Complement or substitute? How AI increases the demand for human skills

Introduction

Recent technological advancements and their adoption have significant implications for labour markets (e.g., [2]; [1]; [7]; [18]; [20]; [28]; [46]). Much of prior work has focused on investigating automation potential and has accrued significant evidence for the use of artificial intelligence (AI) for substitution of routine cognitive and manual skills (e.g., [18]; [15]). More recent evidence has further suggested that these effects now extend to non-routine cognitive tasks that have traditionally been the realm of highly educated and highly paid workers ([23]; [49]; [50]; [64]). Conversely, a smaller body of literature has shown evidence for augmentation effects such as improved productivity or efficiency for human workers in analytical, programming or writing tasks ([7]; [38]; [44]; [71]) or new complementation effects such higher demand or valuation for specific skills used alongside AI ([11]; [78]). In sum, there exists little consensus regarding what skills AI systems excel in, and consequently little agreement regarding specific expected labour market outcomes ([29]). This research seeks to shed light on the nature of AI-driven labour market effects – with particular emphasis on these less-understood complementation effects – and the dynamics of how and where these changes occur.

In order to balance the benefits and concerns associated with “AI at work” ([27]), researchers, policymakers, employers and job-seekers alike require better, empirically-grounded, understanding of these important effects. To date, however, much prior work has had a speculative forecasting approach (e.g., [9]; [80]) and relied on expert opinions or theoretical models for making claims, where the underlying assumptions may not be validated nor broadly accepted ([84]). This may limit the practical applicability of these works for policymakers seeking to inform better labour market policy, employers seeking to engage in strategic workforce planning, and job-seekers seeking to improve their employment prospects. In addition, perhaps given the very recent step change in AI adoption ([66]; [82]), there have thus far been “substantially fewer” studies examining AI ([27]: 4) compared to those investigating the effects of other technologies. Building on prior work demonstrating a strong association between technology adoption and changes in skills demands (e.g., [2]), and following the tenets of skill-biased technological change theory, we assume that artificial intelligence might lead to both complementing and substituting certain skills.

Gathering empirical evidence to measure the effects of technology on labour markets is not easy ([46]). Following [2] and [39], we assume that changes in demand for and valuation of skills can be observed from the ‘footprints’ left by job vacancy postings as employers communicate their evolving skills demands and their willingness to pay for specific skills. This study leverages a dataset of around 12 million ($n=11{,}729{,}662$) online job vacancy postings from the United States spanning a period of January 2018 to December 2023. Scraped from over 65,000 websites globally, including national and local job boards, company career pages, and job vacancy aggregators (Lightcast Data, n.d.), these postings cover the “near-universe” of online vacancies ([2], p.1). In what follows, our paper begins with a question at the forefront of the minds of both economists and policymakers—whether AI technologies are leading to substitution (reducing work) or complementation (increasing demand for specific types of work).

Turning to our findings, in the first part of the paper we first show a rise in demand for AI roles in the data, that is stronger than the general increase in labour demand. Next, we reveal a significant increase in demand for skills identified as complementary to AI technologies, such as digital literacy, team work, or resilience and agility, alongside a rising premium for these skills within AI roles, such as data scientists. Conversely, skills that are considered substitutable by AI, like customer service, summarisation or text review, have declined both in popularity and value within AI-related positions. Finally, we take a deeper look at complementary skills and show that for some skills, both the demand and the premium are more pronounced, that certain combinations of AI skills and complementary skills are more popular, and that there exists a compounding effect of complementary skills where combinations of specific complementary skills may be both more popular and profitable. We further illustrate this with specific AI roles, such as data scientists, to exemplify the compounding effect of complementary skills.

In the second half of this paper, we turn to expanding our analysis to measure the external effects across the broader economy, showing that AI impacts complementary and substitute skills even outside AI roles. Centrally, we find statistically significant external effects at the occupation, industry, and regional level. Specifically, growth in AI roles is associated with increased demand for non-AI complementary roles and a decline in demand for non-AI substitutable roles. Importantly, we find that the complementarity effect is considerably larger than the substitution effect. We repeat this analysis for the UK and Australia confirming the robustness of our findings. Finally, we conclude and discuss implications for stakeholders.

