The Demand for AI Skills in the Labor Market

Journal Pre-proof

The Demand for AI Skills in the Labor Market

Liudmila Alekseeva, José Azar, Mireia Giné, Sampsa Samila,

Bledi Taska

PII: S0927-5371(21)00037-3
DOI: https://doi.org/10.1016/j.labeco.2021.102002
Reference: LABECO 102002

To appear in: Labour Economics

Received date: 15 October 2019

Revised date: 6 May 2021
Accepted date: 10 May 2021

Please cite this article as: Liudmila Alekseeva, José Azar, Mireia Giné, Sampsa Samila,
Bledi Taska, The Demand for AI Skills in the Labor Market, Labour Economics (2021), doi:
https://doi.org/10.1016/j.labeco.2021.102002

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.

© 2021 Published by Elsevier B.V.

• Data on skills required in vacancies are used to estimate the demand for AI labour
• Demand for AI labour rose dramatically over 2010-2019 in the U.S. economy
• Larger, more liquid firms with higher R&D demand more of AI labour
• AI skills carry a wage premium of 11% within firm and of 5% within job title
• Firms demanding AI labour more also offer higher salaries in non-AI jobs
 The Demand for AI Skills in the Labor Market∗

Liudmila Alekseeva, José Azar, Mireia Giné, Sampsa Samila, and Bledi Taska

May 5, 2021

Abstract

Using detailed data on skill requirements in online vacancies, we estimate the demand for AI spe-
cialists across occupations, sectors, and firms. We document a dramatic increase in the demand for
AI skills over 2010-2019 in the U.S. economy across most industries and occupations. The demand is
highest in IT occupations, followed by architecture and engineering, scientific, and management oc-
cupations. Firms with larger market capitalization, higher cash holdings, and higher investments in
R&D have a higher demand for AI skills. We also document a wage premium of 11% for job postings
that require AI skills within the same firm and 5% within the same job title. Managerial occupations
have the highest wage premium for AI skills. Firms demanding AI skills more intensively also offer
higher salaries in non-AI jobs.

Keywords: artificial intelligence; machine learning; wage premium; technology diffusion

∗ We thank Hal Varian, Paul Seabright and participants at the 13th Digital Economics Conference at TSE and at the Technis-

che Universität München.

1
1 Introduction

Artificial Intelligence (AI) is possibly the most promising technology currently in development which

has potential applications across a wide range of industries and functions. While it offers the possibility

to increase productivity and create new jobs, it may also replace humans to reduce costs (Acemoglu

and Restrepo, 2019) and potentially lead to lower labor share, increased inequality (Acemoglu and Re-

strepo, 2018b), and disruptions in many industries (Mateos-Garcia, 2019). Hence, its impact is a matter of

immense interest to both economists and policymakers – especially in economies with declining produc-

tivity growth and shrinking working-age populations. Still, little in known about the level of adoption

of AI technology across firms, its penetration within occupations, and its impact in the labor market.

In this paper, we start by providing some key facts on the demand for AI skills in the United States

across occupations and industries and show how it has evolved over the last decade. Most interestingly,

we examine which firm characteristics predict the demand for AI skills to give a sense of the types of

firms that may have adopted AI earlier. Finally, we estimate a wage premium for AI skills and analyse

how wages and skills in other jobs change when the firm demands AI-skilled labor.

Indeed, the extent to which firms use AI remains a black box. The firm-level adoption of AI is hard

to estimate since it would require a specific knowledge on the type of technology the firm uses in the

production function. However, because the application of AI technology by firms requires a highly-

specialized human capital (Gofman and Jin, 2020), the demand for labor with AI skills can proxy for

the use of AI technology. Based on this insight, our approach is to track the demand for AI skills using

data on the online job vacancies collected by Burning Glass Technologies (BGT) which contain nearly a

universe of online job postings since 2010.

BGT provides us with detailed information on the skill set of each vacancy and a broad range of

other details about the job and the employer.1 AI vacancies are identified using a list of skills that are

directly associated with the knowledge of AI technology or the ability to use AI-related software, such as

"artificial intelligence", "machine vision", "deep learning" or "speech recognition". Thus, an “AI vacancy”

is a vacancy that requires at least one AI-related skill from this list. Our main measure of the demand
1 BGT is one of the most comprehensive database of job vacancies in the United States. Due to its detailed nature, it has

been extensively used in prior research on skill requirements across firms and labor markets (e.g., Hershbein and Kahn, 2018;
Modestino, Shoag and Ballance, 2019, 2016; Deming and Kahn, 2018)

2
for AI skills is the ratio of AI job postings over the total number of job postings. The unique qualities of

the BGT data allow us to follow the demand for AI skills across industries, occupations, firms, and even

job titles.

First, we document a fast growth in the demand for AI skills over time. We show that the demand

for AI skills grew 4 times over the period from 2010 to 2019. Firms are striving to hire more employees

with AI-related skills, especially in the last three years, and the demand is still gaining momentum. As

expected, the demand for AI skills is highest in the Information sector, but it is also high in Professional

Services, Administrative and Support Services, and Finance sectors. Among occupations, Computer

and Mathematical occupations present the highest intensity of the demand for AI skills, followed by

Architecture and Engineering and Science occupations. In sum, the data shows that the demand for

AI skills has been growing fast across a wide range of sectors and occupations – it is not limited to the

IT-related industries and jobs.

Second, we explore which firm characteristics explain the demand for AI skills and find that large

firms with high cash holdings and high R&D investments on average demand AI skills in their vacancies

more. These results are consistent with the predictions of a simple theoretical framework that examines

which firm-level characteristics predict the automation with the use of AI algorithms, and consequently

the demand for AI skills. We develop this framework by extending the task-based model of Acemoglu

and Restrepo (2018a). In our framework, firms differ in terms of size, productivity, and the fixed cost of

automation, and thus will tend to automate with AI differently. Specifically, the framework predicts that

large scale firms will have a higher demand for AI-skilled labor, while firms facing higher fixed costs of

automation will demand less AI skills.

Next, we move on to exploring the variation in pay across vacancies and estimate the wage premium

associated with AI skills. Job postings demanding AI skills do offer an 11% premium within the same

firm and a 5% premium within the same job title. We explore how the premium differs across industries

and occupations. As expected from the fast growing demand for AI skills in Computer and Mathemat-

ical jobs, there is a substantial wage premium for AI skills in these vacancies. Indeed, AI is a valuable

IT-related skill. But to our surprise, we find that the premium for AI skills is the highest for Management

occupations, suggesting that in these jobs AI skills carry even more value for employers. This evidence

of an especially high premium for AI skills in Management jobs highlights that the AI’s value may rest

3
in its combination with a broader organizational change.

Finally, we explore whether firms demanding AI-skilled labor show evidence of changing demand

for other, non-AI, jobs in terms of wages and skill requirements. We find that firms demanding AI skills

more intensively tend to offer higher salaries in their non-AI vacancies. This evidence is consistent with

a view that AI technology can facilitate the demand for high-skilled jobs complementing it. To further

explore how non-AI jobs change with the rise in the demand for AI, we examine which skill groups

go hand-in-hand with AI skills in the firm. We find that jobs that require software, cognitive, social,

project management, and people management skills are complementary to AI jobs, while jobs requiring

customer service skills show evidence of being substituted by AI.

The results presented in the paper provide one of the first aggregate pieces of evidence on the dif-

fusion of AI technology in the United States. Overall, our findings on the pervasiveness of the demand

for AI-skilled labor across industries and occupations suggest that AI is likely to be a driver of signifi-

cant transformations in the labor market over the next decades. These results can inform policymakers

deciding on future education and labor market policies as well as workers deciding about investing in

their human capital that is necessary to fit the profile of the employee of the future.

The paper aims to contribute to the literature analysing the effect of new technologies on labor mar-

ket outcomes. Our primary contribution is to the recent empirical literature analysing the effect of AI

technology on labor. Felten, Raj and Seamans (2019) find that occupations impacted by advances in AI

experience a small positive change in wages, and no change in employment. Using the job vacancies

data, Acemoglu et al. (2020) show that AI changes the composition of occupations in firms, but there is

no evidence of a massive job destruction due to AI. Alderucci et al. (2020) document that the number of

firm’s AI-related patents is positively related to the firm’s employment growth rate, but also is positively

associated with the within-firm wage inequality. We complement this literature by providing evidence

on the wage premium for AI skills and on how the intensity of the demand for AI skills is related to

the average wages and skills required in non-AI vacancies. Brynjolfsson, Mitchell and Rock (2018), Fel-

ten, Raj and Seamans (2019), and Webb (2019) build indices showing the impact of developments in AI

technology on the occupations’ task content. The authors are able to identify jobs in which tasks will be

most and least affected by AI. These indices are useful to classify jobs that are more or less likely to be

transformed and replaced by AI technology. We complement this literature by documenting the extent

4
of the actual demand for AI-skilled labor across firms, industries, and occupations.

We also complement the research on the effect of technology on labor by providing a simple theoret-

ical framework helping to understand which firms are more likely to automate with AI and thus more

likely to demand AI skills in their vacancies. In this, we build on the task-based framework of Acemoglu

and Restrepo (2018b). The authors provide a model that shows how machines can have a dual effect on

labor markets. On the one hand, machines can replace human labor in some tasks and thus reduce labor

share, employment, and wages; on the other hand, they can promote the creation of new tasks in which

labor has a comparative advantage and may counteract the replacement effect. This framework with a

specific application to AI is used in Acemoglu and Restrepo (2018a). Other relevant conceptual work on

labor displacement as the result of automation includes Acemoglu and Autor (2011), Brynjolfsson and

McAfee (2014), and Benzell et al. (2015).

This paper also contributes to a recent but fast-growing stream of literature that documents the dif-

fusion of AI technology (Goldfarb, Taska and Teodoridis, 2019, 2020; Blair and Deming, 2020; Acemoglu

et al., 2020). Our descriptive evidence on the demand for AI skills is close to the findings in the paper

of Goldfarb, Taska and Teodoridis (2019) that documents how widespread AI skills are in job post-

ings across industries and occupations. However, our analysis differs substantially from this research.

The authors’ primary focus lies in comparing AI with a few other new technologies and on identifying

whether AI has features of a general-purpose technology (Bresnahan and Trajtenberg, 1995). Also, the

authors abstract from studying the predictors of the AI adoption or the effects that the demand for AI

labor has on other jobs, while our paper addresses these questions. Our descriptive evidence is also

consistent with a supplementary part of the analysis in Blair and Deming (2020) that shows an increase

in the demand for machine learning skills in job postings. However, their research focuses on the jobs’

upskilling after the Great Recession, and hence we do not intersect in the rest of the analysis. Also,

Acemoglu et al. (2020) show a descriptive evidence on the growth of the share of AI vacancies across

industry sectors and across geographies. Their analysis beyond the descriptive part, however, has a

different focus than ours and we consider our findings highly complementary. The authors show that

firms demanding AI skills change the composition of occupations within the firms by hiring less for

jobs at risk of replacement by AI. At the same time, we provide evidence on the wage effects and skills

requirements in non-AI jobs in firms demanding AI specialists.

5
 Finally, we contribute to the research analysing the evolution in the demand for various skills and

premium for skills. We used papers of Deming and Kahn (2018) and Deming and Noray (2019) to define

10 non-AI skills for which we control in our regressions. We contribute to this literature by compar-

ing the premium for AI skills vis-as-vis premia for other skills and by documenting complementarities

between AI and non-AI skills.

This paper proceeds as follows. Section 2 describes our data and sample and outlines the procedure

we use to estimate the demand for AI skills. Section 3 presents the descriptive analysis on the demand

for AI skills in the overall U.S. economy and across industries and occupations. It also shows which

firm characteristics predict the demand for AI skills. Section 4 analyses the wage premium for AI skills.

Section 5 presents evidence on the effect of the firm-level demand for AI skills on non-AI vacancies.

Section 6 concludes.

2 Measuring the Demand for AI skills

2.1 Data

Our main data source is the online job postings database provided by Burning Glass Technologies, an

employment analytics firm. From January 2010 to July 2019, the BGT data contain details on 192.3

million vacancies, including information on job title, standard occupation classification (SOC) code,

name and industry of the employer, and job location. The data also provide detailed information about

the profile of desired job candidates, including education, work experience, and most importantly a list

of skills required from a potential employee. The BGT database as well contains information on salaries

offered in the vacancies. While posted salaries do not represent the actual salaries paid to hired workers,

they are a good indicator of the firms’ willingness to pay for the particular set of skills demanded in the

vacancy.

Carnevale, Jayasundera and Repnikov (2014) conducted a comprehensive analysis of the BGT data

accuracy and representativeness compared to the overall job market. They conclude that the BGT on-

line vacancies correlate strongly with job openings data in JOLTS and do provide detailed employment

demand in a timely manner. The authors warn that BGT may overrepresent job openings for college

graduates and for industries that demand high-skilled workers. As well, Hershbein and Kahn (2018)

6
find that the aggregate and industry trends of the number of vacancies in BGT are consistent with other

sources of job vacancies information such as CPS and OES, in addition to JOLTS. In sum, the BGT data is

representative of the actual demand of skills in the U.S. economy, though we should be cautious about

interpreting results in occupations and industries employing less skilled workers.

It is worth noting some potential limitations of using vacancy data. While the data show the demand

for workers with specific skills, it tell us neither if those positions get filled, nor what happens with the

current employees in the firm – whether they are dismissed or not. What we can observe are changes in

potential hiring and we consider it a proxy for which type of workforce is likely to grow or shrink in the

firm in the near future. Additionally, Babina et al. (2020) explore the similarity between the demand for

AI-skilled employees in the BGT vacancies with the actual employment of such workers observed in the

employment histories database Cognism. The authors find that the two measures of AI skill intensity

- based on BGT and based on Cognism - have a strong correlation and, therefore, the demand for AI

specialists is a good proxy for AI adoption.

Another limitation of our data lies in not fully capturing the extent of AI adoption since firms may

have several means of obtaining AI expertise that will not manifest themselves in job ads: firms can

acquire externally developed AI algorithms or train AI labor in-house. Nevertheless, even when firms

acquire external AI solutions, they often need internal employees who can help to implement these so-

lutions, prepare data inputs, maintain the AI systems, and educate other employees how to use them.

It is also known that the internal development of AI specialists is difficult, and we observe in our data

that all large technology firms that have internal AI training programs, extensively post AI vacancies.

Therefore, we expect that the extent of AI adoption as proxied by the share of AI vacancies might rep-

resent a lower bound of AI adoption. Given the detailed and dynamic nature of the online job postings

data, it still provides a unique source of information allowing us to trace the demand for AI skills and

the willingness to adopt AI technology.

