson, and Valentinyi 2014), this technological change services before the advent of the 21st century with the out-
led to an increase in the digital workforce across the sourcing and offshoring only part of the revolution (Bhagwati
manufacturing and services sectors (Gallipoli and Makridis 1984; Blinder 2006; Karmarkar 2004). The usage of intensive
2018). Coupled with the rise of globalization (Rodrik 2018), data processing, big data, and AI will become even more
the rise of digital technologies led to an exodus of workers from pronounced in the coming years as AI-based services have a
manufacturing to services. Note that the modernization of the significant impact on the transformation of many industries,
services sector saved it from dragging down aggregate GDP, at including education (Butler-Adam 2018), retail (Teller,
least in the U.S. (Gallipoli and Makridis 2018), where services Kotzab, and Grant 2006), human-resources (Gulliford and
were traditionally known for having lower productivity growth, Parker Dixon 2019), marketing (Rust and Huang 2014), cus-
that is, “Baumol’s cost disease”(Baumol 1967). tomer engagement (Hollebeek, Sprott, and Brady 2021; Paluch
508 Journal of Service Research 25(4)

and Wirtz 2020; Singh and Satish Nambisan, 2021), and at the center of economic policy questions.6 Although there is a
manufacturing (Li et al. 2017). The pace of AI growth has general recognition that GDP per capita and SWB are highly
accelerated over the last decade. correlated (Helliwell, Layard, and Sachs 2012; Stevenson and
Historically, buyers and sellers needed to meet face to face. Wolfers 2008), the correlation can break down, particularly in
Now, many such services can be carried out in one location and lower-income countries where there are many other unob-
consumed in many different places. The internet and other served dis-amenities resulting from weaker institutions and
systems of network technologies like industrial robotics, and poor healthcare infrastructure.
artificial intelligence are providing technical changes to pro- Recent studies have summarized threats and opportunities
duction techniques and business processes. For example, facing community well-being for which AI could have positive
machine translation systems have significantly increased in- or negative ramifications (Musikanski et al. 2020). On one hand,
ternational trade on various online platforms, increasing ex- there are many potential benefits to particular sectors and
ports by 10.9%. This is a result of the substantial reduction in communities, including the automotive sector (Li et al. 2018),
translation costs (Brynjolfsson, Rock, and Syverson 2019). financial services (Wall 2018; West and Allen 2018), medicine
The co-existence of such complex tradable services yields (Topol 2019), education, health, and retail sectors (Breazeal
connections to many other complex activities involving both 2019), among others (Musikanski et al. 2020). On the other
products and services (Mishra, Tewari, and Toosi 2020; hand, there are some concerns, which include the devaluation of
Zaccaria et al. 2018). As AI-based services increasingly be- human skills, declines in self-determination, loss of human
come the main component of all hardware systems, they begin control, and displacement of human responsibility (Floridi et al.
to possess a physical presence like goods which implies that 2018; Musikanski et al. 2020).7 Moreover, there is additional
they can be produced, stored, and consumed across borders evidence that digital technologies, specifically social media and
(Ghani and O’Connell 2017; Loungani et al. 2017). AI-based smartphones have a deleterious impact on youth and the elderly.
services are not only inputs in the global value chain but also These deleterious impacts may include lack of sleep, isolation,
standalone services that could impact growth through their and depression (Twenge and Park 2019). Artificial intelligence
growing tradability. The paper uses dis-aggregated CBSA data also has the potential to compound issues relating to bias,
as an identification mechanism to study AI from the per- privacy, data ownership, personal identity, data governance,
spective of spatial units within a country. The promise of the manipulation, and trustworthiness (Chatila and Havens 2019;
service revolution is that location (or economic geography) IEEE Standards Association 2018).
becomes relatively less important when compared to capital- In addition to productivity—which is reflected in real GDP
intensive manufacturing. The characteristics of cities (in- and per capita income indicators—we consider the effect of
cluding industrial structure, skills, and institutions) are im- AI on SWB. In particular, even if AI leads to greater pro-
portant in understanding how the AI-based service revolution ductivity growth, it could have other unintended conse-
could transform into broad-based economic growth and well- quences on social welfare and the delivery of services,
being. ranging from less social interaction to privacy and security
The globalization of services is the tip of the iceberg (Blinder threats. However, we posit that AI in the long run could
2006; Ghani and Kharas 2010). Services are the largest sector in benefit SWB through its effect on economic activity. For
the world, accounting for more than 70% of global output. The example, if AI leads to productivity growth, as Brynjolfsson,
service revolution has altered the characteristics of services. Rock, and Syverson (2021) predict, then it should pass
Services can now be produced and exported at a low cost through to wages and per capita income (Hornbeck and
(Bhagwati 1984). Technological changes are transforming in- Moretti 2020). Our measure of SWB builds upon a large
tangible services into tangible goods where information tech- literature that has linked the outlook and attitude of re-
nology (IT) serves as a facilitator, enabler, the context, or the spondents about their lives with income across (Clark,
service itself blurring the distinction between goods and ser- Frijters, and Shields 2008) and within (Deaton 2012;
vices (Huang and Rust 2013). Kahneman and Deaton 2010) countries.
Technological changes are transforming intangible services Based on this background, we make the following
into tangible goods where information technology (IT) serves hypotheses.
as a facilitator, enabler, the context, or the service itself H1) As the share of AI jobs increases in a city, productivity
blurring the distinction between goods and services increases.
(Brynjolfsson, Collis, and Eggers 2019; Kahneman and H2) As the share of AI jobs increases in a city, subjective
Deaton 2010). Given the recent divergence between produc- well-being among the average respondent increases.
tivity and median earnings (Stansbury and Summers 2018), H3) The increase in subjective well-being resulting from the
some are concerned about the continued effects of techno- rise of the AI share of jobs is in part a function of the increase
logical change on welfare and polarization in the labor market in productivity.
(Korinek and Stiglitz 2019). Even if AI and automation have In the next section, we begin by introducing the data and
positive effects on productivity, the effects on welfare and the measurement strategy. Then, we take these theoretical hy-
distribution of welfare could be negative. Introducing well- potheses to the data using a sample of CBSAs that we follow
being as a yardstick for measuring the effects of AI puts people over time.
Makridis and Mishra 509