Literature Review

Technological Change and Labour Market Evolution

The question of whether AI is used to complement or substitute human labour is a central debate in the literature. Technological advancements have continually reshaped labour markets, from the mechanisation of the Industrial Revolution to contemporary digital transformations, each has redefined the skills required in the workforce and transformed the economic landscape ([47]; [52]; [75]). First, the Industrial Revolution marked a significant shift as production was mechanised and modernised, creating new manufacturing jobs, dramatically altering workforce structures, catalysing urbanisation and population growth and reshaping the economic foundations of society ([77]). Second, the 20th century saw rapid scientific progress, electrification and the emergence of automation technologies for industry standardisation and mass production, particularly in manufacturing where robotics began to undertake tasks previously performed by humans ([83]). This era saw the displacement of some routine jobs but also the expansion of the service sector, which offered new employment opportunities ([79]). These transitions illustrate the dual effects of technology on labour markets – both destructive and generative. Thirdly, the “Information Age” (e.g., [36]; [57]), underpinned by the development of semiconductors and computing technologies ([24]) saw sources of productivity, innovation and competitiveness in the economy increasingly depend on information technology ([36]).

Finally, recent technological advancement and adoption, also labelled the “fourth industrial revolution”, have continued to reshape labour markets as machines are used for increasingly complicated tasks through a combination of AI, increased computing power and connectivity, and cyber-physical systems such as IoT (e.g., [17]; [75]). Subsequent literature has investigated the impacts of such technologies on employment and wages, especially on routine manual and cognitive tasks (e.g. [18]; [20]; [28]). Though some research has provided evidence for positive impacts, such as improved productivity ([7]) or reorientation to “meaningful” work tasks ([72]) much of the literature warns of its risks or negative effects. These include wage inequality ([17]; [63]), (risks of) worker displacement through automation (e.g., [6]; [48]), and employment polarisation or a “hollowing-out” of the labour market as these changes differently effect high- and low-skilled workers (e.g., [28]). Though “the periodic warning that automation and new technologies are going to replace large numbers of jobs is a recurring theme in economic literature” ([78] p.2), the growing volume of studies on these risks of new technologies to equitable labour markets underscores the importance of research into changes in employment and wages in an context of current technological changes – especially with the recent step-change in artificial intelligence development and adoption ([66]; [82]).

Understanding AI’s Impact on Skills and Work

However, studies on the labour market effects of artificial intelligence on skills are substantially fewer in number ([2]; [27]), with some notable recent exceptions. For example, a Science-published experimental study found evidence of (generative) AI augmentation of human workers for writing tasks ([71]) and another experiment conducted by Harvard researchers at management consultancy Boston Consulting Group found evidence of augmentation in select knowledge work tasks ([38]). Meanwhile, Grennan and Michaely find evidence of AI substitution effects on security analysts, with consequences for employment and compensation ([54]), and Bessen et al show that over 50% of AI startups claim to automate routine work and decrease labour expenses for clients ([22]), suggesting a tension between evidence for augmentation and automation effects, and underscoring the importance of untangling these important effects.

This question of whether the effects of AI are to substitute or complement human work is central to debates on the future of work. An initial wave of occupation-based research focused on the automation of entire occupations (e.g., Bessen, 2016). This approach was instrumental in initiating early debates on the impacts of technology on employment, but has been criticised for its treatment of occupations as homogeneous entities, which risks simplifying complex dynamics occurring within occupations ([46]). Subsequent waves applied a task-based approach (e.g., [13]; [18]; [16]) which focuses on evidence for augmentation or automation of specific tasks within occupations. For example, formal reasoning tasks within Professional, Managerial, and Technical occupations ([1]).

Finally, the skills-based approach (e.g., [1]) suggests that the susceptibility of automation and augmentation of a role is defined by its skills. Proponents of the skills-based approach argue that as skills are used to perform tasks and occupations are best understood as bundles of skills, this approach is both intuitive and powerful in reducing any risk of over-simplifying complex dynamics ([18]; [46]; [78]).

Substitution, Augmentation, and Skill Complementarities in the AI Era

Prior work has accrued significant evidence for the use of AI for automation or substitution of routine cognitive and manual skills (e.g., [18]; [15]). A smaller body of more recent evidence suggests that artificial intelligence, specifically large language models (LLMs), can be used to generate human-level analytical, problem-solving, decision-making, creative, and writing output ([23]; [50]) including passing economics and medical licensing examinations ([49]; [64]) – non-routine cognitive tasks that have traditionally been the realm of highly educated and highly paid workers ([38]).

Yet, other recent work has claimed that AI can be used to augment knowledge work. This can include improving efficiency or accuracy for skills such as analytical, programming or writing skills ([38]; [44]; [71]) or creating demand for new skills ([21]). Others have found evidence for higher demand or valuation for specific skills used alongside AI ([11]; [78]), suggesting that complementarities exist between AI skills and other skills. Such skill complementarities have received growing research attention, for example, in measuring skill synergies via their co-occurrence in job requirements (Anderson, 2017), observing complementarities (and substitutions) of knowledge in the semiconductor industry as measured through patent specifications (Dibiaggio et al., 2014), and investigating of how clustering of complementary skills is driving persistent workforce polarisation and economic inequality (Alabdulkareem et al., 2018). As Neffke (2019) concludes, collectively this body of research underscores that skills and technological expertise are inherently interlinked via interdependencies among skill or knowledge elements.