2.2 A Skill-based Measurement Approach

The strengths of the BGT data are in its granular skill description in each vacancy and the timely addition

of the job postings to the dataset. These features allow us to estimate the demand for workers with AI

skills through time and across firms’, occupations’ and industries’ and hence to capture the extent to

7
which firms use AI without knowing the precise technology each firm is applying (Frank et al., 2019).

We identify vacancies requiring AI-related skills (AI vacancies) based on skill taxonomy developed

by BGT, which relies on the classification of nearly 17,000 unique skills obtained from job postings, re-

sumes, and other sources into non-overlapping categories. BGT’s skill taxonomy development process

is described in Burning Glass Technologies (2019). Overall, BGT identifies AI-related skills based on

the presence of words and phrases commonly associated with the knowledge of AI (e.g., artificial in-

telligence, machine vision, deep learning, image processing, speech recognition, etc.) and AI-related

software/packages (e.g., IBM Watson, TensorFlow, Pybrain, Random Forests, ND4J, etc.) in skill re-

quirements part of a job description. Our goal is to define AI skills in the most straightforward and

encompassing manner. Appendix Table A1 provides a complete list of skill requirements we use to

identify vacancies demanding AI. Goldfarb, Taska and Teodoridis (2019) use an identification of AI-

related jobs that is similar to ours, but include only the most general AI-related skills in their definition.

We build our main measure of the demand for AI skills in firm i and time t as the ratio of job postings

requesting AI skills over the total number of job postings2 :

Number of Job Postings requesting AI skillsi,t

AI Sharei,t = (1)
Total Number of Job Postingsi,t

Our measure of AI Share based on skills as a proxy for AI adoption is complementary to other

measures in the recent literature. Webb (2019) and Alderucci et al. (2020) use patent descriptions and

Felten, Raj and Seamans (2019) and Brynjolfsson, Mitchell and Rock (2018) use surveys to predict which

occupations and job tasks are likely to be transformed by AI. These alternative approaches are based on

mapping firms’ occupational structure to the advances in AI technology.3

The BGT job postings can be characterized by ten additional skills which the literature has identified

as relevant to explain pay differences across labor markets and performance differences across firms

(e.g., Deming and Kahn, 2018). We use these skills as controls in the regression analysis of wages. To

categorize skills in these ten groups, we use the definitions of the skill groups from the Appendix of

Deming and Noray (2019).

2 AI Share for the overall economy and the specific industry or occupation is calculated as the number of AI vacancies in the

economy/industry/occupation divided by the total number of vacancies in the economy/industry/occupation respectively.

3 Brynjolfsson, Mitchell and Rock (2018) acknowledge that their approach may have shortcomings since their method fo-

cuses on the technological feasibility of AI adoption but does not account for the various economic, organizational, cultural,
and legal factors that can influence whether firms in fact implement AI.

8
2.3 Sample Characteristics

Our sample consists of all job vacancies excluding internships from 2010 until 2019 which encompasses

a total of approximately 190.2 million vacancies. Panel A of Table 1 presents summary statistics for

the full sample of job postings used in the descriptive part of the analysis. In our sample, AI skills are

demanded in 0.4% of the total vacancies, while Software skills are demanded in 21.5% and Cognitive

skills in 28.3% of the vacancies.4

Next, we show which skills tend to be demanded in the job postings when AI skills are also required.

Panel B of Table 1 displays the probability of demanding different skills conditioning on each of the

remaining skills.5 For example, in postings that demand AI (column 1), Software skills are required four

times more frequently than in an average job posting in the sample. Cognitive skills are 2.5 times more

frequently requested than on average. Thus, the demand for AI skills in postings is most frequently

accompanied by the demand for Software and Cognitive skills, while Customer Service is the least

frequently required skill together with AI. More generally, this table as well allows us to compare the

probability of skill demand across postings conditionally on different skills. For example, in job postings

that demand Software skills (column 2), Cognitive skills are nearly 1.8 times more frequent than on

average in the sample, though this is lower than the demand for Cognitive skills in AI postings. In terms

of the demand for Cognitive skills, jobs requiring Project Management skills (column 9) are closest to AI

jobs – they require Cognitive abilities around 2.1 times more frequently than the overall sample does on

average.

To answer the question about which firms demand and potentially adopt AI, we study firm char-

acteristics associated with the firm-level AI Share. We match our vacancies sample to Compustat data

based on the employer’s name. We drop observations with missing or negative Total Assets and Sales in

Compustat and keep observations with non-missing data for all Compustat-based variables necessary

for the regression analysis. The resulting sample includes 4,121 unique Compustat’s GVKEYs that are

matched to the BGT data in the 2010-2018 period and have complete data. Panel C of Table 1 shows

financial and operational characteristics of the firms in the Compustat-matched sample. All monetary
4 For the the definition of the 10 non-AI skills we rely on Appendix of Deming and Noray (2019). For instance, Software skills

are skills labeled by BGT simultaneously as "software skills" and "specialized skills". Cognitive skills are the ones containing
"solving", "research", "decision", "thinking", etc. in their description.
5 The results of this analysis are also supported by Table A2 in Appendix that uses logit regressions to show how the

occurrence of a specific skill in vacancies can predict the occurrence of other skills.

9
variables in Compustat are adjusted for inflation and log-transformed, and ratios are winzorised at 5%

and 95% levels (we also check the robustness of the results to winzorising ratios at 1% and 99%). On

average, AI Share in the Compustat-matched sample is 0.58%, and nearly 24% of firm-year observations

in the Compustat-matched sample demand AI skills.

3 The Demand for AI Skills

Figure 1 shows the evolution of the demand for AI skills over the period 2010-2019. Job postings that

require AI skills have grown dramatically over this period, both in the absolute number and as the

proportion of the overall number of vacancies posted. The number of AI job postings has increased

from 20.6 thousand in 2010 to 180.9 thousand in 2018, and based on the projection of the data available

for the first seven months of 2019 can be expected to reach nearly 220 thousand vacancies in 2019. The

proportion of job postings requiring AI skills relative to the total number of posted vacancies grew four

times over the period 2010-2019. AI Share of vacancies in the overall economy increased from 0.18% in

2010 to 0.63% in 2018 and was equal to 0.72% in the first seven months of 2019.

To check that this trend is not driven by the addition of AI-hiring firms to the dataset over time, we

also looked at the demand for AI skills only among firms that were present every year of the 2010-2019

period. Figure A1 in Appendix shows the demand for AI skills over time for this subset of firms. The

increase in the share of postings requiring AI skills for this balanced panel of firms is similar to the

full sample, though it is slightly faster - there was a fivefold increase in AI Share in this sample over

the analysed period. Therefore, this rising trend is not driven by the changes in the universe of firms

included in the BGT data through time.

It is important to benchmark the growth in the demand of AI skills relative to the demand for other

computer skills. Figures 2(a) and 2(b) show analogous charts of the trends in the number and the share

of job postings requiring general Computer skills and specialized Software skills over time. The share

of job postings mentioning Computer skills is relatively flat over the period 2010-2019, while the share

of job postings mentioning Software skills declined slightly over this period.

In spite of the dramatic growth, the demand for AI may seem small as the share of AI vacancies is

only 0.72% in 2019 in the overall BGT data. To evaluate the relevance of the demand for AI skills, we

benchmark it both with the demand for skills related to several new technologies analysed in Goldfarb,

10
Taska and Teodoridis (2019) and with the demand for specific IT skills. We find that robotics is almost

twice less frequently required than AI-related skills in the BGT vacancies, the internet-of-things is 3

times less required, and nanotechnology is 25 times less frequently required than AI skills. Comparison

with popular IT-related skills shows that SQL is 5 times more frequently required and Python is 2.5

times more frequently required than AI-related skills. This comparison shows that the knowledge of

AI is a rather specific skill, still it is just 2.5 to 5 times more “specific” than the knowledge of popular

programming languages.

Finally, to get a sense of the most frequently demanded AI-related skills in vacancies, we plot the top

15 skills based on the last month of our data, July 2019. Figure 3 shows that “Machine learning” is the

most frequently mentioned skill in AI vacancies. As a subset of the artificial intelligence field, machine

learning is used to build mathematical models based on large data sets in order to make predictions

without explicitly programming the task and has a variety of applications in different settings. Other

more specific machine learning skills such as "Deep Learning" and "Image Processing" are also at the

top of the list. The second most common skill is simply “Artificial intelligence”. “Natural language

processing” is the third most demanded skill, followed by “Deep learning” and “Image processing”.

Interestingly, “IBM Watson” is the only AI-related skill in the top-15 list that relates to a specific AI

platform for business and not to a general skill category or a scripting language library, such as “Keras”.

IBM Watson is the information search and analytics solution developed by IBM that has applications in

many industries, including agriculture, health, and finance, among others.

3.1 AI Share across Industries and Occupations

In this section, we explore the presence of AI vacancies across different industries and occupations as

a proxy for AI technology adoption. Figure 4 shows AI Share for broad industry sectors defined by

the 2-digit NAICS codes. As expected, Information industry is the one demanding most AI skill, on

average 2.2% of vacancies in this sector are AI vacancies in 2019. Professional, Scientific, and Technical

Services sector has AI Share just below 2% and the next groups of industries are Finance and Insurance,

Administrative and Support Services, Agriculture, Forestry, Fishing and Hunting, and Manufacturing –

all with AI Share around 1%. The share of AI vacancies increased substantially in most 2-digit industries,

especially in the last three years, which is in line with the AI Share growth trend for the overall economy

11
in Figure 1. In sum, the demand for AI skills is present across a wide array of industries – even Mining,

Educational Services, or Public Administration are displaying an increasing demand for such skills.

Figure 5 shows AI Share across 2-digit SOC occupations. Computer and Mathematical occupations

group has the highest AI Share with 4% of vacancies demanding AI skills. As well, other occupations

are also increasingly posting AI vacancies: Architecture and Engineering, Life/Physical/Social Science,

Management, and Legal occupations are among the top five occupation groups with the highest AI

Share. Even occupations with almost zero AI Share at the beginning of our period, such as Protective

Services or Farming, Forestry, and Fishing, are also requiring a notable share of AI vacancies by 2019.

These results are consistent with a few recent studies analysing the diffusion of AI using job vacancies

data (Goldfarb, Taska and Teodoridis, 2019, 2020; Blair and Deming, 2020; Acemoglu et al., 2020). The

results are also in line with actual employment profiles-based measure of AI relatedness of jobs (based

on Cognism data) across industries in Babina et al. (2020).

3.2 Theoretical Framework

In the previous subsection, we have shown that AI skills are increasingly demanded across a wide array

of occupations and industries. AI is becoming a sought-after skill beyond computer-related occupations

or the Information sector. Given the wide and fast-growing pace of the demand for AI skills across

occupations and industries, we want to examine what type of firms are demanding AI skills.

We use a simple theoretical framework to formally explore and generate testable predictions on

which firms might be more likely to adopt AI. We build on the task-based framework of Acemoglu

and Autor (2011) and Acemoglu and Restrepo (2018b).6 We extend this model by adding a downward-

sloping demand for the firm’s products and a fixed cost of automation with AI in order to be able to

assess the role of firm characteristics in AI adoption process. For example, automation of additional

tasks with AI might require investment in hardware, such as servers, to increase the computational

capacity or in the acquisition of externally produced AI algorithms. Due to adding the fixed cost of

automation, the profit maximization problem of the firm changes since the allocation of factors to tasks

will depend not only on the relative cost of labor and algorithms but also on the ability to cover the fixed
6 Using this framework, Acemoglu and Restrepo (2018a) show how various automation technologies, and AI in particular,
can displace labor from tasks. Acemoglu et al. (2020) empirically test this framework in the setting where AI algorithms
substitute for labor in automatable jobs.

12
costs of automation. Thus, the framework allows us to predict which firms are more likely to replace

labor with AI and hence are more likely to demand AI skills in their vacancies.

In a task-based framework (Acemoglu and Autor, 2011), the aggregate output is produced by com-

bining the output of tasks located on a unit interval [0,1].7 The production function is assumed to be

Cobb-Douglas with a unit elasticity, the environment is static with a unique final good:

h Z 1 i
Q = exp z ln y( x )dx , (2)
0

where Q denotes aggregate output, y( x ) is the output of task x, and z is the firm’s total factor produc-

tivity (TFP).8

In the model, each task can be performed by non-AI labor l ( x ) or by algorithms a( x ). We assume that

one unit of algorithms is produced by some fixed proportions of AI labor and machines. Therefore, tasks

x ∈ [0, I ] are technologically automated and can be produced by non-AI labor or by the combination

of machines and AI-skilled labor9 , while tasks in x ∈ ( I, 1] are not automated and produced only by

non-AI labor:




γL ( x )l ( x ) + γ A ( x ) a( x ) if x ∈ [0, I ]
y( x ) = (3)


γL ( x )l ( x ) if x ∈ ( I, 1],

where γL ( x ) is the productivity of non-AI labor in task x which is assumed to be strictly increasing,

γ A ( x ) is the productivity of algorithms produced by the AI-skilled labor in automated tasks. Following

Acemoglu and Restrepo (2018a), we also assume that γL ( x )/γ A ( x ) increase in task index, so non-AI

labor has a comparative advantage in higher-indexed tasks. In the model, I is a threshold task which

characterises the range of tasks that can be automated using existing technologies which are given ex-

ogenously.

Again, as in Acemoglu and Restrepo (2018a), we assume γL ( I )/γ A ( I ) < w/r (Assumption 1), where

w is a cost of non-AI labor and r is a cost of algorithms. This assumption implies that all tasks below
7 We are not considering the changes in the range of tasks due to the introduction of new tasks as in Acemoglu and Restrepo

(2018b) for simplicity.

8 It means that firms with higher z are more productive in all tasks. Addition of term z is a modification of the production

function notation in Acemoglu and Restrepo (2018a).

9 It means that we classify all AI jobs as directly participating in automation process and displacing non-AI labor from their

tasks.

13
I will be performed by algorithms. Since the price of task, p( x ), equals the unit cost of its production

(tasks are produced competitively), the unit cost of production for tasks x > I equals the effective cost of

labor and the unit cost of production for tasks x ≤ I equals the effective cost of algorithms:



 r
 i f x ∈ [0, I ]
γA (x)
p( x ) = (4)

 w
 i f x ∈ ( I, 1].
γL ( x)

Thus, the threshold task I also represents the task for which the firm’s costs of producing with algorithms

and with non-AI labor are equal. In the rest of the discussion we also assume that the allocation of factors

is not constrained by technology, meaning that the firm can endogenously choose a cost-minimizing

factor allocation to tasks.10 We will denote an equilibrium (endogenous) threshold task by I ∗ .