Data and Measurement sample is less likely to cover industries and firms that search
for recruits through physical means, like the placement of a
Artificial intelligence requires several disparate inputs to be put “help wanted” sign. To identify AI jobs, we search the text of
into production. These inputs include human capital, software, the job posting for at least one AI-related skill.9 The overall
data, computational power, and management practices. Since measurement strategy is fairly robust to the exact taxonomy;
firms are still adjusting inputs to AI and experimenting with the see Acemoglu et al. (2020); Alekseeva et al. (2021) as ex-
technology, current observed investments are only noisy indi- amples of studies that have used this taxonomy. Generally,
cators (Tambe 2014). For instance, rapid changes to AI tech- data on job postings have been used a lot within labor eco-
nologies can change the value of investments in AI-related skills nomics to study the demand for skills across space and time
(Rock 2019). The challenges in measuring AI apply to both AI (Hershbein and Kahn 2018). We focus on CBSAs as our
inputs and AI outputs as AI is often intangible and changing at a primary geography of analysis since they best reflect con-
rapid pace. For example, open-source software and data are not solidated labor markets.
transacted through physical markets—so their value is not al- The growth in AI labor demand exhibits two notable trends
ways fully reflected in the national accounts.8 Even beyond across sectors and regions. According to AI Index (Perrault et al.
quantitative measurement challenges, the scientific under- 2019), the United States’ share of AI jobs grew from 0.3% in
standing of cognitive technologies and the microeconomic 2012 to 0.8% of total jobs posted in 2019. Across sectors, AI
processes involved with broader system interactions create labor demand is growing especially in high-tech services and
barriers to a more complete understanding of AI. the manufacturing sector. Across regions, evidence points to no
It is important to note that robotics is not necessarily AI. For clear convergence in AI labor demand growth, that is, many
example, studies show that robot adopters are larger and grow small metropolitan areas with a low initial stock of AI jobs also
faster than their competitors, experience significant declines in experienced fast growth in AI labor demand (2010–19). The
labor share and the share of production workers in employment, growth of AI labor demand in smaller cities and regions of the
and exhibit increases in value-added and productivity US illustrates the potential of AI to generate new types of work
(Acemoglu et al., 2020). Studies on robotic automation show across sectors and regions. Figure 1 shows the percent of AI jobs
that automation has a stronger impact on employment in de- posted as a share of total online jobs posted in 2018–19 across
veloping regions. For example, the impact of robotization on the top 100 metropolitan areas. Artificial intelligence jobs are
employment in emerging economies is more than ten times relatively concentrated in certain pockets of the US. The chart
those of advanced economies. The impact comes from both a presents a yellow–green color spectrum where yellow indicates
direct effect of automation on low-skilled jobs and a reshoring the share of AI jobs as a percent of total jobs in the district is
effect (Barbieri et al., 2021). Additionally, robot per thousand below 0.01%, and the green indicates the share of AI jobs as a
workers leads to a reduction of the employment-to-population percent of total jobs is over 1%.
ratio by 0.18–0.34 percentage points and wages by 0.25–0.50 The data on AI job postings are combined with two addi-
percentage points between 1990 and 2007 (Acemoglu and tional sources. First, traditional measures of economic perfor-
Restrepo 2020). However, by studying a cross-panel of mance from the Bureau of Economic Analysis (BEA), including
countries over the same period, Graetz and Michaels (2018) per capita income, employment, and GDP. We deflate per capita
find that the expansion of robotics led to an increase in labor income by the 2012 personal consumption expenditure index
productivity growth, but only a decline in the employment and GDP by the 2012 implicit price deflator from the St. Louis
share for low-skilled workers. Industrial robots increase labor Federal Reserve. Second, dimensions of well-being and human
productivity, total factor productivity, and wages. Although flourishing from Gallup, the United States’ premier polling
they do not significantly change total hours worked, they may service. We specifically leverage their U.S. Daily Poll, which
be a threat to low- and middle-skilled workers. Economic conducted daily surveys of roughly 1000 U.S. adults on well-
studies of AI should be explicit about defining AI, as many being between 2008 and 2018.10 Gallup’s current polling relies
robotic statistics may not reflect AI. Indeed, many industrial on live interviews (roughly 200 interviewers) with dual-frame
robot shipments have very little (or no) AI in them, which sampling (including random-digit-dial [RDD]) landline and
makes it a poor metric for progress in AI (Shoham et al. 2018). wireless phone sampling with randomly sampled respondents
To address these measurement challenges, researchers are (age 18 or over) from all 50 states and the District of Columbia.
using AI labor demand from job postings (Mishra, Tewari, and The sampling method also uses a three-call design to reach
Toosi 2020; Perrault et al. 2019) and skill taxonomies to respondents who do not pick up on the original attempt.
generate measures of investments in skills related to data and We focus on the respondents who receive the “well-being
algorithmic decision-making since job postings map to labor track,” which captures five components of well-being: career,
resources (Brynjolfsson et al. 2020; Tambe 2014; Tambe, Ye, community, physical, financial, and social. We enumerate each
and Cappelli 2020). of the inputs into these overall indices in Section A.1 of the
We measure AI from the universe of online job postings Online Appendix, but we briefly summarize them here. Career
scraped by Burning Glass Technologies, spanning over well-being focuses on how people enjoy their work, community
45,000 online job sites and representing over 85% of the total focuses on their satisfaction with their city of residence,
labor demand (Burning Glass Technologies 2019). Our physical focuses on their health and medical conditions,
510 Journal of Service Research 25(4)

Figure 1. The heterogeneity in AI job postings across the USA, 2018–2019. Source: Burning Glass Technologies, 2020.

Figure 2. Subjective well-being and real per capita income, 2014–2018. Source: Bureau of Economic Analysis, Gallup, 2014–2018. The figure
plots the association between standardized (mean zero, standard deviation of one) overall well-being and logged real per capita income. Section
A.1 of the Online Appendix describes the specific questions used in these indices in more detail.

financial focuses on their ability to pay bills and purchase what been used to study the relationship between income and hap-
they need, and social focuses on their friendships and optimism. piness within (Kahneman and Deaton 2010) and across
Although these are subjective, and individuals are not always countries (Stevenson and Wolfers 2008).
well-informed about their options (Benjamin et al. 2012), these Figure 2 builds upon these past studies by plotting a binned
measures are reliable (Krueger and Schkade 2008) and have scatterplot across the major CBSAs. Although the relationship
Makridis and Mishra 511

between logged real per capita income and well-being is pos- could also be increases in average productivity (e.g., Graetz
itive, it is far from perfect (only 0.15), suggesting that SWB and Michaels (2018)), if these technologies are primary
likely detects meaningful variation that is not captured by complements to capital and benefit capital holders more
standard macroeconomic indicators, such as income or than laborers. To our knowledge, we are not aware of ex-
GDP. Such data on SWB are especially important when isting work that has related SWB with AI exposure at the
considering the effects of AI and automation because there micro-level.

Table 1. Summary statistics on AI growth, CBSAs, and demographics.

Full sample High-tech services Low-tech services

Mean SD Mean SD Mean SD

AI intensity
AI, % IT 0.031 0.033 0.029 0.021 0.034 0.040
AI, % total 0.0016 0.0022 0.0023 0.0029 0.0010 0.0012
AI, % IT change 0.006 0.029 0.007 0.015 0.005 0.037
AI, % total change 0.000 0.001 0.001 0.001 0.000 0.001
Subjective well-being
Overall SWB 62 2 62 2 61 2
Career SWB 61 2 60 2 61 3
Community SWB 62 4 62 3 61 4
Physical SWB 60 2 61 2 60 3
Financial SWB 61 4 61 3 61 4
Social SWB 61 2 61 2 61 3
Demographics
Population 628,823 1,122,251 1,137,733 1,520,760 239,817 326,495
Male 0.49 0.01 0.49 0.01 0.49 0.01
Age under 18, % 0.24 0.03 0.24 0.03 0.24 0.03
Age 17–34, % 0.24 0.05 0.24 0.04 0.24 0.05
Age 35–54, % 0.28 0.02 0.28 0.02 0.27 0.02
Age 55–64, % 0.11 0.01 0.11 0.01 0.11 0.02
Age 65+, % 0.13 0.03 0.13 0.03 0.13 0.03
White, % 0.80 0.12 0.78 0.11 0.81 0.12
Black, % 0.11 0.11 0.12 0.11 0.10 0.11
Married, % 0.51 0.04 0.51 0.04 0.51 0.04
Some college, % 0.30 0.30 0.30 0.04 0.30 0.04
College, % 0.16 0.04 0.19 0.04 0.14 0.03
Masters, % 0.07 0.02 0.08 0.02 0.06 0.02
Professional, % 0.02 0.01 0.02 0.01 0.01 0.00
Doctorate, % 0.01 0.01 0.01 0.01 0.01 0.01
Industry composition
Information services, % 0.02 0.01 0.02 0.01 0.02 0.00
FIRE, % 0.06 0.02 0.07 0.02 0.05 0.01
Professional services, % 0.09 0.02 0.11 0.02 0.07 0.01
Wholesale/Retail trade, % 0.15 0.02 0.15 0.01 0.15 0.02
Education/Health, % 0.23 0.04 0.23 0.04 0.24 0.05
Economic activity
Real GDP, millions 2012 USD 43 110 84 158 11 13
Income/Capital, 2012 USD 45,211 7162 48,082 7152 43,018 6351
Real GDP growth 0.02 0.03 0.02 0.02 0.01 0.03
Income/Capita growth 0.02 0.03 0.03 0.02 0.02 0.03
Observations 1318 571 747
Height
Notes. Source: Burning Glass Technologies (BGT), Bureau of Economic Analysis (BEA), Census Bureau, 2010–2018. The table reports the means and standard
deviations of the main variables of interest: AI job intensity, subjective well-being (SWB), demographics over the city, industry composition of the city, and economic
activity (deflated in 2012 terms). High-tech refers to whether the CBSA has an employment share in professional services, information services, and FIRE that is
above the median; low-tech refers to whether the CBSA has an employment share in wholesale trade, retail trade, education, and healthcare above the median.
512 Journal of Service Research 25(4)

Table 1 presents summary statistics of our main variables of AI. This is consistent with the large macroeconomic literature
interest. We report the mean and standard deviations of these on the transition of developed countries from agriculture
variables separately for the full sample, the CBSAs with high-tech (Herrendorf, Rogerson, and Valentinyi 2014). Although we see
services, and the CBSAs with lower-tech services. High-tech refers a null association with the manufacturing share, the correlation
to whether the CBSA has an employment share in professional with low-skilled services (retail and wholesale trade) is strongly
services, information services, or finance, insurance, and real estate negative (p =
0.43), and the correlation with high-skilled
(FIRE) that is above the median. Low-tech refers to whether the services (information, professional, and finance, real estate, and
CBSA has an employment share in wholesale trade, retail trade, insurance) is strongly positive (p = 0.60), consistent with the rise
education, and healthcare above the median. Starting with the AI in digital workers and the changing composition of the labor
share, we see that the AI share of IT jobs is 3.1%. Interestingly, market (Gallipoli and Makridis 2018).
CBSAs with more low-tech services have a slightly higher share of The knowledge and technology spillovers of modern ser-
IT jobs. However, the AI share of total jobs is higher in the CBSAs vices are robust. In the absence of such industrial structure and
with high-tech services, but only 0.16% in the aggregate. This endowments, AI-based labor demand could remain an isolated
follows from the fact that many IT jobs reside in the low-tech enclave, rather than serve as a catalyst for economy-wide
services economy, which require fewer cognitive skills (e.g., call growth like we posit in our first hypothesis. In more dis-
center jobs). We also report our measures of SWB, demographics, torted economies, there may be less scope for inter-sectoral and
industry composition of the cities, and economic activity. intra-sectoral resource allocation, as well as knowledge and
Armed with data on the intensity of AI jobs at a local level, technology spillovers. To test this hypothesis (the first hy-
the results identify the links between AI job growth, the service pothesis), the formal regression specification is as follows:
economy, economic growth, and well-being. Figure 3 docu-
yu ¼ γAlu þ f ðXl , θÞ þ fl þ λt þ εu (1)
ments the correlation between various employment shares in
different industries at a CBSA level between 2014 and 2018 and where yu denotes our measure of productivity (log real GDP and
the growth in AI jobs as a share of total jobs between 2011 and per capita income) in location (CBSA) l and year t, Al denotes
2019 for 343 CBSAs. It is not surprising that we see a strong the AI share of jobs, f ðXl θÞ denotes a semi-parametric function
negative correlation (p =
0.21) between the employment share of location-specific characteristics, such as demographics and
in agriculture, mining, and forestry and the employment share in industry composition, f and λ denote fixed effects on location

Figure 3. Employment composition and the 2011–2019 growth in the share of AI jobs. Source: Burning Glass Technologies (BGT) and Census Bureau.
The figure plots the correlation (population-weighted) between the share of workers in the (i) agriculture, mining, and forestry sector; (ii) the
manufacturing sector; (iii) the low-tech (wholesale and retail trade) services sector; and (iv) the high-tech (finance, professional, and information)
services sector and the 2011 to 2019 percentage point change in the share of artificial intelligence jobs in a core business statistical area (CBSA).
 Makridis and Mishra

Table 2. The relationship between the AI share of jobs and productivity.

log(Real GDP) log(Real Income/Capita)

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13)

AI share of total 25.32∗∗∗ 23.80∗∗∗ 7.10∗∗∗ 8.57∗∗∗
16.61∗∗∗ 8.93∗∗∗
9.71∗∗ 9.50∗∗∗ 10.44∗∗∗ 2.77∗∗∗
7.68∗∗∗ 3.28∗∗∗
4.45∗∗∗
[4.64] [5.05] [1.94] [2.09] [3.69] [2.28] [3.90] [2.69] [3.00] [0.99] [1.82] [1.22] [1.69]
× high manufacturing
2.64
[3.89]
× high skilled services 118.20∗∗∗ 52.13∗∗∗
[16.76] [8.41]
× low skilled services
6.25∗∗
1.75
[2.62] [1.59]
× high college 50.84∗∗∗ 21.85∗∗∗
[11.06] [5.01]
R-squared 0.952 0.955 0.999 0.999 0.999 0.999 0.999 0.480 0.535 0.964 0.964 0.964 0.964
Sample size 2565 2565 2563 2563 2563 2563 2563 2565 2565 2563 2563 2563 2563
CBSA controls Yes Yes No No No No No Yes Yes No No No No
Industry shares No Yes No No No No No No Yes No No No No
CBSA FE No No Yes Yes Yes Yes Yes No No Yes Yes Yes Yes
Year FE No No Yes Yes Yes Yes Yes No No Yes Yes Yes Yes
height

Notes. Source: Burning Glass Technologies (BGT), Bureau of Economic Analysis (BEA), 2010–2018. The table reports the coefficients associated with regressions of logged real GDP (in 2012 prices) and logged per
capita income on the share of AI jobs in the CBSA × year. CBSA controls include: logged population, the share male, the age distribution (under age 18, 18–34, 35–54, 55–64, and 65+), race (white and black), the
education distribution (some college, college, masters, professional, and PhD). Industry controls include: the share in agriculture, mining, forestry, the share in construction, the share in manufacturing, the share in
wholesale and retail trade, the share in information services, the share in finance, real estate, and insurance, the share in professional services, and the share in education and healthcare. High manufacturing, high
skilled services (information, professional, and FIRE), low skilled services (retail and wholesale trade, education, and healthcare), and college or more are indicators constructed to equal one if the CBSA share is
above the median in the sample of CBSAs in the matched BGT/Census data. Standard errors are clustered at the CBSA-level.
513
514 Journal of Service Research 25(4)