Then, the firm maximizes profits:

π = P( Q) Q − Qµ − f I ∗ , (5)

where P( Q) is an output price schedule that we assume to have form αQ−1/θ (α and θ are constants), Q is

aggregate output, µ is a marginal cost of production, f is a fixed cost of automation expansion described

above, and I ∗ is the equilibrium threshold task.

Our goal is to understand how I ∗ changes with respect to firm heterogeneity. In particular, we con-

sider two firm characteristics that can determine the firm’s automation threshold choice: firm size that

is determined by both the level of the market demand-shifter parameter, α, and the firm’s productivity,

z, and the fixed cost of automation, f .

Model’s solution presented in Appendix provides us with predictions about the relationships we are

interested in. We summarize them as follows:

Proposition 1. Consider the task-based model of automation developed above, in which firms face downward-

sloping demand curves and a cost of automation. The (firm-specific) automation threshold I ∗ is non-decreasing in

the productivity parameter z and the demand-shifter parameter α, and non-increasing in the cost of automation f .
10 This
case corresponds to one of the two scenarios considered in Acemoglu and Restrepo (2018b). In the second scenario,
when factor allocation is constrained by the available automation technology, firms will have to take the factor allocation that is
allowed by the technology instead of a cost-minimizing option. This scenario therefore does not allow to generate predictions
about how firm heterogeneity affects AI adoption.

14
 Thus, the framework shows that firms facing a higher market size and firms with higher productivity

are likely to automate more. At the same time, firms that face higher fixed costs of automation are likely

to automate less. How do we translate the model predictions into the predictions about the relationships

between the firm’s intensity of using AI and firm characteristics?

Empirically, both parameters of the model, market demand α and productivity z, are associated

with the firm’s scale. Both firms facing a larger market demand and having a higher productivity have

strong incentives to grow. Therefore, based on the model predictions, we hypothesise that larger firms

will automate more. Empirically, we measure firm size in various ways: via sales, employment, the

number of vacancies, and market capitalization.

Fixed costs of automation in turn can represent various financial as well as organizational costs of ex-

panding the range of tasks performed by the algorithms. Brynjolfsson, Mitchell and Rock (2018) name

technological, economic, organizational, cultural, and legal factors that can affect the adoption of AI

technology. A straightforward example of such cost is a financing constraint. Financially constrained

firms have a higher cost of capital and thus have a higher costs of attracting financing for new invest-

ments. For example, Ersahin (2017) shows that a better access to borrowing leads to firms adopting

more efficient production technologies, while Gorodnichenko and Schnitzer (2013) showed that finan-

cially constrained firms innovate less and thus cannot catch up to the technological frontier. Hence, we

expect firms facing higher financial constraints, such as higher leverage, to be lagging in the introduc-

tion of new automation technologies using AI and thus demand AI specialists in their vacancies less

intensively. Conversely, higher cash holdings create a higher financial flexibility for a firm and thus

are associated with a higher demand for AI skills. Finally, firms that extensively invest in R&D can be

expected to have lower costs of introducing new automation technologies, since they are more open to

innovations. AI algorithms is an example of a new technology that promises substantial productivity

gains and a boost in firms’ innovation capabilities, thus we expect that firms with higher innovation

efforts will demand AI skills more.

3.3 Firm Characteristics and the Demand for AI

The framework predicts that firms characterised by a larger scale are likely to automate their tasks with

AI more and thus will have a higher demand for AI-skilled labor. Meanwhile, firms that face higher fixed

15
costs of automation are likely to automate their tasks with AI less, and thus their demand for AI skills

will be lower. We use these predictions to motivate our empirical tests using the Compustat-matched

sample of firms for which we have data on a range of financial and operational characteristics.

We explore the relationship between AI Share and firm characteristics using the following specifica-

tion:

AI Sharei,t+1 = β 1 Firm Chari,t + γs + ζ t + ε i,t , (6)

where AI Share is the percentage of job posting demanding AI skills over the total number of job postings

in year t + 1, as defined in equation (1), Firm Chari,t is the vector of Compustat-based financial and

operational characteristics and the logarithm of BGT-based total number of job postings of firm i in year

t, γs and ζ t are 2-digit NAICS industry and year fixed effects, and ε i,t is an error term.

Table 2 shows the results of the regression analysis weighted by firm employment, with the firm-

level AI Share as the dependent variable. All independent variables in the regression are lagged by one

year. We include independent variables one-by-one and then report results of the multivariate regres-

sion: column (1) shows the results for the log market capitalization, column (2) the log of employment,

column (3) the log sales, column (4) the market-to-book ratio, column (5) the return on assets, column

(6) the cash holdings ratio, column (7) the book leverage, column (8) the R&D expenses over sales, col-

umn (9) the capital expenditures over assets, column (10) the PP&E over assets, column (11) the log of

total vacancies, column (12) presents the multivariate specification. Columns (11) and (12) have fewer

observations due to the lagged logarithm of vacancies not available in 2009 and occasionally missing in

other periods; excluding the lagged logarithm of vacancies from column (12) does not affect the signifi-

cance of the coefficients - the results are not driven by the decrease in the sample size. In our data, sales,

employment, the number of vacancies, and market capitalization measure various aspects of the firm’s

scale parameters in the model. At the same time, leverage measures the firm’s fixed costs of automation,

while cash holdings ratio and R&D intensity (as R&D expenditures over sales) are the various measures

of the firm’s flexibility which decreases the automation fixed costs, as discussed in the previous subsec-

tion. As well, we include several firm characteristics without making specific predictions about their

effect.

The results show that there is a positive association between different measures of firms’ size – that

is, market capitalization, employment, and sales – and the firm-level demand for AI skills. As well,

16
cash holdings and R&D intensity are positively associated with the demand for AI skills. Finally, there

is also a negative association with tangible assets as proxied by PP&E over assets and AI Share. These

results are in line with Babina et al. (2020) who show that growth in AI adoption is positively related

to the initial firm’s sales, cash holdings ratio, and R&D intensity. Column (12) shows that the positive

association between AI Share and market capitalization, cash holdings and R&D intensity is robust even

when we control for all relevant firm characteristics in the multivariate specification. In terms of the

economic significance, the coefficients in column (12) suggests that a one standard deviation increase in

the logarithm of market capitalization is associated with a 0.34 percentage points increase in AI Share,

or nearly a 60% increase in the sample mean of AI Share. As well, cash rich firms on average have

a higher AI Share. A one standard deviation increase in the cash-to-assets ratio is associated with a

0.28 percentage points or nearly a 50% increase in average AI Share. Finally, a one standard deviation

increase in the R&D intensity is associated with a 1.25 percentage points increase in AI Share. This effect

is also economically meaningful, as it represents over a 200% increase relative to the sample mean AI

Share.

Figure 6 shows the evolution of the average AI Share by the groups of firms based on three important

characteristics: their size, R&D expenditure, and cash holdings. The figures are consistent with the

regression results – the largest firms, firms with positive R&D, and high-liquidity firms have the largest

AI Share during the whole analysed period. Interestingly, with time, the gap between the top and

middle/bottom group is increasing.

These results are robust to a linear probability regression specification where we explore the relation-

ship between the probability that a firm will search for AI specialists and its financial and operational

characteristics. To use the variation in the requirements for AI skills across labor markets, we measure

whether the firm is searching for AI specialists in each 6-digit SOC-CBSA labor market. We estimate

the association between the probability to post AI vacancies in the labor market and firm characteristics

using the regression model similar to equation (2) but now with a dummy variable equal one if firm

i posted at least one vacancy demanding AI skills in the 6-digit SOC-CBSA labor market sm in year t

and zero otherwise (AI Dummyi,sm,t+1 ) as the dependent variable and add the 6-digit SOC-CBSA labor

market fixed effects. Again, Table A3 shows a positive association between the firms’ size, cash-to-assets

ratio, and R&D over sales and the likelihood to demand AI skills.

17
4 The AI Wage Premium

We have shown that AI skills are increasingly demanded across a wide range of sectors and occupations.

As well, we documented that firms with a larger market capitalization, higher cash holdings, and higher

investments in R&D on average demand AI skills more intensively. Our next goal is to explore the

variation in pay across vacancies and to estimate the wage premium associated with AI skills.

If jobs requiring AI skills are expected to be more productive due to these skills, then such vacancies

will offer a significant wage premium. As well, we may observe differences in the AI skills premium

across different occupations. For example, managers with AI knowledge might focus on the strategy

of implementing AI in their organization and adapting the current operations of the firm. In contrast,

employees in computer occupations with additional AI skills might be devoted to writing scripts and

developing AI systems within the organization. Hence, these two occupations requiring AI can be

expected to deliver differential benefits to the employer. We can rely on employers preference for skills

and the corresponding salaries offered in vacancies to estimate the expected benefits from these hires.

We estimate the wage premia using standard regressions of wages on employee skills and use job

posting-level data from 2016 until 2019 to capture the most up-to-date effect. We estimate the following

regression:

Log(Wagev,i,sm,t ) = β 1 AIv,i,sm,t + β 2 Other Skillsv,i,sm,t + γi + δsm + ζ t + νv,i,sm,t , (7)

where Log(Wage)v,i,sm,t is the logarithm of salary offered in the job vacancy v posted by firm i in the mar-

ket sm and year-month t, AIv,i,sm,t is the indicator variable equal one if the vacancy requires at least one

AI-related skill and zero otherwise, Other Skillsv,i,sm,t is the vector of ten indicator variables that equal

one if the vacancy requires one of the ten skills we use as controls following prior literature (Deming

and Kahn, 2018) and zero otherwise, γi , δsm , and ζ t are firm, market (6-digit SOC-CBSA labor market),

and year-month fixed effects, and νv,i,sm,t is an error term.

Table 3 presents the estimated premia for AI and the set of skills that recent literature has deemed

relevant.11 We are especially interested in the premium for AI skills vis-a-vis Software as well as Cogni-
11 We
use the classification of jobs by 10 skill categories (i.e., Software, Computer, Cognitive, Social, Character, Writing,
Customer Service, Project Management, People Management, Finance and Accounting) following Deming and Kahn (2018)
and Deming and Noray (2019). These skill groups are present across a wide range of jobs and they were defined by the authors

18
tive skills because as shown in Table 1 Panel B, AI skills are demanded in conjunction with Software and

Cognitive skills. Column 1 shows results accounting for labor market and industry fixed effects which

allows us to control for the differences in pay across the narrowly defined labor markets and to account

for the differences across industries. Column 1 shows that AI skills command a substantial premium for

a vacancy demanding such skills: in the same labor market and industry, an employee with AI skills is

offered on average a 16% higher salary than a comparable employee with no such skills. We may want to

benchmark this effect vis-a-vis Software skills, which offer on average an 8% premium. Column 2 shows

the specification with time and firm fixed effects. Thus, within firm, vacancies demanding AI skills post

20% higher salaries than vacancies not demanding AI skills. Column 3 adds labor market fixed effects to

the firm fixed effects to account for both, constant firm characteristics and pay differences across 6-digit

SOC-CBSA markets. In this specification, vacancies demanding AI skills offer an 11% higher salary than

vacancies with no demand for AI skills, while Software skills display a 6% premium.

Finally, when we introduce job title fixed effects in column 4, the premium for AI is 5.1% over salaries

for the same job titles but with no AI skill requirements – note that we are also controlling for firm and

labor market fixed effects. This decrease in the coefficient due to job title fixed effects suggests that other

unobserved higher-paid skills are typically required together with AI skills in the same job title. Still,

the premium for AI skills is notably higher than the premium for other Software skills. In fact, the skills

with the most similar wage premia are Project Management and People Management suggesting that

the productivity differential from AI skills is closer in magnitude to management skills than to software

skills.12

It is noteworthy to compare the AI and Software skills premia to the more general Cognitive skills

premium. Across specifications, Cognitive skills do display a positive premium, however it is much

smaller. Accounting for market and firm fixed effects (column 3), Cognitive skills deliver a 3.4% wage

premia, less than one third of the AI premium.

To further understand whether the perceived value of AI specialists differs across occupations, we
following an extensive literature defining jobs tasks (i.e., Autor, Levy and Murnane, 2003; Heckman and Kautz, 2012; Deming,
2017). These skill groups were also found to explain the wage differentials across labor markets and firms (Deming and Noray,
2019). Hence, the use of these skills as controls in our analysis allows us to avoid an arbitrary choice of benchmarks for AI
skills and thus allows to compare AI jobs with wide but well understood categories of jobs.
12 We also examined the variation in wages for the sub-sample of professional occupations - those with 2-digit SOC codes

11-29 and 41-43, following Deming and Kahn (2018) and Deming and Noray (2019). The sample of professional occupations
comprises nearly 65% of our overall sample of vacancies. Table A4 in Appendix shows that the magnitude and significance of
the coefficients of interest do not change significantly and the results remain qualitatively the same.

19
run the analysis for the set of occupations that display the highest demand for AI skills in Figure 5 -

that is, Computer and Mathematical occupations (SOC 15), Architecture and Engineering occupations

(SOC 17), Life, Physical, and Social Science occupations (SOC 19), Management occupations (SOC 11),

Legal occupations (SOC 23), and Business and Financial Operations occupations (SOC 13). The main

specification for this analysis accounts for time, SOC-CBSA market, and firm fixed effects to ensure we

are comparing job vacancies within labor markets and that we control for constant unobserved firms

characteristics. The results are presented in Table 4.

As expected, AI is very relevant for vacancies in Computer and Mathematical occupations (column

1), offering a 8.8% higher salary than similar jobs with no AI skill requirement. Note that within this

occupation, Software has a larger coefficient than AI, delivering a stronger premium than AI skills.

Interestingly, Management occupations (column 4) show the highest AI skill premium: job postings for

managers that demand AI skills offer a 10.2% higher salary than job postings for managers that do not

require such skills. However, in these postings, the premium for Software skills is much smaller (only

3.5%) and is less relevant than Cognitive and Social skills. This especially high premium for AI skills

in Management jobs suggests that the ability to understand the impact of AI on broader aspects of the

business has a lot of value for employers.

Jobs in Architecture and Engineering occupations offer a 4.7% premium; Life, Physical, and Social

Science occupations a 7.5%, and Business and Financial Operations a 7.3% higher salary than similar

jobs not requiring AI skills. Finally, the AI coefficient is insignificant for vacancies in Legal occupations.

Next, we explore the premia across 2-digit NAICS industries with the highest AI Shares (as in Fig-

ure 4). Table 5 presents the results for all occupations in the top six industries demanding AI skills.

The largest premium for AI skills is displayed in the Administrative and Support services sector which

encompasses credit bureau, document management, call centers, and other business support type of

organizations. In this sector, a vacancy demanding AI skills offers a 17.7% higher salary than a vacancy

with no demand for AI. This sector has experienced the most sudden increase in the demand for AI

skills. The Information industry and the Finance and Insurance industry offer comparable 10.8% and

10.4% premia for vacancies with AI skill requirements respectively. Overall, across the main industries

that demand AI skills, a substantial premia is offered, ranging from 8% in Manufacturing and Profes-

sional Services to 17.7% in Support Services.

20
 The results in previous section showed that operational and financial firm characteristics are corre-

lated with the firms’ demand for AI skills. Therefore, we additionally explore whether these character-

istics could explain the wage premium for AI skills that we find. We repeat the analysis in Table 3 for

the sample of vacancies that can be matched to Compustat firms and include firm-level characteristics

as controls in the regressions. Table A5 in Appendix reports the results for this sample. Notably, the

sample of vacancies that we can use in this analysis is significantly smaller: we use 10-12 million va-

cancies with available data on wages in regressions in Table 3, but we have only 0.4 million vacancies

with available wage data in the Compustat-matched sample. The results show a wage premium that is

significant and positive, though slightly smaller in magnitude than in Table 3. Further analysis reveals

that this magnitude difference is mainly driven by the sample restriction, not by the additional control

variables.13

One potential concern with our approach to estimating the AI wage premium is that it could be

overstated when highly correlated and well-paid skills such as Software, and Cognitive skills, are omit-

ted from the vacancy description. To further explore this concern, we conduct some additional tests.

First, we rerun the analysis using only postings where AI and Software skills are both present and find

similar results to our baseline. Second, and more interestingly, we explore the interactions across skills

to understand if the complementarities between two skills, such as AI and Software, for example, are

priced in wages. Table A6 shows the results of our baseline wage regression specification to which we

add interactions of AI with Software, Cognitive, and Computer skills one-by-one and then all together.

In none of the regressions, the interactions are significant. However, we observe that in the regression

with AI variable interacted with Software, the individual coefficient of AI becomes smaller. Thus, when

Software skills are not mentioned in the vacancy together with AI, the size of the AI premium is slightly

lower, by 2 percentage points. Thus, a higher AI premium is generated mostly by jobs that mention both

AI and Software skills. In the specifications where AI is interacted with Cognitive and Computer, AI

coefficient increases slightly, suggesting that when Cognitive or general Computer skills are not men-

tioned in the vacancy, the premium for AI skills is, in fact, slightly higher.

Overall, our results on the AI wage premium show that across a myriad of both occupations and
13 Allspecifications in Tables 4 and 5 already control for firm and labour market fixed effects. Thus, any difference between
full sample results and the results in the Compustat-matched sample would be mainly explained by the sample restriction and
not by the inclusion of firm characteristics as controls. Re-estimation of Tables 4 and 5 in smaller Compustat-matched sample
does not alter significantly the inference made using the full data and thus the results are not tabulated.

21
industries AI skills are highly valued by employers. Since the differences in wage premium are likely to

reflect the variation in expected benefits from the skill projected by firms, these results can give a hint on

where AI is expected to bring the highest value. A substantial wage premium in IT occupations confirms

our expectations that AI is a valuable IT-related skill, which is also consistent with a fast increasing

demand for computer and math specialists with AI skills, as shown in Figure 5. However, an even

higher premium for AI skills in management jobs may suggest that the AI’s value can rather lie in its

combination with a broader organizational change.

5 Firm-level AI Share and Non-AI Wages

In this section we explore whether differences in the firm-level demand for AI skills is associated with

salary variation for the rest of posted vacancies that do not require AI skills. On the one hand, we have

seen that firms that demand AI are larger in size, tend to be more R&D intensive and are cash rich.

Hence, demanding AI in job vacancies can signal a certain quality of the firm that may be reflected

in higher salaries for jobs not requiring AI skills. As well, demanding AI can enable the creation of

new tasks that increase the demand for other high-skilled jobs that may complement this technology

(Acemoglu and Restrepo, 2018a). On the other hand, if AI algorithms are replacing workers within a

firm, we should see lower wages in non-AI vacancies in firms with a higher AI Share.

We estimate a regression model similar to equation 2 substituting AI dummy variable for the firm-

year level of AI Share. Our sample is now comprised of vacancies that do not request AI skills. The

results of this analysis are presented in Table 6. Columns 1-3 show that the increase in the firm-level

AI Share has a strong positive association with the wage level offered to non-AI vacancies. Column 1

shows a strong positive correlation when accounting for labor market, industry, and year-month fixed

effects. Columns 2 and 3 show that the positive correlation is present even after including firm fixed

effects. In column 4, job title fixed effects are added to the specification and the statistical significance

of AI Share disappears. This may suggest that the positive association with non-AI wages comes from

firms posting high-paid jobs with new titles rather than from offering higher wages "within" job titles.

So far, the evidence suggests that firms demanding more AI, that is, firms with higher AI Share,

offer higher wages to non-AI positions even after controlling for constant firms’ characteristics. This

is consistent with the view that AI technology can facilitate the creation of new tasks that increase the

22
demand for high-skilled jobs complementing it. We explore this potential explanation by running the

analysis for the sub-samples of non-AI positions across occupations and industries. As shown in Table

A7 of Appendix, the increase in the firm-level AI Share is associated with the increase in wages offered

in non-AI vacancies in specific occupations such as Computer and Mathematical, Management, and

Business and Financial occupations. Table A8 of Appendix also shows that the increase in the firm-level

AI Share is positively correlated with the level of wages offered in non-AI vacancies in specific industries

such as Professional, Scientific, and Technical Services and Administrative and Support Services. These

results show that the complementarity with AI skills is not evenly distributed across occupations and

sectors. In sum, so far we mostly find a positive or no effect of the demand for AI on wages offered in

non-AI jobs.

To check robustness of the result to including the firm-level characteristics as controls, we repeat the

analysis for the sample of vacancies matched to Compustat firms. Results in column 1 of Table A9 in

Appendix show that in the cross-section of Compustat firms, higher AI Share is associated with higher

wages in non-AI jobs even controlling for varying firm-level characteristics. The coefficient’s magnitude

is close to the one in the full sample. Specifications controlling for firm fixed effects in columns 2-4 show

no significant correlation between AI Share and non-AI wages. Thus, within-firm positive correlation

between AI Share and non-AI wages found in Table 6 is not observed in the smaller sample of Compustat

firms.

To further explore how non-AI jobs change as firms demand AI-skilled labor, we examine which

skill groups are more demanded when firms have a higher share of AI vacancies. As Acemoglu and

Restrepo (2018a) suggest, AI technology not only threatens to replace jobs, but also seems to create roles

that are complementary to AI. "Trainers", "explainers", and "sustainers" of AI are new tasks created in

the result of AI technology development. Acemoglu et al. (2020) find no evidence that AI destroys jobs

but show that it is changing the firms’ occupational structure. We, in turn, study how the structure of

skill demand is changing with the firms’ growing demand for AI skills.

Panel A of Table 7 shows that firms with a higher AI Share are more likely to require Software and

Cognitive skills in non-AI vacancies. They are also less likely to require Customer Service skills. Panel

B accounts for firm fixed effects and shows a similar pattern: firm-level AI Share is positively associated

with the probability to demand Software and Cognitive skills in non-AI job postings. Social skills and

23
People and Project Management skills as well display a complementarity with AI skills “within” firms.

And we see some evidence that when firms have a higher AI Share they are less likely to demand

Customer Service skills, even though the coefficient is significant only at the 10% level. Thus, the firms’

skill structure seems to shift towards more software-intensive, cognitive, social, and people management

jobs, and towards less customer service-intensive jobs. This evidence suggests that Software, Cognitive,

Social, and People Management jobs are complementary to AI jobs in firms.

AI Share’s positive correlation with the probability to demand Software and Cognitive skills and neg-

ative correlation with the probability to demand Customer Service skills is robust to using the smaller

sample of Compustat-matched vacancies and to controlling for varying firm-level characteristics (not

tabulated). The result is consistent in specifications with and without firm fixed effects.

One might be concerned that the association between AI Share and the likelihood to demand specific

skills in vacancies is driven by the overall increase in the demand for all skills due to a general upskilling

(or downskilling) or due to changes in a preferred job description length that also correlate with the

demand for AI. Current specifications partially address this concern by controlling for time fixed effects

that account for the overall economy-wide trends in the demand for skills. As well, by including labour

market or firm fixed effect we control for market or firm-specific probability to demand specific skills.

However, there still can be across-job variation in the length of vacancy description and we might want

to control for this. Therefore, we repeat the analysis in Table 7 controlling for the number of skills

required by the vacancy and present the results in Table A10 in Appendix.

We find that most AI Share’s correlations with other skills are robust to controlling for the length

of the skill requirements in the vacancy. The cross-sectional results in Panel A show that AI Share’s

positive correlation with the likelihood to demand Software, Cognitive, Social, Project Management

and Customer Service skills is robust to controlling for the number of skills demanded in the vacancy.

AI Share’s positive correlation with Software and People Management skills and negative correlation

with Customer Service skills is also preserved in specifications with firm fixed effects. The results of this

analysis, however, should be interpreted with caution. If the change in the number of skills required in

the vacancy is caused by the AI skills adoption in the firm, then by including this control we partially

eliminate the effect of AI. However, robustness of most effects to controlling for the number of skills

required in the vacancy provides additional support for the effects we find.

24
 Overall, these results seem to give a better understanding of the types of skills which go hand-in-

hand with AI skills.

6 Conclusion

Given the increasing importance of artificial intelligence for economic activity and the lack of knowledge

about its effects, we have examined the diffusion of AI through looking at the demand for AI-related

skills in the labor market. We have used data from Burning Glass Technologies (BGT) consisting nearly

of the universe of online job postings in the United States.

Our findings suggest a dramatic increase in the demand for AI skills. The number of positions re-

quiring AI skills has increased ten-fold from 2010 to 2019 and four-fold as a share of all job postings.

While a large proportion of the demand for AI skills is focused in Information Technology, Professional

Services, Finance, and Manufacturing sectors, the increase in demand is evident in most 2-digit NAICS

industries. Likewise, demand for AI skills is concentrated in Computer, Engineering, and Science occu-

pations, but the increase is evident in a large number of 2-digit SOC occupations, including for instance

Farming, Forestry, and Fishing which at the beginning of the observation period had almost zero de-

mand for AI skills.

We further find that there are clear differences between firms that are looking for AI skills with

different intensity. In a cross-sectional analysis, firms with higher market values, higher R&D intensity,

and higher cash stock have a higher share of AI vacancies.

Our results also indicate the existence of a significant wage premium for AI skills. Controlling for

occupation-CBSA labor market and industry fixed effects, in job postings with comparable other skills,

the addition of AI skill requirement increases the offered wage by 16%. Controlling for firm fixed effects,

the AI wage premium is almost 20%. Further controlling for occupation-CBSA and job title fixed effects

in addition to firm fixed effects lowers the premium to 11% and 5% respectively. However, this premium

is still substantially higher than for other skills in our analysis, being for instance more than twice as

high as for other software skills. Within occupations, we find that the highest premium for AI skills

is offered in Management occupations, suggesting that the ability to understand the impact of AI on

broader aspects of the business carries a lot of value.

We also find a significant wage premium in non-AI positions in companies that are looking for AI

25
skills more intensively. Positive association between the firm-level AI Share and non-AI wages could be

a signal of the quality of the firm and is also consistent with AI facilitating demand for well-paid com-

plementary positions. Our findings suggest that jobs that require Software, Cognitive, Social, Project

Management, and People Management skills are complementary to AI jobs, while jobs requiring Cus-

tomer Service skills might be substituted by AI.

Overall, our findings clearly indicate a dramatic increase in the recruitment of employees with AI

skills. Such recruitment appears to be more intensive in larger firms, firms that invest in R&D, and

firms that have more cash holdings. AI skills carry a significant wage premium across industries and

occupations and firms looking for AI specialists more intensively pay higher wages for other positions

as well.

26
References

Acemoglu, Daron, and David Autor. 2011. “Skills, tasks and technologies: Implications for employment

and earnings.” In Handbook of labor economics. Vol. 4, 1043–1171. Elsevier.

Acemoglu, Daron, and Pascual Restrepo. 2018a. “Artificial intelligence, automation and work.” Na-

tional Bureau of Economic Research.

Acemoglu, Daron, and Pascual Restrepo. 2018b. “The race between man and machine: Implications of

technology for growth, factor shares, and employment.” American Economic Review, 108(6): 1488–1542.

Acemoglu, Daron, and Pascual Restrepo. 2019. “The Wrong Kind of AI? Artificial Intelligence and the

Future of Labor Demand.” National Bureau of Economic Research.

Acemoglu, Daron, David Autor, Joe Hazell, and Pascual Restrepo. 2020. “AI and Jobs: Evidence from

Online Vacancies.” Mimeo. Massachusetts Institute of Technology.

Alderucci, Dean, Lee Branstetter, Eduard Hovy, Andrew Runge, and Nikolas Zolas. 2020. “Quanti-

fying the Impact of AI on Productivity and Labor Demand: Evidence from U.S. Census Microdata.”

mimeo.

Autor, David H, Frank Levy, and Richard J Murnane. 2003. “The skill content of recent technological

change: An empirical exploration.” The Quarterly journal of economics, 118(4): 1279–1333.

Babina, Tania, Anastassia Fedyk, Alex Xi He, and James Hodson. 2020. “Artificial Intelligence, Firm

Growth, and Industry Concentration.” Firm Growth, and Industry Concentration (July 14, 2020).

Benzell, Seth G, Laurence J Kotlikoff, Guillermo LaGarda, and Jeffrey D Sachs. 2015. “Robots are us:

Some economics of human replacement.” National Bureau of Economic Research.

Blair, Peter Q, and David J Deming. 2020. “Structural Increases in Demand for Skill after the Great

Recession.” Vol. 110, 362–65.

Bresnahan, Timothy F, and Manuel Trajtenberg. 1995. “General purpose technologies ‘Engines of

growth’?” Journal of Econometrics, 65(1): 83–108.

27
Brynjolfsson, Erik, and Andrew McAfee. 2014. The second machine age: Work, progress, and prosperity in a

time of brilliant technologies. WW Norton & Company.

Brynjolfsson, Erik, Tom Mitchell, and Daniel Rock. 2018. “What can machines learn, and what does it

mean for occupations and the economy?” Vol. 108, 43–47.

Burning Glass Technologies. 2019. Mapping the Genome of Jobs: The Burning Glass Skills Taxonomy.

Carnevale, Anthony P, Tamara Jayasundera, and Dmitri Repnikov. 2014. “Understanding online job

ads data.” Georgetown University, Center on Education and the Workforce, Technical Report (April).

Deming, David, and Kadeem Noray. 2019. “STEM Careers and the Changing Skill Requirements of

Work.”

Deming, David, and Lisa B Kahn. 2018. “Skill requirements across firms and labor markets: Evidence

from job postings for professionals.” Journal of Labor Economics, 36(S1): S337–S369.