and year. Although the location-specific characteristics are CBSAs that rank above the median in their share of college degree
collinear with our location-fixed effects, we layer our specifi- holders. These results are consistent with macroeconomic models
cation sequentially. Standard errors are clustered at the CBSA of capital-skill complementarity and skill-biased technical change
level to allow for auto-correlation (Bertrand et al. 2004). (Katz and Murphy 1992).
Table 2 documents the results associated with equation (1). The results are related to economic literature that shows how
Starting with columns 1 and 8, we see that there is a strong changes in the relative prices between sectors can lead to a shift
positive association between economic growth and the AI share in industrial composition (Ngai and Pissarides 2007). If AI is
of jobs: a 1% rise in the AI share of jobs is associated with a 25% more of a complement to higher-skilled and digitally intensive
and 9.5% increase in real GDP and income per capita. Im- services sectors (Gallipoli and Makridis 2018), then the ex-
portantly, these estimated coefficients are obtained after con- pansion of AI jobs could lead to an increase in their share of the
trolling for a semi-parametric function of location-specific economy. Moreover, increases in the skill premium (Buera,
demographic characteristics, such as the age, education, and Kaboski, and Rogerson 2015), particularly for AI and tech-
race distributions. Columns 2 and 9 subsequently introduce nology workers, could also lead to changes in industrial
industry composition controls, but the coefficients are fairly composition.
invariant. We then add CBSA and year-fixed effects in columns Figure 4 plots the raw data across CBSAs grouped using a
3 and 10, which control for time-invariant heterogeneity across binned scatterplot. Panel A shows that there is a correlation of
space and time. Consistent with our concerns about upward 0.22 between the 2014 and 2018 increase in the share of AI jobs
bias, the gradients decline in magnitude, but remain statistically and the growth of overall well-being and Panel B shows that
and economically meaningful: a 1 percentage point rise in the there is a correlation of 0.51 between the 2010 and 2017 in-
AI share of jobs is associated with a 7% and 2.7% increase in crease in the share of AI jobs and the growth in per capita
real GDP and per capita income. income. Admittedly, the correlation could be confounded by a
To understand if these associations only reflect the pro-
ductivity gains that have arisen from the broader digital rev-
olution, that is, those discussed in Gallipoli and Makridis
(2018), we parse our data to identify a broader set of digital
(information technology) jobs, subsequently running a horse
race between the two job shares. When our outcome is logged
real GDP, we find a much lower and statistically insignificant
association of 0.09 (ρ-value = 0.785). We also find a negative
association when our outcome is logged per capita outcome. In
sum, these results show that variation in the share of AI jobs
captures systematically different movements in the data than the
share of IT jobs (the correlation is only 0.54), suggesting that
AI-driven innovation in the modern service economy has dif-
ferent and new prospects for the emerging digital economy.
We now turn toward several dimensions of heterogeneity. We
create indicators for whether the CBSA has above the median
share of workers in the manufacturing sector, the high skilled
services sector (information, professional, and finance, real
estate, and insurance), low skilled services (retail trade,
wholesale trade, education, and healthcare), and college at-
tainment. We do not find evidence of heterogeneous treatment
effects across CBSAs that rank higher in their employment
share of manufacturing workers, but we do find large and
statistically significant effects for CBSAs that rank above the
median with respect to differences in the industrial composition
of services and educational attainment.
For example, a 1 percentage point rise in the AI share of jobs is
associated with a large 118% rise in real GDP and a 52% rise in real
income per capita for CBSAs that have above the median em- Figure 4. Changes in the AI share of jobs, well-being, and income.
ployment share of high skilled services. However, there is a slightly Source: Bureau of Economic Analysis, (Perrault et al. 2019), Gallup,
smaller increase for those that rank below the median. In contrast, 2010–2018. Panel A plots the 2014–2017 growth in CBSA well-being
with the percentage point change in the AI share of jobs across CBSAs
we find that a similar increase in the AI share of jobs is associated weighted by the number of respondents in the CBSA. Panel B plots
with only a slight positive increase in economic activity for CBSAs the 2010-2017 growth in CBSA per capita income with the percentage
that rank above the median employment shares of low-skilled point change in the AI share of jobs across CBSAs weighted by
services. Finally, we find substantially larger effects of AI in employment.
Makridis and Mishra 515

Table 3. The relationship between the AI share of jobs and subjective well-being.
Overall Career Community Physical Financial Social

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

AI share of total 44.89∗∗ 10.36 24.23 27.72 42.30∗∗ 5.70 31.31∗ 34.66∗ 59.94∗∗∗ 9.82 34.99∗∗
18.61
[18.68] [17.34] [20.75] [19.75] [17.69] [19.54] [17.39] [18.63] [20.24] [18.08] [17.54] [18.23]
log(Real GDP) 1.30 0.52 1.33 0.82 0.54 0.47
[0.97] [1.08] [0.82] [0.83] [1.04] [1.22]
log(Real Income/Capita) 1.40 1.68∗ -.34 1.67∗∗ 2.43∗∗ 0.53
[0.95] [1.00] [1.42] [0.85] [1.21] [1.17]
R-squared 0.55 0.63 40 0.48 0.70 0.75 0.55 0.63 0.48 0.55 0.42 0.45
Sample size 1666 1307 1666 1307 1666 1307 1666 1307 1666 1307 1666 1307
CBSA FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Height

Notes. Source: Burning Glass Technologies (BGT), Gallup, Bureau of Economic Analysis (BEA), 2010–2018. The table reports the coefficients associated with
regressions of standardized z-scores of subjective well-being, conditional on CBSA and year fixed effects. Subjective well-being is measured in five ways: career,
community, purpose, financial, and social, which combine (unweighted average) into overall well-being. Section A.1 of the Online Appendix describes the specific
questions used in these indices in more detail.

number of unobserved forces, but we find the raw correlations comes with the added benefit of allowing us to identify how
motivating and useful as a starting point. The regressions that much of the overall effect is driven by productivity gains, like
follow in the subsequent paragraphs endeavor to control for real GDP. We find that the proportion mediated is 77.6%
potentially spurious factors and explore whether these results (ρ-value = 0.062). This suggests that the majority of the dif-
remain robust. ferences we see in subjective well-being are driven by pro-
We use a different regression model to understand the effects ductivity effects, but not all, which offers an even more
of AI on well-being as a broader measure of welfare. Also, this optimistic message.
model controls for economic activity to assess whether the Even with granular controls and fixed effects, there are
impact of AI on well-being is coming through economic ac- still other confounding effects we may not be addressing. The
tivity. The formal regression specification is as follows correlation between AI and SWB declines in statistical
significance after we control for economic characteristics,
SWBu ¼ γAlu þ ζ yu þ fl þ λt þ εu (2) suggesting that economic performance mediates the benefits
where SWB denotes our measure of standardized subjective of AI on SWB. Nonetheless, it is hard to distinguish between
well-being and y denotes our measures of productivity as ad- all the underlying mechanisms at play. For example, there
ditional control variables. could be knowledge spillovers that stem from universities
Table 3 documents the results associated with equation (2). and other innovative environments (Moretti 2004), leading to
Although we find a statistically significant positive association the effective deployment and application of AI in a com-
between increases in the AI share of jobs and well-being across munity. Moreover, there could be complementarities between
every specification without the productivity-based controls certain industrial compositions and specific AI-related skills
(odd columns), these correlations decline for all except (Autor 2015). Furthermore, demographic factors, especially
physical well-being. In particular, a 1 percentage point rise in age distribution, might influence the returns to human capital
the AI share of jobs is associated with a 34SD (standard and the acquisition of AI skills (Acemoglu and Restrepo
deviation) rise in physical well-being. Since the average AI 2018b).
share of jobs is 0.00,186, then the association comes to In unreported regressions, we also explore the relationship
0.063SD when evaluated at the mean. The fact that the cor- between the share of high-tech services and well-being in the
relations decline in statistical significance after adding the cross-section (i.e., without CBSA fixed effects). We find a
productivity-related controls suggests that increases in the AI strong positive correlation of 0.27. The correlation remains
share of jobs affect well-being primarily through productivity strong and statistically significant until the share of college
effects. graduates and/or professional degree holders is controlled for,
To more precisely understand the mediating role of pro- which makes the association statistically insignificant at con-
ductivity on subjective well-being, we use methods from ventional levels. This is consistent with the view that modern
Tingley et al. (2014) who introduced an approach for causal services are associated with greater well-being, but they alone
mediation analysis. In brief, we find a marginal effect of 46.35, are not a causal factor. Indeed, only when they are coupled with
which is statistically indistinguishable from our main result in the introduction of technology do they advance well-being and
the manuscript of 44.59 in column 1 of Table 3. This approach productivity, as we have shown above.
516 Journal of Service Research 25(4)