Deming, David J. 2017. “The growing importance of social skills in the labor market.” The Quarterly

Journal of Economics, 132(4): 1593–1640.

Ersahin, Nuri. 2017. “Creditor rights, technology adoption, and productivity: Plant-level evidence.” The

Review of Financial Studies.

Felten, Edward W., Manav Raj, and Robert Seamans. 2019. “The Occupational Impact of Artificial

Intelligence: Labor, Skills, and Polarization.”

Frank, Morgan R, David Autor, James E Bessen, Erik Brynjolfsson, Manuel Cebrian, David J Deming,

Maryann Feldman, Matthew Groh, José Lobo, Esteban Moro, et al. 2019. “Toward understanding the

impact of artificial intelligence on labor.” Proceedings of the National Academy of Sciences, 116(14): 6531–

6539.

Gofman, Michael, and Zhao Jin. 2020. “Artificial Intelligence, Human Capital, and Innovation.”

Goldfarb, Avi, Bledi Taska, and Florenta Teodoridis. 2019. “Could Machine Learning Be a General-

Purpose Technology? Evidence from Online Job Postings.”

28
Goldfarb, Avi, Bledi Taska, and Florenta Teodoridis. 2020. “Artificial Intelligence in Health Care? Evi-

dence from Online Job Postings.” Vol. 110, 400–404.

Gorodnichenko, Yuriy, and Monika Schnitzer. 2013. “Financial constraints and innovation: Why poor

countries don’t catch up.” Journal of the European Economic association, 11(5): 1115–1152.

Heckman, James J, and Tim Kautz. 2012. “Hard evidence on soft skills.” Labour economics, 19(4): 451–

464.

Hershbein, Brad, and Lisa B Kahn. 2018. “Do recessions accelerate routine-biased technological

change? Evidence from vacancy postings.” American Economic Review, 108(7): 1737–72.

Mateos-Garcia, Juan. 2019. The economics of Artificial Intelligence today. Nesta.

Modestino, Alicia Sasser, Daniel Shoag, and Joshua Ballance. 2016. “Downskilling: changes in em-

ployer skill requirements over the business cycle.” Labour Economics, 41: 333–347.

Modestino, Alicia Sasser, Daniel Shoag, and Joshua Ballance. 2019. “Upskilling: Do Employers De-

mand Greater Skill When Workers are Plentiful?” The Review of Economics and Statistics, 0(ja): 1–46.

Topkis, Donald M. 1978. “Minimizing a submodular function on a lattice.” Operations research,

26(2): 305–321.

Webb, Michael. 2019. “The impact of artificial intelligence on the labor market.” Available at SSRN

3482150.

29
 300 0.8%

0.7%
250
219.7 0.6%
200
180.9 0.5%

150 0.4%

103.5 0.3%
100
73.3
65.4 0.2%
50 37.6 42.9
35.0
20.6 26.5 0.1%

0 0.0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019*

AI share of the Total number of vacancies

Number of vacancies requiring AI skills (in thousands)

Figure 1. Demand for AI skills over time. The figure shows the number of vacancies requiring AI skills over time (bars) and
the ratio of the number of vacancies requiring AI skills to the total number of vacancies (line). The total number of vacancies
requiring AI in 2019 is a projection based on the annualized number of vacancies posted in January-July, 2019. Sample is a full
BGT data, excluding internships.

30
 9,000 26%
24%
8,000 7,690
22%
7,000 6,882
20%
18%
6,000 5,475
5,027 5,236 16%
5,000 4,470 14%
4,312
4,000 12%
3,438 3,594
10%
3,000 2,568 8%
2,000 6%
4%
1,000
2%
0 0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019*

Computer skills share of the Total number of vacancies

Number of vacancies requiring Computer skills (in thousands)

(a) The number and share of vacancies demanding Computer skills over time.

9,000 26%
24%
8,000
22%
7,000 6,725 20%
6,199
18%
6,000
16%
5,000 4,507 4,585 14%
4,411
4,000 3,653 3,646 3,699 12%
3,550
10%
3,000 2,797
8%
2,000 6%
4%
1,000
2%
0 0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019*

Software skills share of the Total number of vacancies

Number of vacancies requiring Software skills (in thousands)

(b) The number and share of vacancies demanding Software skills over time.

Figure 2. Demand for Computer and Software skills over time. The figure shows the number of vacancies requiring the
analysed skills over time (bars) and the ratio of the number of vacancies requiring these skills to the total number of vacancies
(line). The total number of vacancies requiring Computer/Software skills in 2019 is a projection based on the annualized
number of vacancies posted in January-July, 2019. Sample is a full BGT data, excluding internships.

31
 0 2,000 4,000 6,000 8,000 10,000 12,000

Machine Learning

Artificial Intelligence

Natural Language Processing

Deep Learning

Image Processing

Computer Vision

Neural Networks

Speech Recognition

Decision Trees

Keras

Random Forests

IBM Watson

Text Mining

Pattern Recognition

Tokenization

Figure 3. Most frequently required AI skills. The figure shows the top 15 AI skills sorted by the number of vacancies in
which they are demanded in the BGT dataset in July, 2019.

32
 0.0% 0.5% 1.0% 1.5% 2.0% 2.5%

Information

Professional, Scientific, and Technical Services

Finance and Insurance

Administrative and Support Services

Agriculture, Forestry, Fishing and Hunting

Manufacturing

Management of Companies and Enterprises

Mining, Quarrying, and Oil and Gas Extraction

Educational Services

Public Administration

Wholesale Trade

Retail Trade

Utilities

Real Estate and Rental and Leasing

Health Care and Social Assistance

Transportation and Warehousing

Arts, Entertainment, and Recreation

Accommodation and Food Services

Other Services (except Public Administration)

Construction

2019 2018 2017 2016 2015 2014 2013 2012 2011 2010

Figure 4. AI Share by industry over time. The figure shows the share of AI vacancies in the total number of vacancies
by 2-digit NAICS industries in 2010-2019. Industries are ranked by AI Share in 2019. Sample is a full BGT data, excluding
internships. Data only includes job postings with non-missing 2-digit NAICS codes.

33
 0% 1% 2% 3% 4% 5%

Computer and Mathematical

Architecture and Engineering

Life, Physical, and Social Science

Management

Legal

Business and Financial Operations

Arts, Design, Entertainment, Sports, and Media

Protective Service

Education, Training, and Library

Farming, Fishing, and Forestry

Office and Administrative Support

Sales and Related

Healthcare Practitioners and Technical

Community and Social Service

Military Specific

Personal Care and Service

Installation, Maintenance, and Repair

Healthcare Support

Production

Food Preparation and Serving

Building and Grounds Cleaning and Maintenance

Construction and Extraction

Transportation and Material Moving

2019 2018 2017 2016 2015 2014 2013 2012 2011 2010

Figure 5. AI Share by occupation over time. The figure shows the share of AI vacancies in the total number of vacancies by
2-digit occupations (first two digits of SOC code) in 2010-2019. Occupations are ranked by AI Share in 2019. Sample is a full
BGT data, excluding internships. Data only includes job postings with non-missing SOC codes.

34
 1.0%

0.9%

0.8%

0.7%

0.6%

0.5%

0.4%

0.3%

0.2%

0.1%

0.0%
2010 2011 2012 2013 2014 2015 2016 2017 2018
Bottom tercile Middle tercile Top tercile

(a) AI Share by terciles of market capitalization.

1.2%

1.1%

1.0%

0.9%

0.8%

0.7%

0.6%

0.5%

0.4%

0.3%

0.2%

0.1%

0.0%
2010 2011 2012 2013 2014 2015 2016 2017 2018
Zero or missing R&D Positive R&D

(b) AI Share by R&D expenses.

1.1%

1.0%

0.9%

0.8%

0.7%

0.6%

0.5%

0.4%

0.3%

0.2%

0.1%

0.0%
2010 2011 2012 2013 2014 2015 2016 2017 2018
Bottom tercile Middle tercile Top tercile

(c) AI Share by terciles of cash-to-assets ratio.

Figure 6. AI Share heterogeneity across firms over time. Data source: BGT merged with Compustat.

35
Table 1. Summary statistics.

(a) Panel A: Demand for Skills in Vacancies.

The sample is a full BGT data in 2010-2019 period, excluding internships. Each observation is one vacancy.

Variable Obs. Mean Std. Min Max

AI 190,218,763 0.0038 0.0611 0 1
Software 190,218,763 0.2154 0.4111 0 1
Computer 190,218,763 0.2391 0.4266 0 1
Cognitive 190,218,763 0.2826 0.4502 0 1
Social 190,218,763 0.3979 0.4895 0 1
Character 190,218,763 0.2039 0.4029 0 1
Writing 190,218,763 0.0510 0.2199 0 1
Customer Service 190,218,763 0.3889 0.4875 0 1
Project Management 190,218,763 0.0757 0.2646 0 1
People Management 190,218,763 0.1371 0.3439 0 1
Finance & Accounting 190,218,763 0.1524 0.3594 0 1
Salary 33,822,194 57,307 42,722 10,005 350,000

(b) Panel B: Ratio of Conditional to Unconditional Probabilities of Skill Demand.

Each cell in the table presents the ratio of a probability to demand the skill (in rows) conditionally on another skill (in columns) to the
unconditional probability to demand the skill (in rows) in the overall sample (i.e., the sample mean in Panel A). The sample is a full BGT data
in 2010-2019 period, excluding internships. Each observation is one vacancy.
Conditioning on Skill
Customer Project People Finance &
AI Software Computer Cognitive Social Character Writing
Service Management Management Accounting
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)

AI 4.00 1.18 2.46 1.49 0.97 1.66 0.69 1.98 1.29 0.92
Software 4.00 1.72 1.77 1.42 1.35 2.00 0.84 2.68 1.29 1.69
Computer 1.18 1.72 1.68 1.66 1.97 1.99 1.22 1.81 1.45 1.86
Cognitive 2.46 1.77 1.68 1.65 1.61 1.94 1.21 2.13 1.81 1.90
Social 1.49 1.42 1.66 1.65 1.79 1.62 1.35 1.74 1.61 1.54
Character 0.97 1.35 1.97 1.61 1.79 1.77 1.36 1.55 1.36 1.52
Writing 1.66 2.00 1.99 1.94 1.62 1.77 0.99 2.66 1.64 2.00
Customer Service 0.69 0.84 1.22 1.21 1.35 1.36 0.99 0.93 1.28 1.17
Project Management 1.98 2.68 1.81 2.13 1.74 1.55 2.66 0.93 2.28 2.68
People Management 1.29 1.29 1.45 1.81 1.61 1.36 1.64 1.28 2.28 2.23
Finance & Accounting 0.92 1.69 1.86 1.90 1.54 1.52 2.00 1.17 2.68 2.23

36
(c) Panel C: Firm characteristics - Compustat-matched sample.
The sample includes firms that have Compustat and BGT data available. BGT vacancies data is for period 2010-2018, Com-
pustat data is for 2009-2017, as used in the AI adoption and firm characteristics regression analysis (characteristics are lagged
by 1 year). We only keep observations with non-missing data for all Compustat-based variables included in the analysis. All
Compustat-based ratios are winsorized at 5% and 95% and monetary variables are deflated by the CPI and log-transformed.

Variable Obs. Mean Std. Min Max

AI Share 20,760 0.58% 3.66% 0.00% 100.00%
AI Dummy 20,760 23.82% 42.60% 0.00% 100.00%

Log(Market Cap) 20,760 6.932 2.175 -4.778 13.522

Log(Sales) 20,760 6.554 2.367 -6.213 13.076
Log(Employment) 20,760 0.830 2.111 -6.908 7.741
Log(Vacancies) 16,599 4.577 2.388 0 12.024

ROA 20,760 -0.010 0.150 -0.576 0.171

Cash / Assets 20,760 0.196 0.200 0.004 0.754
Book Leverage 20,760 0.227 0.200 0 0.731
R&D / Sales 20,760 0.094 0.234 0 1.554
PP&E / Assets 20,760 0.481 0.385 0.025 1.385
Tobin’s Q 20,760 1.991 1.235 0.758 6.338
CAPEX / Sales 20,760 0.072 0.100 0.003 0.475

Market Cap 20,760 7,967 27,392 0 790,050

Sales 20,760 5,997 21,645 0 483,521
Employment 20,760 16 61 0 2,300
Vacancies 16,599 1,165 5,128 1 166,724

37
 Table 2. Demand for AI and Firm Characteristics.
The table presents the results of the OLS regressions weighted by the firm employment. The sample includes firms that have Compustat and BGT data available and the independent
variables are lagged by one year: BGT vacancies data is for period 2010-2018, Compustat data if for period 2009-2017. Each observation is at the firm-year level. We only keep
observations with non-missing data for all Compustat-based variables included in the analysis. All Compustat-based ratios are winsorized at 5% and 95% and monetary variables are
deflated by the CPI and log-transformed. Independent variables: column (1) is the log market capitalization, column (2) the log of employment, column (3) the log sales, column (4)
the market-to-book ratio, column (5) the return on assets, column (6) the cash holdings to assets ratio, column (7) the book leverage, column (8) the R&D expenses over sales, column
(9) the capital expenditures over sales, column (10) the PP&E over assets, column (11) the log of total vacancies. Column (12) presents the multivariate specification. Standard errors in
parentheses are clustered by firm. Significance levels are denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: AI Sharet+1

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

Log(Market Cap) 0.00116*** 0.00158**

(0.000268) (0.000671)
Log(Employment) 0.000550**
(0.000240)
Log(Sales) 0.000773***
(0.000268)
Tobin’s Q 0.00155***
(0.000567)
ROA 0.00867 -0.00842
(0.0110) (0.0175)
Cash/Assets, 0.0261*** 0.0142**

38
(0.00791) (0.00694)
Book Leverage -0.00453 0.000422
(0.00292) (0.00379)
R&D/Sales 0.0697*** 0.0532***
(0.0131) (0.0114)
Capex/Sales -0.00743 -0.00646
(0.00876) (0.00783)
PPE/Assets -0.00622*** -0.00273
(0.00174) (0.00171)
Log(Vacancies) - 0.00005 -0.000561
(0.000512) (0.000650)

Observations 20,760 20,760 20,760 20,760 20,760 20,760 20,760 20,760 20,760 20,760 16,599 16,599
R-squared 0.045 0.040 0.041 0.042 0.040 0.049 0.040 0.057 0.039 0.044 0.043 0.072