Discussion and Conclusion identifying and creating a community-accepted definition of

skills/technologies to classify what is an AI job and what is not.
This paper presents the first empirical evidence on the rela- Second, we only consider the channel of AI labor demand in this
tionship between AI jobs, the service revolution, economic study. There are other potential channels to measure AI growth,
growth, and subjective well-being (SWB). The results highlight including publications, patents, papers, and private, or public
the central role that service-based cities play to translate the investment. Third, future research could include attributes of
benefits of AI job growth to subjective well-being. Cities with income/wealth distribution besides well-being. However, data
greater growth in AI job postings also exhibit greater economic and measurement on all these attributes are difficult, especially
growth. The relationship between AI job growth and economic at a local level. We leave these for future research.
growth is driven by cities that had a higher concentration of
modern (or professional) services. Artificial intelligence job
Declaration of Conflicting Interests
growth also leads to an increase in the state of well-being in US
cities. The transmission channel of AI job growth to increased The author(s) declared no potential conflicts of interest with respect to
subjective well-being is explained by the positive relationship the research, authorship, and/or publication of this article.
between AI jobs and economic growth. These results are
consistent with models of structural transformation where Funding
technological change leads to improvements in well-being The author(s) received no financial support for the research, authorship,
through improvements in economic activity. Our results and/or publication of this article.
counter the prevailing concern that AI and automation will have
net-negative effects on social welfare, and suggest AI-driven ORCID iD
economic growth could have a positive effect on well-being—
Saurabh Mishra  https://orcid.org/0000-0002-5949-9446
concentrated in cities with a greater initial industrial structure of
modern services.
Supplemental Material
Policymakers should take these results as a source of opti-
mism about the effect of AI on not only the economy but also Supplemental material for this article is available online.
people. Attempts to curb or over-regulate AI activities may
adversely affect both economic growth and consumer welfare. Notes
Instead, policymakers should continue establishing frameworks 1. See Korinek and Stiglitz (2019) for a detailed explanation over the
for AI so that a market for “AI as a service” can emerge and help distributional consequences of AI on workers and the resulting
organizations and people make better and more effective de- implications for unemployment and inequality.
cisions. A good example involves the principles of trustworthy 2. https://thefederalist.com/2020/07/07/why-millions-of-americans-
AI established in 2020, which articulates a set of ethical might-be-about-to-ditch-city-living/
guidelines for the development and expansion of AI.11 3. Section A.1 of the Online Appendix discusses the items that are
One limitation of these results, however, is that we are still in used to construct the SWB index.
the infancy of AI—that is, the vast majority of jobs still do not 4. Although real GDP remains a fairly strong proxy for overall
require AI skills. There could be a non-linear relationship be- welfare, there are some departures, particularly for less developed
tween economic & social outcomes and the AI share. For countries (Jones and Klenow 2016).
example, when AI job shares are low, the economy and society 5. Traditional economic statistics do not explicitly measure the
boom since it is an important complement to digital services, but technology content of service activities. Inman (1989); Baumol
as the AI share grows, it can create greater inequality and (1967) laid the conceptual foundation for classifying the different
polarization in the labor market. Although we recognize the characteristics of services. Modern impersonal progressive ser-
possibility that higher levels of AI jobs could have different vices are communication, banking, insurance, business-related
effects than the ones we document, we emphasize the role of services, remote access services, transcribing medical records,
continued human capital accumulation in the workforce. As call centers, education, etc. These services differ significantly from
long as individuals are upskilling, that should shield them the traditional personal services, which demand face-to-face in-
from displacement. In conclusion, regions with higher shares teraction. Economists have used data on value added or export
of AI job growth are likely to reap the economic and social statistics to classify modern services as communication, insurance,
benefits that come with it. Therefore, local and national finance, computer & information, royalties and license fees and
policymakers should leverage endowments in their service other business services. Traditional service are Transport, Travel,
capacity to maximize the positive influence of technological Construction, Personal, cultural and recreational services (Ghani
changes to help shape the well-being of citizens for long- and Kharas 2010; Mishra, Lundström, and Anand 2011; Mishra,
term prosperity. Tewari, and Toosi 2020). Similarly, we use CBSA data to label
Nonetheless, some measurement issues still need to be in- modern (high-tech) and traditional (low-tech) services to present
vestigated further to understand the effects of AI on the stylized facts.
economy and society. First, AI is a diverse technology and 6. Moreover, Jones and Klenow (2016) show that there can be de-
contains many fundamental measurement challenges in partures between measures of per capita income/GDP and social
Makridis and Mishra 517