Year FE X X X X X X X X X X X X
Industry FE X X X X X X X X X X X X
Table 3. AI Wage Premia.
The table presents the results of the regressions of log wages on AI and other skills dummies. The sample is a full BGT data in 2016-2019
period, excluding internships and vacancies with missing salary data. Each observation is one vacancy. The main independent variable, AI, is
an indicator variable equal one if the vacancy requests at least one AI-related skill and zero otherwise. Industry is defined as 3-digit NAICS
industry, SOC is a 6-digit SOC code. Standard errors in parentheses are clustered by CBSA, SOC, and 3-digit NAICS. Significance levels are
denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Log Wage

(1) (2) (3) (4)

AI 0.164*** 0.197*** 0.107*** 0.0514***

(0.0130) (0.0346) (0.0131) (0.00680)
Software 0.0812*** 0.120*** 0.0595*** 0.0302***
(0.0119) (0.0212) (0.00873) (0.00510)
Computer -0.0848*** -0.0736*** -0.0509*** -0.0300***
(0.0109) (0.0173) (0.00840) (0.00573)
Cognitive 0.0314*** 0.0722*** 0.0341*** 0.0157***
(0.00838) (0.00777) (0.00449) (0.00346)
Social 0.00893* 0.0449*** 0.0210*** 0.00708***
(0.00487) (0.00603) (0.00341) (0.00255)
Character -0.0635*** -0.0671*** -0.0360*** -0.0197***
(0.00872) (0.00811) (0.00465) (0.00333)
Writing 0.0196** 0.00320 0.00776 0.00463
(0.00973) (0.0133) (0.00489) (0.00480)
Customer Service -0.0170* -0.0182 -0.0157*** -0.0120***
(0.00908) (0.0192) (0.00522) (0.00366)
Project Management 0.132*** 0.182*** 0.109*** 0.0636***
(0.0149) (0.0190) (0.0114) (0.00716)
People Management 0.0789*** 0.161*** 0.0813*** 0.0488***
(0.00728) (0.0106) (0.00464) (0.00376)
Finance & Accounting 0.0973*** 0.138*** 0.0914*** 0.0460***
(0.00929) (0.0210) (0.00739) (0.00497)

Observations 12,247,697 9,978,996 9,923,378 6,424,944

R-squared 0.507 0.578 0.701 0.757

Year-month FE X X X X
Industry FE X
CBSA× SOC FE X X X
Firm FE X X X
Job Title FE X

39
Table 4. AI Wage Premia Across Occupations.
The table presents the results of the regressions of log wages on AI and other skills dummies by the top 2-digit SOC occupations ranked by
AI Share. The sample is a full BGT data in 2016-2019 period, excluding internships and vacancies with missing salary data. Each observation
is one vacancy. The main independent variable, AI, is an indicator variable equal one if the vacancy requests at least one AI-related skill and
zero otherwise. Columns are labeled using the following 2-digit occupation names and abbreviations: Computer & Math - Computer and
Mathematical (SOC 15), Architecture and Engineering (SOC 17), Science - Life, Physical, and Social Science (SOC 19), Management (SOC 11),
Legal (SOC 23), Business & Financial - Business and Financial Operations (SOC 13). Standard errors in parentheses are clustered by CBSA and
3-digit NAICS. Significance levels are denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Log Wage

Computer & Architecture & Business &

Science Management Legal
Math Engineering Financial
(1) (2) (3) (4) (5) (6)

AI 0.0878*** 0.0472*** 0.0746*** 0.102*** -0.0399 0.0732***

(0.00695) (0.0158) (0.0227) (0.0159) (0.0536) (0.0103)
Software 0.140*** 0.0341*** 0.0169 0.0345*** 0.00135 0.0306***
(0.0129) (0.00991) (0.0118) (0.00662) (0.0106) (0.00568)
Computer -0.0994*** -0.0267*** -0.0826*** -0.0856*** -0.0659*** -0.0696***
(0.00465) (0.00536) (0.0157) (0.0112) (0.0125) (0.00531)
Cognitive 0.0206** 0.0318*** 0.0251*** 0.0506*** 0.0394** 0.0419***
(0.00992) (0.00608) (0.00859) (0.00661) (0.0169) (0.00607)
Social 0.0323*** 0.0262*** 0.0335*** 0.0414*** 0.0235*** 0.0325***
(0.00613) (0.00382) (0.00576) (0.00404) (0.00838) (0.00404)

Observations 544,130 191,761 109,140 954,759 62,341 534,600

R-squared 0.613 0.731 0.738 0.628 0.757 0.634

Other Skills controls X X X X X X

Year-month FE X X X X X X
CBSA× SOC FE X X X X X X
Firm FE X X X X X X

40
Table 5. AI Wage Premia Across Industries.
The table presents the results of the regressions of log wages on AI and other skills dummies by the top 2-digit NAICS industries ranked by
AI Share. The sample is a full BGT data in 2016-2019 period, excluding internships and vacancies with missing salary data. Each observation
is one vacancy. The main independent variable, AI, is an indicator variable equal one if the vacancy requests at least one AI-related skill
and zero otherwise. Columns are labeled using the following 2-digit NAICS industry names and abbreviations: Information (NAICS 51),
Professional Services - Professional, Scientific, and Technical Services (NAICS 54), Finance & Insurance - Finance and Insurance (NAICS
52), Administrative & Support - Administrative and Support Services (NAICS 56), Agriculture - Agriculture, Forestry, Fishing and Hunting
(NAICS 11), Manufacturing (NAICS 31-33). Standard errors in parentheses are clustered by CBSA and 6-digit SOC. Significance levels are
denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Log Wage

Professional Finance & Admin &

Information Agriculture Manufacturing
Services Insurance Support
(1) (2) (3) (4) (5) (6)

AI 0.108*** 0.0816*** 0.104*** 0.177*** -0.0143 0.0896***

(0.0280) (0.0101) (0.0101) (0.0193) (0.0290) (0.0202)
Software 0.113*** 0.0793*** 0.0570*** 0.0433*** 0.0258** 0.0662***
(0.0131) (0.00791) (0.0120) (0.0127) (0.0123) (0.00807)
Computer -0.0819*** -0.0817*** -0.0782*** -0.0224** -0.0306 -0.0326***
(0.00768) (0.00484) (0.00646) (0.0111) (0.0193) (0.00701)
Cognitive 0.0256** 0.0192*** 0.0154** 0.0181 0.0163 0.0289***
(0.00991) (0.00469) (0.00741) (0.0112) (0.0114) (0.00528)
Social 0.0139 0.0279*** -0.00139 0.0212*** 0.0414*** 0.0264***
(0.0107) (0.00295) (0.00914) (0.00274) (0.00990) (0.00444)

Observations 240,022 829,057 571,423 675,318 21,539 730,758

R-squared 0.791 0.716 0.714 0.762 0.871 0.736

Other Skills controls X X X X X X

Year-month FE X X X X X X
CBSA× SOC FE X X X X X X
Firm FE X X X X X X

41
Table 6. Firm AI Adoption and Non-AI Wages.
The table presents regressions of log wages in non-AI vacancies on the firm-level AI Share. The sample is a full BGT data in 2016-2019 period,
excluding AI vacancies, internships, and vacancies with missing salary data. Each observation is one vacancy. The main independent variable,
Firm-Level AI Share, is calculated for each firm-year. Industry is defined as 3-digit NAICS industry, SOC is a 6-digit SOC code. Standard errors
in parentheses are clustered by CBSA, SOC, and 3-digit NAICS. Significance levels are denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Non-AI Log Wage

(1) (2) (3) (4)

Firm-Level AI Share 1.235*** 0.591** 0.319** 0.169

(0.123) (0.278) (0.133) (0.110)

Software 0.0785*** 0.119*** 0.0592*** 0.0300***

(0.0115) (0.0210) (0.00869) (0.00507)
Computer -0.0840*** -0.0734*** -0.0508*** -0.0300***
(0.0108) (0.0174) (0.00844) (0.00575)
Cognitive 0.0306*** 0.0720*** 0.0341*** 0.0157***
(0.00834) (0.00780) (0.00450) (0.00347)
Social 0.00833* 0.0449*** 0.0210*** 0.00713***
(0.00481) (0.00607) (0.00343) (0.00256)
Character -0.0630*** -0.0670*** -0.0360*** -0.0198***
(0.00863) (0.00812) (0.00465) (0.00335)
Writing 0.0200** 0.00361 0.00801 0.00472
(0.00965) (0.0134) (0.00487) (0.00482)
Customer Service -0.0169* -0.0179 -0.0157*** -0.0119***
(0.00916) (0.0192) (0.00523) (0.00366)
Project Management 0.131*** 0.184*** 0.109*** 0.0640***
(0.0147) (0.0189) (0.0113) (0.00714)
People Management 0.0789*** 0.161*** 0.0813*** 0.0489***
(0.00731) (0.0105) (0.00462) (0.00376)
Finance & Accounting 0.0977*** 0.139*** 0.0917*** 0.0460***
(0.00933) (0.0211) (0.00737) (0.00497)

Observations 12,212,048 9,948,114 9,892,498 6,409,438

R-squared 0.506 0.577 0.700 0.757

Year-month FE X X X X
Industry FE X
CBSA× SOC FE X X X
Firm FE X X X
Job Title FE X

42
Table 7. AI Share and The Demand for Other Skills.
The table presents the results of the linear probability regressions of the demand for other skills on the firm-level AI Share. The sample is
a full BGT data in 2016-2019 period, excluding AI vacancies and internships. Each observation is one vacancy. The dependent variable is a
dummy equal 1 if the skill is required by the vacancy and zero otherwise. The main independent variable, Firm-Level AI Share, is calculated
for each firm-year. Standard errors in parentheses are clustered by CBSA, 6-digit SOC, and 3-digit NAICS. Significance levels are denoted as:
*** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Skill Dummy (Non-AI Vacancies)

Customer Project People Finance &
Software Computer Cognitive Social Character Writing
Service Management Management Accounting
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Panel A: Industry & Year-Month fixed effects

Firm-Level AI Share 3.737*** -0.223 2.520*** 1.760*** -0.327 0.237* -1.105*** 1.300*** 0.852*** 0.508*
(0.378) (0.339) (0.385) (0.282) (0.434) (0.126) (0.347) (0.245) (0.178) (0.282)

Observations 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778
R-squared 0.144 0.064 0.090 0.086 0.045 0.033 0.128 0.064 0.024 0.094

Year-month FE X X X X X X X X X X
Industry FE X X X X X X X X X X

Panel B: Employer & Year-Month fixed effects

Firm-Level AI Share 0.646*** -0.0359 0.362** 0.255** 0.0509 0.0206 -0.290* 0.217* 0.235*** 0.107
(0.155) (0.0865) (0.158) (0.113) (0.104) (0.0321) (0.162) (0.112) (0.0798) (0.0982)

Observations 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320
R-squared 0.316 0.249 0.278 0.313 0.259 0.145 0.313 0.161 0.138 0.208

Year-month FE X X X X X X X X X X
Firm FE X X X X X X X X X X

43
A Appendix

200 0.9%

0.8%
155.7
0.7%
150
124.6 0.6%

0.5%
100
78.3 0.4%

57.4 0.3%
51.5
50
0.2%
29.3 28.5 32.1
22.0
16.6 0.1%

0 0.0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019*

AI share of the Total number of vacancies

Number of vacancies requiring AI skills (in thousands)

Figure A1. Demand for AI skills over time (balanced sample). The figure shows the number of vacancies requiring AI skills
over time (bars) and the ratio of the number of vacancies requiring AI skills to the total number of vacancies (line). The data
includes only firms that were present in BGT data each year of 2010-2019. The total number of vacancies requiring AI in 2019
is a projection based on the annualized number of vacancies posted in January-July, 2019.

44
Table A1. List of skills in the Burning Glass Technologies job vacancies dataset used to identify AI vacancies.

N Skill N Skill
1 AI ChatBot 31 Machine Translation (MT)
2 AI KIBIT 32 Machine Vision
3 ANTLR 33 Madlib
4 Apertium 34 Mahout
5 Artificial Intelligence 35 Microsoft Cognitive Toolkit
6 Automatic Speech Recognition (ASR) 36 MLPACK (C++ library)
7 Caffe Deep Learning Framework 37 Mlpy
8 Chatbot 38 Modular Audio Recognition Framework

(MARF)
9 Computational Linguistics 39 MoSes
10 Computer Vision 40 MXNet
11 Decision Trees 41 Natural Language Processing
12 Deep Learning 42 Natural Language Toolkit (NLTK)
13 Deeplearning4j 43 ND4J (software)
14 Distinguo 44 Nearest Neighbor Algorithm
15 Google Cloud Machine Learning Platform 45 Neural Networks
16 Gradient boosting 46 Object Recognition
17 H2O (software) 47 Object Tracking
18 IBM Watson 48 OpenCV
19 Image Processing 49 OpenNLP
20 Image Recognition 50 Pattern Recognition
21 IPSoft Amelia 51 Pybrain
22 Ithink 52 Random Forests
23 Keras 53 Recommender Systems
24 Latent Dirichlet Allocation 54 Semantic Driven Subtractive Clustering

Method (SDSCM)
25 Latent Semantic Analysis 55 Semi-Supervised Learning
26 Lexalytics 56 Sentiment Analysis / Opinion Mining
27 Lexical Acquisition 57 Sentiment Classification
28 Lexical Semantics 58 Speech Recognition
29 Libsvm 59 Supervised Learning (Machine Learning)
30 Machine Learning 60 Support Vector Machines (SVM)
Continued on next page

45
 Table A1 – Continued from previous page
N Skill N Skill
61 TensorFlow 67 Virtual Agents
62 Text Mining 68 Vowpal
63 Text to Speech (TTS) 69 Wabbit
64 Tokenization 70 Word2Vec
65 Torch (Machine Learning) 71 Xgboost
66 Unsupervised Learning

46
Table A2. Skills Required Together In Vacancies.
The table presents the results of the logit regressions of the demand for a specific skill on the demand for each of the remaining skills. The
sample is a full BGT data in 2016-2019 period, excluding internships. Each observation is one vacancy. Each coefficient represents the coefficient
of the one-to-one logit regression where the dependent variable is in a column and the independent variable is in a row: dependent variable is
a dummy equal 1 if the skill (in columns) is required by the vacancy and zero otherwise; independent variable is a dummy which equals one
if the skill (in rows) is required and zero otherwise.

Dependent Variable: Skill Dummy

Customer Project People Finance &
AI Software Computer Cognitive Social Character Writing
Service Management Management Accounting
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)

AI 24.94*** 1.176*** 5.876*** 2.277*** 0.881*** 1.790*** 0.560*** 2.292*** 1.446*** 0.945***
(0.105) (0.00381) (0.0184) (0.00671) (0.00318) (0.00923) (0.00181) (0.00922) (0.00536) (0.00390)
Software 24.94*** 3.084*** 3.606*** 2.506*** 1.729*** 3.206*** 0.767*** 6.051*** 1.552*** 2.676***
(0.105) (0.00170) (0.00193) (0.00131) (0.00100) (0.00314) (0.000411) (0.00502) (0.00107) (0.00169)
Computer 1.176*** 3.084*** 3.260*** 4.363*** 4.141*** 3.149*** 1.621*** 2.779*** 1.932*** 3.404***
(0.00381) (0.00170) (0.00167) (0.00227) (0.00225) (0.00305) (0.000798) (0.00228) (0.00125) (0.00208)
Cognitive 5.876*** 3.606*** 3.260*** 4.558*** 2.720*** 3.351*** 1.670*** 4.375*** 3.231*** 3.820***
(0.0184) (0.00193) (0.00167) (0.00225) (0.00143) (0.00324) (0.000780) (0.00362) (0.00201) (0.00231)
Social 2.277*** 2.506*** 4.363*** 4.558*** 5.313*** 2.958*** 2.574*** 3.864*** 3.203*** 2.982***
(0.00671) (0.00131) (0.00227) (0.00225) (0.00294) (0.00298) (0.00113) (0.00339) (0.00204) (0.00181)
Character 0.881*** 1.729*** 4.141*** 2.720*** 5.313*** 2.367*** 2.034*** 1.988*** 1.584*** 2.094***
(0.00318) (0.00100) (0.00225) (0.00143) (0.00294) (0.00235) (0.00104) (0.00170) (0.00108) (0.00133)
Writing 1.790*** 3.206*** 3.149*** 3.351*** 2.958*** 2.367*** 0.987*** 3.596*** 1.899*** 2.678***
(0.00923) (0.00314) (0.00305) (0.00324) (0.00298) (0.00235) (0.000966) (0.00445) (0.00221) (0.00284)
Customer Service 0.560*** 0.767*** 1.621*** 1.670*** 2.574*** 2.034*** 0.987*** 0.863*** 1.707*** 1.403***
(0.00181) (0.000411) (0.000798) (0.000780) (0.00113) (0.00104) (0.000966) (0.000718) (0.00104) (0.000827)
Project Management 2.292*** 6.051*** 2.779*** 4.375*** 3.864*** 1.988*** 3.596*** 0.863*** 3.252*** 4.895***
(0.00922) (0.00502) (0.00228) (0.00362) (0.00339) (0.00170) (0.00445) (0.000718) (0.00292) (0.00415)
People Management 1.446*** 1.552*** 1.932*** 3.231*** 3.203*** 1.584*** 1.899*** 1.707*** 3.252*** 3.604***
(0.00536) (0.00107) (0.00125) (0.00201) (0.00204) (0.00108) (0.00221) (0.00104) (0.00292) (0.00250)
Finance & Accounting 0.945*** 2.676*** 3.404*** 3.820*** 2.982*** 2.094*** 2.678*** 1.403*** 4.895*** 3.604***
(0.00390) (0.00169) (0.00208) (0.00231) (0.00181) (0.00133) (0.00284) (0.000827) (0.00415) (0.00250)

47
 Table A3. Demand For AI Skills And Firm Characteristics.
The table presents the linear probability regression results of the demand for AI skills on firm characteristics. The sample includes firms that have Compustat and BGT data available in
2010-2018 period. Each observation is at the Firm-Year-Market (6-digit SOC-CBSA) level. Independent variables are lagged by one year. We only keep observations with non-missing
data for all Compustat-based variables included in the analysis. All Compustat-based ratios are winsorized at 5% and 95% and monetary variables are deflated by the CPI and log-
transformed. Independent variables: column (1) is the log market capitalization, column (2) the log of employment, column (3) the log sales, column (4) the market-to-book ratio, column
(5) the return on assets, column (6) the cash holdings to assets ratio, column (7) the book leverage, column (8) the R&D expenses over sales, column (9) the capital expenditures over
sales, column (10) the PP&E over assets, column (11) the log of total vacancies. Column (12) presents the multivariate specification. Standard errors in parentheses are clustered by firm,
CBSA, and 6-digit SOC. Significance levels are denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: AI Dummyt+1

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

Log(Market Cap) 0.00228*** 0.00301***

(0.000598) (0.000822)
Log(Employment) 0.00202***
(0.000543)
Log(Sales) 0.00216***
(0.000572)
Tobin’s Q 0.00102**
(0.000500)
ROA 0.0227** -0.00300
(0.00887) (0.0104)
Cash/Assets 0.0118** 0.0175**

48
(0.00528) (0.00839)
Book Leverage 0.00005 0.00556
(0.00212) (0.00359)
R&D/Sales 0.0691*** 0.0884***
(0.0209) (0.0333)
Capex/Sales 0.00989 0.0138
(0.0114) (0.0177)
PPE/Assets -0.00187 0.000521
(0.00121) (0.00218)
Log(Vacancies) 0.0119*** 0.0113***
(0.00356) (0.00340)

Observations 4,561,055 4,561,055 4,561,055 4,561,055 4,561,055 4,561,055 4,561,055 4,561,055 4,561,055 4,561,055 2,026,055 2,026,055
R-squared 0.123 0.122 0.123 0.121 0.121 0.121 0.121 0.122 0.121 0.121 0.151 0.154
Year FE X X X X X X X X X X X X
CBSA× SOC FE X X X X X X X X X X X X
Industry FE X X X X X X X X X X X X
Table A4. AI Wage Premia in Professional Occupations.
The table presents the results of the regressions of log wages on AI and other skills dummies. The sample consists of professional occupations
(2-digit SOC codes 11-29 and 41-43) in BGT data in 2016-2019 period, excluding internships and vacancies with missing salary data. Each
observation is one vacancy. The main independent variable, AI, is an indicator variable equal one if the vacancy requests at least one AI-
related skill and zero otherwise. Industry is defined as 3-digit NAICS industry, SOC is a 6-digit SOC code. Standard errors in parentheses are
clustered by CBSA, SOC, and 3-digit NAICS. Significance levels are denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Log Wage

(1) (2) (3) (4)

AI 0.166*** 0.180*** 0.103*** 0.0493***

(0.0131) (0.0317) (0.0122) (0.00608)
Software 0.0825*** 0.103*** 0.0577*** 0.0294***
(0.0127) (0.0205) (0.00895) (0.00532)
Computer -0.0994*** -0.118*** -0.0628*** -0.0356***
(0.0113) (0.0155) (0.00769) (0.00510)
Cognitive 0.0324*** 0.0597*** 0.0344*** 0.0155***
(0.00902) (0.00730) (0.00505) (0.00367)
Social 0.0121** 0.0389*** 0.0224*** 0.00714**
(0.00551) (0.00640) (0.00398) (0.00324)
Character -0.0726*** -0.0778*** -0.0411*** -0.0236***
(0.00883) (0.00649) (0.00447) (0.00342)
Writing 0.0169 -0.00618 0.00372 0.00423
(0.0105) (0.0130) (0.00520) (0.00470)
Customer Service -0.0148* -0.0391* -0.0166*** -0.0127***
(0.00875) (0.0201) (0.00520) (0.00391)
Project Management 0.129*** 0.167*** 0.103*** 0.0615***
(0.0158) (0.0179) (0.0113) (0.00711)
People Management 0.0755*** 0.143*** 0.0785*** 0.0463***
(0.00763) (0.0105) (0.00480) (0.00373)
Finance & Accounting 0.0940*** 0.110*** 0.0858*** 0.0465***
(0.00919) (0.0198) (0.00706) (0.00363)

Observations 8,131,465 6,486,762 6,453,258 3,993,719

R-squared 0.503 0.557 0.688 0.771

Year-month FE X X X X
Industry FE X
CBSA× SOC FE X X X
Firm FE X X X
Job Title FE X

49
Table A5. AI Wage Premia - Compustat-matched Sample.
The table presents the results of the regressions of log wages on AI, other skills dummies, and firm-level characteristics. The sample is BGT data
in 2016-2019 period, excluding internships and vacancies with missing salary data, matched to Compustat firms that have data on required
firm characteristics. Each observation is one vacancy. The main independent variable, AI, is an indicator variable equal one if the vacancy
requests at least one AI-related skill and zero otherwise. Industry is defined as 3-digit NAICS industry, SOC is a 6-digit SOC code. Standard
errors in parentheses are clustered by CBSA, SOC, and 3-digit NAICS. Significance levels are denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Log Wage

(1) (2) (3) (4)

AI 0.0880** 0.268*** 0.105*** 0.0376***

(0.0388) (0.0547) (0.0264) (0.0142)

Software 0.114*** 0.202*** 0.0925*** 0.0529***

(0.0209) (0.0335) (0.0160) (0.0123)
Computer -0.0672*** -0.0838*** -0.0415** -0.0218**
(0.0222) (0.0308) (0.0186) (0.00966)
Cognitive 0.00377 0.0499*** 0.0210** 0.00732
(0.0214) (0.0181) (0.0101) (0.00822)
Social 0.0115 0.0267 0.0230 -0.00180
(0.0209) (0.0184) (0.0193) (0.0116)

Log(Market Cap) 0.0200*** 0.0122 0.0240 0.0491***

(0.00670) (0.0333) (0.0169) (0.0173)
ROA -1.081** 0.148 -0.0815 -0.363***
(0.473) (0.267) (0.137) (0.111)
Cash/Assets 0.214 0.0178 0.303 0.271
(0.195) (0.484) (0.431) (0.395)
Leverage 0.181** 0.100 0.0674 -0.0471
(0.0903) (0.222) (0.141) (0.146)
R&D/Sales -0.295* -1.763** -1.080 -0.973
(0.170) (0.800) (0.759) (0.747)
Capex/Sales -0.266 -0.422 -0.371 -0.0772
(0.241) (0.446) (0.271) (0.247)
PPE/Assets -0.146 -0.134 -0.0206 -0.0836
(0.0891) (0.292) (0.280) (0.349)
Log(Vacancies) -0.0217** -0.0229 -0.0317*** -0.0292***
(0.0103) (0.0163) (0.0114) (0.0105)

Observations 395,885 410,846 395,563 277,959

R-squared 0.668 0.614 0.760 0.812

Other Skills controls X X X X

Year-month FE X X X X
Industry FE X
CBSA× SOC FE X X X
Firm FE X X X
Job Title FE X

50
Table A6. The Effect of Skill Complementarity to AI on Wages.
The table presents the results of the regressions of log wages on AI dummy and its interactions with dummies of the related skills. The
sample is a full BGT data in 2016-2019 period, excluding internships and vacancies with missing salary data. Each observation is one vacancy.
Standard errors are clustered by CBSA, 6-digit SOC, and 3-digit NAICS. Significance levels are denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Log Wage

(1) (2) (3) (4) (5)

AI 0.107*** 0.0866*** 0.113*** 0.122*** 0.105***

(0.0131) (0.0235) (0.0155) (0.0177) (0.0234)
AI x Software 0.0244 0.0269
(0.0267) (0.0275)
AI x Computer -0.0212 -0.0205
(0.0162) (0.0168)
AI x Cognitive -0.0210 -0.0208
(0.0145) (0.0146)
Software 0.0595*** 0.0594*** 0.0595*** 0.0595*** 0.0593***
(0.00873) (0.00880) (0.00873) (0.00873) (0.00880)
Computer -0.0509*** -0.0509*** -0.0508*** -0.0509*** -0.0508***
(0.00840) (0.00840) (0.00842) (0.00841) (0.00843)
Cognitive 0.0341*** 0.0341*** 0.0341*** 0.0342*** 0.0342***
(0.00449) (0.00449) (0.00449) (0.00449) (0.00450)

Observations 9,923,378 9,923,378 9,923,378 9,923,378 9,923,378

R-squared 0.701 0.701 0.701 0.701 0.701

Other Skills controls X X X X X

Year-month FE X X X X X
CBSA× SOC FE X X X X X
Firm FE X X X X X

51
Table A7. Firm AI Adoption and Non-AI Wages by Occupation.
The table presents the results of the regressions of log wages in non-AI vacancies on the firm-level AI Share by the top 2-digit SOC occupations
ranked by AI Share. The sample is a full BGT data in 2016-2019 period, excluding AI vacancies, internships, and vacancies with missing salary
data. Each observation is one vacancy. The main independent variable, Firm-Level AI Share, is calculated for each firm-year. Columns are
labeled using the following occupation names and abbreviations: Computer & Math - Computer and Mathematical (SOC 15), Architecture
and Engineering (SOC 17), Science - Life, Physical, and Social Science (SOC 19), Management (SOC 11), Legal (SOC 23), Business & Financial -
Business and Financial Operations (SOC 13). Standard errors in parentheses are clustered by CBSA and 3-digit NAICS industry. Significance
levels are denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Computer & Architecture & Business &

Science Management Legal
Math Engineering Financial
(1) (2) (3) (4) (5) (6)

Firm-Level AI Share 0.196** -0.00773 0.318 0.187* 1.081 0.356**

(0.0821) (0.129) (0.323) (0.0947) (1.053) (0.136)

Software 0.141*** 0.0341*** 0.0173 0.0345*** 0.00162 0.0305***

(0.0130) (0.00991) (0.0120) (0.00664) (0.0105) (0.00560)
Computer -0.101*** -0.0267*** -0.0820*** -0.0858*** -0.0659*** -0.0696***
(0.00471) (0.00534) (0.0157) (0.0112) (0.0124) (0.00531)
Cognitive 0.0217** 0.0319*** 0.0250*** 0.0505*** 0.0396** 0.0418***
(0.0101) (0.00604) (0.00852) (0.00657) (0.0169) (0.00611)
Social 0.0329*** 0.0264*** 0.0328*** 0.0415*** 0.0231*** 0.0328***
(0.00628) (0.00373) (0.00596) (0.00407) (0.00839) (0.00398)

Observations 527,291 190,566 108,382 951,049 62,286 532,732

R-squared 0.610 0.731 0.739 0.627 0.757 0.634

Other Skills controls X X X X X X

Year-month FE X X X X X X
CBSA× SOC FE X X X X X X
Firm FE X X X X X X

52
Table A8. Firm AI Adoption and Non-AI Wages by Industry.
The table presents the results of the regressions of log wages in non-AI vacancies on the firm-level AI Share by the top 2-digit NAICS industries
ranked by AI Share. The sample is a full BGT data in 2016-2019 period, excluding AI vacancies, internships, and vacancies with missing salary
data. Each observation is one vacancy. The main independent variable, Firm-Level AI Share, is calculated for each firm-year. Columns are
labeled using the following 2-digit NAICS industry names and abbreviations: Information (NAICS 51), Professional Services - Professional,
Scientific, and Technical Services (NAICS 54), Finance & Insurance - Finance and Insurance (NAICS 52), Administrative & Support - Adminis-
trative and Support Services (NAICS 56), Agriculture - Agriculture, Forestry, Fishing and Hunting (NAICS 11), Manufacturing (NAICS 31-33).
Standard errors in parentheses are clustered by CBSA and 6-digit SOC occupations. Significance levels are denoted as: *** p<0.01, ** p<0.05,
* p<0.1.