welfare. More research is required to understand the extent that Acemoglu, Daron and Restrepo Pascual (2018b), “Demographics and
economic measures might also fail to reflect differences in well- Automation,” NBER working paper. Oxford Academic. DOI: 10.
being that stem from the adoption of AI and automation, but 1093/RESTUD/RDAB031.
already there is some evidence that benefits are underestimated Acemoglu, Daron and Restrepo Pascual (2019), “Automation and new
(Brynjolfsson, Collis, and Eggers 2019). tasks: How technology displaces and reinstates labor, ” Journal of
7. The welfare effects of automation-inducing technologies have Economic Perspectives. 33, 3-30, DOI: 10.1257/jep.33.2.3.
been debated for many years. For example, Bainbridge (1983) Acemoglu, Daron and Restrepo Pascual (2020), “Robots and jobs:
speculated that automation-inducing technologies could lead to Evidence from us labor markets, ” Journal of Political Economy,
boredom and laziness. Moreover, 73% of the public is worried that 128. 2188-2244, DOI: 10.1086/705716.
robots and computers could do many of the jobs currently done by Aghion, P. and BF Jones, (2018), CI Jones-The economics of artificial
humans (Smith and Anderson 2017). intelligence: An and undefined 2018. “Artificial intelligence and
8. There is, however, a large literature involved in measuring in- economic growth,” nber.org.
tangible capital (Corrado, Haltiwanger, and Sichel 2005; Corrado, Aghion, Philippe, Céline Antonin, and Bunel Simon (2019), “Artificial
Hulten, and Sichel 2009; Corrado and Hulten 2010). intelligence, growth and employment: The role of policy,”
9. These include: AI ChatBot, AI KIBIT, artificial intelligence, IBM Economie et Statistique, 2019. DOI: 10.24187/ecostat.2019.510t.
Watson, IPSoft Amelia, Ithink, virtual agents, machine vision, 1994.
text mining, pattern recognition, image recognition, image Agrawal, Ajay, Joshua Gans, and Avi Goldfarb (2018), “Prediction
processing, computer vision, decision trees, deep learning, machines: the simple economics of artificial intelligence,” Har-
deeplearning4j, Google Cloud machine learning platform, gra- vard Business Press, 250.
dient boosting, H20 (software), Keras, Libsvm, machine learn- Agrawal, Ajay, Joshua Gans, and Avi Goldfarb (2019), “Economic
ing, madlib, mahout, Microsoft Cognitive Toolkit, MLPACK policy for artificial intelligence,” Innovation Policy and the
(C++ library), Mlpy, MXNet, ND4J (software), neural networks, Economy, University of Chicago Press, 19, 139-159. DOI: 10.
object tracking, openCV, Pybrain, random forests, recommender 1086/699935.
systems, Semantic Driven Subtractive Clustering Method Alekseeva, L., J. Azar, M. Gine, and S. Samila, Taska-Labour eco-
(SDSCM), support vector machines, vowpal, wabbit, xgboost, nomics and undefined 2021 (2021), “The demand for AI skills in
Caffe Deep Learning Framework, object recognition, semi- the labor market,” Elsevier.
supervised learning, supervised learning (machine learning), Autor, David H. (2014), “Skills, education, and the rise of earnings
TensorFlow, Torch (machine learning), unsupervised learning, inequality among the ‘other 99 percent, Science. 344, 843-851.
ANTLR, Apertium, automatic speech recognition (ASR), chat- DOI: 10.1126/science.1251868.
bot, computational linguistics, Distinguo, latent Autor, David H. (2015), “Why are there still so many jobs? the history
dirichlet allocation, latent semantic analysis, lexalytics, lexical and future of workplace automation,” Journal of Economic
acquisition, lexical semantics, machine translation (MT), mod- Perspectives. 29, 3-30. DOI: 10.1257/jep.29.3.3.
ular audio recognition framework (MARF), MoSes, natural Autor, David H. (2019), “Work of the Past, Work of the Future, ” AEA
language processing, natural language toolkit (NLTK), nearest Papers and Proceedings, 109, 1-32, DOI: 10.1257/pandp.
neighbor algorithm, OpenNLP, sentiment analysis/opinion 20191110.
mining, sentiment classification, speech recognition, text to Autor, David H. and David Dorn (2013), “The Growth of Low Skill
speech (TTS), tokenization, word2vec. Service Jobs and the Polarization of the U.S. Labor Market, ”
10. Our coverage of CBSAs is fairly comprehensive. Over the 2014– American Economic Review, 103, 1553-1597.
2018 sample, the median number of respondents within a CBSA is Bainbridge, Lisanne (1983), “Ironies of automation,” Automatica, 19,
1271 (mean = 3794). The 10th and 90th percentiles are 537 and 3. DOI: 10.1016/0005-1098(83)90046-8.
7528, respectively. Roughly 83% of the sample is located in a Barbieri, Laura, Chiara Mussida, Mariacristina Piva, and Marco
CBSA. Vivarelli (2021), “Testing the Employment Impact of Automa-
11. https://trumpwhitehouse.archives.gov/presidential-actions/executive- tion, Robots and AI: A Survey and Some Methodological Issues, ”
order-promoting-use-trustworthy-artificial-intelligence-federal- SSRN Electronic Journal. DOI: 10.2139/ssrn.3457656.
government/ Baumol, William J. (1967), “Macroeconomics Of Unbalanced Growth:
The Anatomy Of Urban Crisis,” The American Economic Review,
57, 10.
References Benjamin, Daniel J., Ori Heffetz, and Miles S, (2012), Kimball and Alex
Acemoglu, Daron, David Autor, Jonathon Hazell, and Restrepo Pascual, Rees-Jones (2012), “What do you think would make you happier?
Andreas Mueller, Rob Seamans and Betsey Stevenson (2020), “AI What do you think you would choose?, ” American Economic
and Jobs: Evidence from Online Vacancies,” Cambridge, MA: Review, 102. 2083-2110, DOI: 10.1257/aer.102.5.2083.
NBER Working Paper w28257. DOI: 10.3386/W28257. Berg, Andrew G. and Jonathan D. Ostry (2011), “Equality and effi-
Acemoglu, Daron and Restrepo Pascual (2018a), “The race between ciency,” Finance and Development.
man and machine: Implications of technology for growth, factor Bertrand, Marianne, Esther Duflo, Sendhil Mullainathan, We Thank,
shares, and employment,” American Economic Review. DOI: 10. Abadie Alberto, Daron Acemoglu, Joshua Angrist, Abhijit
1257/aer.20160696. Banerjee, Victor Chernozhukov, Kei Hirano, Imbens Guido,
518 Journal of Service Research 25(4)