Dependent Variable: Non-AI Log Wage

Professional Finance & Admin &

Information Agriculture Manufacturing
Services Insurance Support
(1) (2) (3) (4) (5) (6)

Firm-Level AI Share -0.0161 0.169*** 0.162 0.323** -0.0840 0.00361

(0.405) (0.0464) (0.121) (0.159) (0.277) (0.149)

Software 0.113*** 0.0793*** 0.0567*** 0.0429*** 0.0249** 0.0658***

(0.0130) (0.00791) (0.0121) (0.0127) (0.0125) (0.00808)
Computer -0.0824*** -0.0823*** -0.0780*** -0.0222** -0.0300 -0.0324***
(0.00777) (0.00488) (0.00645) (0.0111) (0.0192) (0.00702)
Cognitive 0.0260*** 0.0194*** 0.0153** 0.0182 0.0171 0.0289***
(0.0100) (0.00466) (0.00745) (0.0112) (0.0118) (0.00529)
Social 0.0135 0.0284*** -0.00141 0.0213*** 0.0406*** 0.0265***
(0.0108) (0.00297) (0.00920) (0.00273) (0.00977) (0.00446)

Observations 237,674 817,165 567,675 674,399 21,400 727,869

R-squared 0.789 0.714 0.714 0.761 0.870 0.735

Other Skills controls X X X X X X

Year-month FE X X X X X X
CBSA× SOC FE X X X X X X
Firm FE X X X X X X

53
Table A9. Firm AI Adoption and Non-AI Wages - Compustat-matched Sample.
The table presents regressions of log wages in non-AI vacancies on the firm-level AI Share and firm-level characteristics. The sample is
BGT data in 2016-2019 period, excluding internships and vacancies with missing salary data, matched to Compustat firms that have data on
required firm characteristics. Each observation is one vacancy. The main independent variable, Firm-Level AI Share, is calculated for each
firm-year. Industry is defined as 3-digit NAICS industry, SOC is a 6-digit SOC code. Standard errors in parentheses are clustered by CBSA,
SOC, and 3-digit NAICS. Significance levels are denoted as: *** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Non-AI Log Wage

(1) (2) (3) (4)

Firm-Level AI Share 1.344** 1.405 0.609 0.525

(0.604) (1.087) (0.537) (0.626)

Software 0.111*** 0.201*** 0.0920*** 0.0529***

(0.0207) (0.0334) (0.0159) (0.0122)
Computer -0.0662*** -0.0834*** -0.0414** -0.0219**
(0.0223) (0.0311) (0.0192) (0.0100)
Cognitive 0.00387 0.0500*** 0.0214** 0.00766
(0.0213) (0.0181) (0.0102) (0.00832)
Social 0.0128 0.0263 0.0232 -0.00165
(0.0209) (0.0182) (0.0195) (0.0117)

Log(Market Cap) 0.0168*** 0.00989 0.0228 0.0483***

(0.00622) (0.0338) (0.0169) (0.0172)
ROA -1.074** 0.154 -0.0755 -0.359***
(0.469) (0.268) (0.138) (0.113)
Cash/Assets 0.202 0.0373 0.314 0.274
(0.203) (0.481) (0.434) (0.394)
Leverage 0.173** 0.0955 0.0638 -0.0497
(0.0859) (0.225) (0.142) (0.145)
R&D/Sales -0.318* -1.624** -1.056 -0.905
(0.162) (0.779) (0.809) (0.842)
Capex/Sales -0.282 -0.457 -0.382 -0.0878
(0.238) (0.434) (0.264) (0.242)
PPE/Assets -0.148* -0.149 -0.0293 -0.0904
(0.0885) (0.290) (0.278) (0.348)
Log(Vacancies) -0.0215** -0.0222 -0.0315*** -0.0290***
(0.0104) (0.0164) (0.0116) (0.0104)

Observations 394,019 408,962 393,694 276,935

R-squared 0.667 0.613 0.759 0.811

Other Skills controls X X X X

Year-month FE X X X X
Industry FE X
CBSA× SOC FE X X X
Firm FE X X X
Job Title FE X

54
Table A10. AI Share and The Demand for Other Skills Controlling for Skill Count.
The table presents the results of the linear probability regressions of the demand for other skills on the firm-level AI Share. The sample is
a full BGT data in 2016-2019 period, excluding AI vacancies and internships. Each observation is one vacancy. The dependent variable is a
dummy equal 1 if the skill is required by the vacancy and zero otherwise. The main independent variable, Firm-Level AI Share, is calculated
for each firm-year. Standard errors in parentheses are clustered by CBSA, 6-digit SOC, and 3-digit NAICS. Significance levels are denoted as:
*** p<0.01, ** p<0.05, * p<0.1.

Dependent Variable: Skill Dummy (Non-AI Vacancies)

Customer Project People Finance &
Software Computer Cognitive Social Character Writing
Service Management Management Accounting
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Panel A: Industry & Year-Month fixed effects

Firm-Level AI Share 2.919*** -1.270*** 1.301*** 0.395** -1.125*** -0.0264 -1.848*** 0.907*** 0.200 -0.148
(0.312) (0.240) (0.251) (0.181) (0.391) (0.102) (0.350) (0.222) (0.159) (0.220)
Skill Count 0.0195*** 0.0250*** 0.0291*** 0.0326*** 0.0191*** 0.00630*** 0.0178*** 0.00938*** 0.0156*** 0.0157***
(0.00154) (0.00165) (0.00113) (0.00161) (0.00122) (0.000716) (0.00234) (0.00105) (0.00126) (0.00147)

Observations 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778 54,115,778
R-squared 0.268 0.223 0.278 0.286 0.143 0.072 0.187 0.129 0.112 0.182

Year-month FE X X X X X X X X X X
Industry FE X X X X X X X X X X

Panel B: Employer & Year-Month fixed effects

Firm-Level AI Share 0.479*** -0.241** 0.130 0.00646 -0.103 -0.0332 -0.429** 0.137 0.0984* -0.0303
(0.125) (0.0941) (0.102) (0.0723) (0.108) (0.0342) (0.169) (0.0968) (0.0532) (0.0753)
Skill Count 0.0191*** 0.0234*** 0.0265*** 0.0283*** 0.0176*** 0.00615*** 0.0158*** 0.00918*** 0.0157*** 0.0157***
(0.00135) (0.00164) (0.00124) (0.00171) (0.00120) (0.000740) (0.00212) (0.000887) (0.00141) (0.00159)

Observations 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320 46,309,320
R-squared 0.405 0.350 0.389 0.423 0.318 0.172 0.346 0.205 0.199 0.272

Year-month FE X X X X X X X X X X
Firm FE X X X X X X X X X X

55
B Theoretical Framework - Solution

In Section 3.2, we provided the setup of the theoretical framework. In this section of Appendix, we

provide the model’s solution. We first obtain an expression for the marginal cost, µ, as a function of the

productivity parameter z. This will allow us to substitute this expression for µ in the profit equation.

Thus, we first solve a cost minimization problem. The firm minimizes the production costs, which equal

the sum of the costs required to produce each task’s output:

Z 1
min p( x )y( x )dx, (8)
0

subject to constraint

h Z 1 i
exp z ln y( x )dx = 1.
0

Alternatively, taking the natural logarithm of both sides, the constraint can be expressed as:

Z 1
z ln y( x )dx = 0. (9)
0

The Lagrangian of this problem is

Z I∗ Z 1 Z 1
r w
L= y( x )dx + y( x )dx − µ[z ln y( x )dx ].
0 γA (x) I∗ γL ( x) 0

Taking derivative of L with respect to y( x ) and rearranging, the expression for y( x ) is:




zµ γ A (x) if x ∈ [0, I ∗ ]
r
y( x ) =


zµ γL (x) if x ∈ ( I ∗ , 1]
w

Then, taking a natural logarithm and integrating, we obtain:

Z 1 Z I∗ Z 1
ln y( x )dx = (ln z + ln µ + ln γ A ( x ) − ln r )dx + (ln z + ln µ + ln γL ( x ) − ln w)dx =
0 0 I∗

56
 Z I∗ Z I∗ Z 1 Z 1
= ln z + ln µ + ln γ A ( x )dx − ln rdx + ln γL ( x )dx − ln wdx = 0.
0 0 I∗ I∗

Rearranging, we get the expression for ln µ:

Z I∗ Z 1 
ln µ = −ln z − ln γ A ( x )dx + ln γL ( x )dx + I ∗ ln r + (1 − I ∗ )ln w,
0 I∗

which after calculating an exponent gives us the following expression for µ:

I∗
w wr
µ= , (10)
z exp(Γ)

R I∗ R1
where Γ represents 0
ln γ A ( x )dx + I∗
ln γL ( x )dx.

Solving the cost minimization problem gave us the minimum cost of producing the given level of

output. Having obtained it, we can return to the profit maximization problem - we will find an expres-

sion for the optimal output, Q, as a function of marginal cost and the demand schedule parameters:

∂π ∂P( Q) 1 1
= Q + P( Q) − µ = αQ− θ (1 − ) − µ = 0. (11)
∂Q ∂Q θ

Then, Q can be expressed as

 µ −θ
Q= . (12)
α(1 − 1θ )

Substituting expressions (13) and (15) into profit equation in (5), we can rewrite the firm’s profit equa-

tion as a function of the chosen automation threshold I ∗ , market demand-shifter α, firm’s productivity

z, and the fixed automation costs f :

1
 1 θ 1
π ( I ∗ , α, z, f ) = αQ− θ Q − Qµ − f I ∗ = αθ µ1−θ 1 − − f I∗ =
θ θ−1

h w r
I i 1− θ 
∗

w 1 θ 1
=α θ
1− − f I∗. (13)
z exp(Γ) θ θ−1

57
The expression for the derivative of π with respect to threshold I ∗ is then:

∂π ∂µ1−θ θ  1 θ 1 −θ ∂µ θ
 1 θ 1
= α 1 − − f = ( 1 − θ ) µ α 1 − − f.
∂I ∗ ∂I ∗ θ θ−1 ∂I ∗ θ θ−1

Derivative ∂µ/∂I ∗ is:

∗
w r I

∗
w r I
∂µ z w exp(Γ) ln wr − z w exp(Γ)(ln γ A ( I ∗ ) − ln γL ( I ∗ ))
= =
∂I ∗ exp(Γ)2

∗
w r I

z w ln γ (r I ∗ ) − ln w
γL ( I ∗ )  r w 
A
= = µ ln − ln .
exp(Γ) γA ( I ∗ ) γL ( I ∗ )

Then, the resulting expression for the derivative of π with respect to the threshold I ∗ is:

∂π  1 θ 1  r w 
= α θ
1 − (1 − θ )µ−θ µ ln − ln −f =
∂I ∗ θ θ−1 γA ( I )
∗ γL ( I ∗ )

 1  θ 1− θ  w r 
= αθ 1 − µ ln − ln − f. (14)
θ γL ( I )
∗ γA ( I ∗ )

As stated earlier, we want to understand how threshold task I ∗ will change when firm size and

fixed automation costs change. That is, we look for the comparative statics of I ∗ with respect to the

market demand, α, the firm’s productivity, z, and the level of the fixed cost of automation, f . A common

method for obtaining comparative statics is applying the implicit function theorem. We, however, rely

on the theory by Topkis (1978) that allows to reach the same comparative statics results as the implicit

function theorem but involves fewer assumptions. The theory shows that if the problem satisfies the

required assumptions, it will be sufficient to identify the signs of the cross partial derivatives ∂2 π/∂I ∗ ∂α,

∂2 π/∂I ∗ ∂z, and ∂2 π/∂I ∗ ∂ f in order to show the signs of ∂I ∗ /∂α, ∂I ∗ /∂z, and ∂I ∗ /∂ f respectively.

To use the Topkis (1978)’s theorem, we have to show that π ( I ∗ , α, z, f ) is upper semi-continuous and

that the choice set is a lattice, i.e. a partially ordered set. The former condition is satisfied since π is a

continuous function. The latter condition is also clearly satisfied since tasks are ordered by index and

are located on a unit interval [0,1] - the choice set is a bounded lattice. Thus, if the partial cross derivative

58
of π with respect to the parameter (i.e., α, z, or f ) is positive, then I ∗ is non-decreasing in the parameter,

and if the partial cross derivative is negative, then I ∗ is non-increasing.

1. Change in I ∗ with respect to α:

 θ  
∂2 π ∂[αθ 1 − 1θ µ1−θ ln w
γL ( I ∗ )
− ln r
γA ( I ∗ )
− f] ∂αθ  1  θ 1− θ  w r 
= = 1− µ ln − ln =
∂I ∗ ∂α ∂α ∂α θ γL ( I )
∗ γA ( I ∗ )

1
θ 1− θ  w r 
= θαθ −1 1 −
µ ln − ln > 0. (15)
θ γL ( I )
∗ γA ( I ∗ )
 
Since θ > 1 (elastic demand), α > 0, µ > 0, and ln γ w( I ∗ ) − ln γ (r I ∗ ) > 0 (from Assumption 1), the
L A

expression is positive. Therefore, the automation threshold I∗ is non-decreasing in α. Thus, firms

facing a larger market size are likely to automate more.

2. Change in I ∗ with respect to z:

 θ  
∂2 π ∂[αθ 1 − 1θ µ1−θ ln w
γL ( I ∗ )
− ln r
γA ( I ∗ )
− f] ∂µ1−θ θ 1
θ  w r 
= = α 1− ln − ln =
∂I ∗ ∂z ∂z ∂z θ γL ( I )
∗ γA ( I ∗ )

1 1
θ  w r 
= − (1 − θ ) µ 1− θ α θ 1 − ln − ln =
z θ γL ( I )
∗ γA ( I ∗ )

θ − 1 1− θ θ 1
θ  w r 
µ α 1−
= ln − ln > 0. (16)
z θ γL ( I )
∗ γA ( I ∗ )
 
Since z > 0, θ > 1 (elastic demand), µ > 0, α > 0, and ln γ w( I ∗ ) − ln γ (r I ∗ ) > 0 (from Assumption 1),
L A

the expression is positive. Therefore, the automation threshold I∗ is non-decreasing in z. Thus, more

productive firms are likely to automate more.

3. Change in I ∗ with respect to f :

 θ  
∂2 π ∂[αθ 1 − 1θ µ1−θ ln w
γL ( I ∗ )
− ln r
γA ( I ∗ )
− f]
= = −1 < 0. (17)
∂I ∗ ∂ f ∂f

59
Therefore, the automation threshold I ∗ is non-increasing in f . Thus, firms facing higher fixed costs

of automation will automate less.

60