Larry Katz, Je Rey Kling, Kevin Lang, Steve Levitt, Kevin Clark, Andrew E., Frijters Paul, and Michael A. Shields (2008),
Murphy, Emmanuel Saez, Doug Staiger, Bob Topel, Cepr “Relative income, happiness, and utility: An explanation for the
Nber, and Nber Nber (2004), “How much should we trust Easterlin paradox and other puzzles, ” Journal of Economic
differences-in-differences estimates?,” Quarterly Journal of Literature, 46. 95-144, DOI: 10.1257/jel.46.1.95.
Economics, 119, 249-275. Corrado, Carol A. and Charles R. Hulten (2010), “How do you measure
Berube, Alan (2007), MetroNation: How US metropolitan areas fuel a ‘technological revolution’?” In American Economic Review.
American prosperity, Metropolitan Policy Program at Brookings. DOI: 10.1257/aer.100.2.99.
Bhagwati, Jagdish N. (1984), “Splintering and Disembodiment of Corrado, Carol, John Haltiwanger, and Daniel Sichel (2005), Mea-
Services and Developing Nations,” World Economy, 7, 17. DOI: suring Capital in the New Economy, Measuring Capital in the
10.1111/j.1467-9701.1984.tb00265.x. New Economy, University of Chicago Press. DOI: 10.7208/
Blinder, Alan S. (2006), “Offshoring: The next industrial revolution?, ” chicago/9780226116174.001.0001.
Foreign Affairs. 85, 113, DOI: 10.2307/20031915. Corrado, Carol, Charles Hulten, and Daniel Sichel (2009), “Intangible
Breazeal, Cynthia (2019), “Living and flourishing with AI,” vimeo. capital and U.S. economic growth,” Review of Income and Wealth.
com/313938302. Wiley, DOI: 10.1111/j.1475-4991.2009.00343.x.
Brueckner, Jan, Matthew Kahn, and Gary Lin (2021), “A New Spatial Deaton, Angus (2012), “The financial crisis and the well-being of
Hedonic Equilibrium in the Emerging Work-from-Home Econ- Americans, ” Oxford Economic Papers, Oxford Economic Papers,
omy?” NBER Working Paper. 64. 1-26, DOI: 10.1093/oep/gpr051.
Brynjolfsson, Erik, Avinash Collis, and Felix Eggers (2019), “Using Fernald, John G. and Charles I. Jones (2014), “The future of US
massive online choice experiments to measure changes in well- economic growth,” In American Economic Review. Wiley, DOI:
being,” Proceedings of the National Academy of Sciences of the 10.1257/aer.104.5.44.
United States of America, 116, 27. DOI: 10.1073/pnas. Floridi, Luciano, Josh Cowls, Monica Beltrametti, Raja Chatila, Patrice
1815663116. Chazerand, Virginia Dignum, Christoph Luetge, Robert Madelin,
Brynjolfsson, Erik and Andrew Mcafee (2014), “The digitization of just Pagallo Ugo, Francesca Rossi, Burkhard Schafer, Peggy Valcke,
about everything,” The second machine age: work, progress, and and Effy Vayena (2018), “AI4PeopleAn Ethical Framework for a
prosperity in a time of brilliant technologies. Goodreads. Good AI Society: Opportunities, Risks, Principles, and Recom-
Brynjolfsson, Erik and Kristina McElheran (2016), “Data in Action: mendations,” Minds and Machines, 28, 1, DOI: 10.1007/s11023-
Data-Driven Decision Making in U.S. Manufacturing, ” SSRN 018-9482-5.
Electronic Journal. 56. DOI: 10.2139/ssrn.2722502. Frank, Morgan R., David Autor, James E. Bessen, Erik Brynjolfsson,
Brynjolfsson, Erik, Tom Mitchell, and Daniel Rock (2018), “What Can Manuel Cebrian, David J. Deming, Maryann Feldman, Matthew
Machines Learn and What Does It Mean for Occupations and the Groh, José Lobo, Esteban Moro, Dashun Wang, Hyejin Youn, and
Economy?, ” AEA Papers and Proceedings, 108. 43-47, DOI: 10. Iyad Rahwan (2019),“Toward understanding the impact of arti-
1257/pandp.20181019. ficial intelligence on labor,” Proceedings of the National Academy
Brynjolfsson, Erik, Daniel Rock, and Chad Syverson (2021), “The of Sciences of the United States of America. DOI: 10.1073/pnas.
Productivity J-Curve: How Intangibles Complement General 1900949116.
Purpose Technologies, ” American Economic Journal: Macro- Frey, Carl Benedikt and Michael A. Osborne (2017), “The future of
economics, 13. 333-372, DOI: 10.1257/mac.20180386. employment: How susceptible are jobs to computerisation? ”
Brynjolfsson, Erik, Prasanna Tambe, Lorin M. Hitt, and Daniel Rock Technological Forecasting and Social Change, 114. 254-280,
(2020), “IT, Intangible Capital , and Superstar Firms,” NBER. DOI: 10.1016/j.techfore.2016.08.019.
NBER Working Paper. Gallipoli, Giovanni and Christos A. Makridis (2018), “Structural transfor-
Buera, Francisco J., Joseph P. Kaboski, and Richard Rogerson (2015), mation and the rise of information technology, ” Journal of Monetary
“Skill-Biased Structural Change,” NBER. NBER Working Paper. Economics, 97. 91-110, DOI: 10.1016/j.jmoneco.2018.05.005.
DOI: 10.1093/restud/rdab035. Gallup U.S. Poll (2017), The U.S. Daily Poll, 2008-2017. Gallup U.S.
Burning Glass Technologies (2019), “Mapping the Genome of Jobs: Poll.
The Burning Glass skills taxonomy,” Burning Glass Technologies. Ghani, Ejaz and Homi Kharas (2010), “The Service Revolution in
Butler-Adam, John (2018), “The Fourth Industrial Revolution and South Asia: an Overview,” The Service Revolution in South Asia.
education, ” South African Journal of Science. 114, DOI: 10. South Asia Region.
17159/sajs.2018/a0271. Ghani, Ejaz and Stephen D. O’Connell (2017), “Can services be a
CEA (2019a), Adapting to Technological Change with Artificial In- growth escalator in low-income countries? ”, OKR. 24.
telligence while Mitigating Cyber Threats, Washington, DC: Graetz, Georg and Guy Michaels (2018), “The Review of Economics
CEA. and Statistics ROBOTS AT WORK,” MIT Press, 100.
CEA (2019b), Economic Report of the President, Washington, DC: Gulliford, Fred and Amy Parker Dixon (2019), “AI: the HR revolu-
CEA. tion,” Strategic HR Review, 18. DOI: 10.1108/shr-12-2018-0104.
Chatila, Raja and John C. Havens (2019), “The IEEE global initiative Helliwell, J., R. Layard, and J. Sachs (2012), “World happiness report.”
on ethics of autonomous and intelligent systems,” In Intelligent Herrendorf, Berthold, Richard Rogerson, and Ákos Valentinyi (2014),
Systems, Control and Automation: Science and Engineering. “Growth and Structural Transformation,” In Handbook of Eco-
DOI: 10.1007/978-3-030-12524-0_2. nomic Growth. DOI: 10.1016/B978-0-444-53540-5.00006-9.
Makridis and Mishra 519

Hershbein, Brad and Lisa B. Kahn (2018), “Do recessions accelerate Lewis, W. Arthur (2013), The theory of economic growth, The Theory
routine-biased technological change? evidence from vacancy of Economic Growth. MIT Press. DOI: 10.4324/9780203709665.
postings,” American Economic Review, 108. 1737-1772, DOI: Li, Bo hu, Bao cun Hou, Wen tao Yu, Xiao bing Lu, and Chun wei Yang
10.1257/aer.20161570. (2017), “Applications of artificial intelligence in intelligent
Hirschman, I. I. (1959), “On multiplier transformations, ” Duke manufacturing: a review,” Frontiers of Information Technology and
Mathematical Journal, 26. DOI: 10.1215/S0012-7094-59-02623-7. Electronic Engineering. 18, 86-96. DOI: 10.1631/FITEE.1601885.
Hollebeek, Linda D., David E. Sprott, and Michael K. Brady (2021), Li, Jun, Hong Cheng, Hongliang Guo, and Shaobo Qiu (2018), “Correction
“Rise of the Machines? Customer Engagement in Automated to: Survey on Artificial Intelligence for Vehicles, ” Automotive In-
Service Interactions, ” Journal of Service Research. 24, 3-8, DOI: novation, 1. 390, DOI: 10.1007/s42154-018-0034-8.
10.1177/1094670520975110. Loungani, MP, MS Mishra, MC Papageorgiou, and K. Wang (2017),
Hornbeck, Richard and Enrico Moretti (2020), The Growth of Low Skill World trade in services: Evidence from a new dataset, Interna-
Service Jobs and the Polarization of the U.S. Labor Market. tional Monetary Fund.
Huang, Ming Hui and Roland T. Rust (2013), “IT-Related Service: A Mishra, S., S. Lundström, and R. Anand (2011), “Service export so-
Multidisciplinary Perspective, ” Journal of Service Research, 16. phistication and economic growth,” World Bank Policy Research
251-258, DOI: 10.1177/1094670513481853. Working Paper.
Huang, Ming Hui and Roland T. Rust (2018), “Artificial Intelligence in Mishra, Saurabh, Ishani Tewari, and Siavash Toosi (2020), “Economic
Service, ” Journal of Service Research, 21. 155-172, DOI: 10. complexity and the globalization of services,” Structural Change and
1177/1094670517752459. Economic Dynamics, 53. DOI: 10.1016/j.strueco.2020.03.002.
IEEE Standards Association (2018), “The IEEE global initiative on Moretti, Enrico (2004), “Workers’ education, spillovers, and pro-
ethics of autonomous and intelligent systems,” IEEE Standards ductivity: Evidence from plant-level production functions,”
Association. https:/1standards.ieee.org/developlindconn/ec/ American Economic Review. DOI: 10.1257/
autonomous_systems 0002828041464623.
Inman, Robert P. (1989), Managing the Service Economy : Prospects Musikanski, Laura, Bogdana Rakova, Bradbury James, Rhonda
and Problems, Cambridge Books. Phillips, and Margaret Manson (2020), “Artificial Intelligence
Jones, Charles I. and Peter J. Klenow (2016), “Beyond GDP? Welfare and Community Well-being: A Proposal for an Emerging Area of
across countries and time, ” American Economic Review, 106. Research, ” International Journal of Community Well-Being, 3.
2426-2457, DOI: 10.1257/aer.20110236. 39-55, DOI: 10.1007/s42413-019-00054-6.
Kahneman, Daniel and Angus Deaton (2010), “High income improves Ngai, L. Rachel and Christopher A. Pissarides (2007), “Structural
evaluation of life but not emotional well-being,” Proceedings of change in a multisector model of growth, ” American Economic
the National Academy of Sciences of the United States of America, Review, 97. 429-443, DOI: 10.1257/aer.97.1.429.
107. 16489. DOI: 10.1073/pnas.1011492107. Okun, Arthur M. and Lawrence Summers (2015), Equality and effi-
Kaldor, N. (1967), “Strategic factors in economic development.”The ciency: The big tradeoff. DOI: 10.2307/2148425.
University of Chicago Press. Paluch, Stefanie and Jochen Wirtz (2020), “Artificial Intelligence and
Karmarkar, Uday (2004), “Will you survive the services revolution? ” Robots in the Service Encounter, ” Journal of Service Manage-
Harvard Business Review, 82, 100-107. ment Research, 4. 3-8, DOI: 10.15358/2511-8676-2020-1-3.
Karmarkar, Uday (2015), “OM Forum- The service and information Perrault, Raymond, Yoav Shoham, Erik Brynjolfsson, Jack Clark, John
economy: Research opportunities,” Manufacturing and Service Etchemendy, Barbara Grosz, Terah Lyons, Manyika James,
Operations Management. 17, 20. DOI: 10.1287/msom.2015.0525. Saurabh Mishra, and Juan Carlos Niebles (2019), “Artificial In-
Karmarkar, Uday S. and Uday M. Apte (2007), “Operations man- telligence Index 2019 Annual Report,” Human-Centered AI In-
agement in the information economy: Information products, stitute, Stanford University.
processes, and chains, ” Journal of Operations Management, 25. Rock, Daniel (2019), “Engineering Value: The Returns to Techno-
438-453, DOI: 10.1016/j.jom.2006.11.001. logical Talent and Investments in Artificial Intelligence, ” SSRN
Karmarkar, Uday S., Kihoon Kim, and Hosun Rhim (2015), “Indus- Electronic Journal. DOI: 10.2139/ssrn.3427412.
trialization, Productivity and the Shift to Services and Informa- Rodrik, Dani (2018), New technologies, global value chains, and
tion, ” Production and Operations Management, 24. 1675-1695, developing economies. NBER.
DOI: 10.1111/poms.12379. Russell, Stuart and Norvig Peter (2010), Artificial intelligence a
Katz, Lawrence F. and Kevin M. Murphy (1992), “Changes in relative modern approach, London: Pearson Education Inc.
wages, 1963-1987: Supply and demand factors, ” Quarterly Rust, Roland T. and Ming Hui Huang (2014), “The service revolution
Journal of Economics, 107. 35-78, DOI: 10.2307/2118323. and the transformation of marketing science, ” Marketing Science,
Korinek, Anton and Joseph E. Stiglitz (2019), “AI and its implications for 33. 206-221, DOI: 10.1287/mksc.2013.0836.
income distribution and unemployment,” University of Chicago Press. Rust, Roland T. and Ming-Hui Huang (2021), The Feeling Economy,
Krueger, Alan B. and David A. Schkade (2008), “The reliability of The Feeling Economy, Springer International Publishing. DOI:
subjective well-being measures, ” Journal of Public Economics, 10.1007/978-3-030-52977-2.
92. 1833-1845, DOI: 10.1016/j.jpubeco.2007.12.015. Sen, Amartya (1988), “Chapter 1 The concept of development,” In
Kuznets, S. and JT Murphy (1966), Modern economic growth: Rate, Handbook of Development Economics. Elsevier. DOI: 10.1016/
structure, and spread, New Haven, CT: Yale University Press. S1573-4471(88)01004-6.
520 Journal of Service Research 25(4)

Sen, Amartya (1993), “Capability and well-being73, ” The Quality of Tambe, Prasanna, Xuan Ye, and Peter Cappelli (2020), “Paying to
Life, 30, 1–445. program? Engineering brand and high-tech wages, ” Management
Shoham, Yoav, Raymond Perrault, Erik Brynjolfsson, Jack Clark, Science, 66. 3010-3028, DOI: 10.1287/mnsc.2019.3343.
Manyika James, Juan Carlos Niebles, Terah Lyons, John Teller, Christoph, Kotzab Herbert, and David B. Grant (2006), “The
Etchemendy, Barbara Grosz, and Zoe Bauer (2018), The AI in- consumer direct services revolution in grocery retailing: An ex-
dex 2018 annual report. Stanford University. ploratory investigation, ” Managing Service Quality, 16. 78-96,
Jagdip, Singh and R. Satish Nambisan 2021, Gary Bridge and Jürgen DOI: 10.1108/09604520610639973.
Kai Uwe Brock (2021), “One-Voice Strategy for Customer En- Tingley, Dustin, Teppei Yamamoto, Kentaro Hirose, Luke Keele, and Kosuke
gagement, ” Journal of Service Research, 24. 42-65, DOI: 10. Imai (2014), “Mediation: R package for causal mediation analysis, ”
1177/1094670520910267. Journal of Statistical Software, 59. DOI: 10.18637/jss.v059.i05.
Smith, Aaron and Monica Anderson (2017), “Americans’ attitudes Topol, Eric J. (2019), “High-performance medicine: the convergence of
toward a future in which robots and computers can do many human and artificial intelligence, ” Nature Medicine. 25, 44-56,
human jobs,” Pew Research Center. DOI: 10.1038/s41591-018-0300-7.
Spatafora, MN, R. Anand, and MS Mishra (2012), Structural trans- Twenge, Jean M. and Heejung Park (2019), “The Decline in Adult Activities
formation and the sophistication of production. IMF. Among U.S. Adolescents, 1976–2016,” In Child Development,
Stansbury, Anna and Lawrence H. Summers (2018), “Productivity and Blackwell Publishing Inc., 90, 638–654. DOI: 10.1111/CDEV.12930.
Pay: Is the Link Broken?, ” SSRN Electronic Journal. . DOI: 10. Wall, Larry D. (2018), “Some financial regulatory implications of
2139/ssrn.3192609. artificial intelligence, ” Journal of Economics and Business, 100.
Stevenson, Betsey and Justin Wolfers (2008), “Economic growth and 55-63, DOI: 10.1016/j.jeconbus.2018.05.003.
subjective well-being: Reassessing the Easterlin paradox,” In West, Darrell and John Allen (2018), “How artificial intelligence is
Brookings Papers on Economic Activity. NBER. DOI: 10.2139/ transforming the world,” UK: cemas.org.
ssrn.1273524. Zaccaria, Andrea, Saurabh Mishra, Masud Cader, and Luciano
Tambe, Prasanna (2014), “Big data investment, skills, and firm value, ” Pietronero (2018), Integrating Services in the Economic Fit-
Management Science, 60. 1452-1469, DOI: 10.1287/mnsc.2014.1899. ness Approach. DOI: 10.1596/1813-9450-8485.