It is important to note that there are two ways to look at complementation and substitution: popularity (how much this type of work is in demand) and valuation (how much this type of work is being valued). For instance, some research suggests that while substitution of work may not be observable in demand changes, it could be evident in devaluation. For example, Goldin and Katz’ seminal work on wages in the United States links wage differentials to skill-biased technological change (Goldin & Katz, 2007a), and Bessen links computerised occupations with “substantially greater” wage inequality (Bessen, 2016, p.1).

In sum, there is no consensus regarding what skills AI systems excel in, and thus no consensus on specific expected labour market outcomes (Brynjolfsson & Mitchell, 2017). In fact, it is most likely that substitution and complementation effects co-exist and that AI adoption is associated with both a decline in demand and valuation for some skills, as well as increasing demand for related skills or new skills (Acemoglu & Restrepo, 2018b; Autor et al., 2003). As succinctly explained by Acemoglu et al:

“…if different tasks require different skills, then the adoption of AI technologies may also change the set of skills that exposed establishments demand (and list in their vacancies). Skills relevant for tasks now performed by algorithms will be demanded less frequently, and new skills necessary for working alongside AI algorithms may also start being included in vacancies” (Acemoglu et al., 2021, p.8)

Unresolved Questions in AI’s Skill Impacts

Finally, scholars have drawn attention to two key shortcomings of prior work.

Firstly, while this literature review has sought to refer only to studies that adhere to the highest standards in research design, there is a large portion of the literature (as well as public press) that relies on theoretical models, expert opinions and predictions for their claims, where the underlying assumptions may not be validated nor broadly accepted (Yilmaz et al., 2023). Much prior work on the impact of AI seeks to forecast substitution effects on these bases (e.g., Agrawal et al., 2019; Webb, 2019), with vastly varying magnitudes and directions in these forecasts. This approach may introduce bias for at least two reasons – even among experts there exists no widely shared consensus on the skill-complementing or skill-substituting effects of AI (Brynjolfsson & Mitchell, 2017; Frank et al., 2019), and even experts underestimate the variety of skills deployed within occupations, which may cause over- or underestimations (Arntz et al., 2016). Gathering empirical evidence to generate and prove hypotheses is not easy – as explained by Frank, Autor, Bessen, Brynjolfsson and others, “major barriers that inhibit scientists from measuring the effects of AI on the future of work … include the lack of high-quality data about the nature of work (e.g., the dynamic requirements of occupations) [and] lack of empirically informed models of key microlevel processes (e.g., skill substitution and human-machine complementarity)” (Frank et al., 2019, p.6531).

Second, and relatedly, there is little consensus regarding what skills AI systems excel in or expected labour market outcomes (e.g., Acemoglu & Restrepo, 2018b; Autor et al., 2003; Brynjolfsson & Mitchell, 2017), with fierce debates as to which skills, tasks and occupations are affected by AI, in what ways, and to what extent. The strongest evidence points to complex dual substitution and complementation effects, with diminishing demand for substitution of skills now executable by AI and increasing demand for skills required for working in combination with AI (e.g., Acemoglu et al., 2021).

Building on prior work which has demonstrated a strong association between technology adoption and changes in skills demands (Acemoglu et al., 2021) and mindful of the shortcomings laid out above, this paper seeks to extend our understanding of this emerging and important field of research. In light of these unresolved questions and the critical need for empirical evidence, this study seeks to determine whether the growing demand for AI-related work not only reshapes skill requirements within AI-intensive roles but also exerts a broader influence on associated labor markets. More specifically, we ask: Does the increasing demand for AI work enhance the demand and valuation of skills that complement AI, both within AI-focused positions and in their surrounding ecosystems of related occupations, industries, and regions?

Hypotheses

It is clear that the growing integration of AI technologies into various facets of economic activity has important implications for the skill sets demanded in the labor market. We conceptualize these implications as arising through both “internal” and “external” effects on skill requirements. Internal effects refer to changes in skill demands within AI-intensive roles, while external effects describe how AI influences skill requirements beyond strictly AI-focused occupations, extending to other workers within an occupation, across competing firms in an industry, and even across entire regions.

Internal Effects

Within roles that directly engage with AI—such as data scientists or machine learning engineers—the advent of increasingly complex and opaque models alters the bundle of skills needed for effective job performance. For example, when working with advanced AI algorithms that produce outcomes less readily interpretable than conventional statistical models, these workers must now possess stronger ethical reasoning skills to identify and mitigate potential biases and broader unintended consequences. Similarly, heightened emphasis on algorithmic accountability and interpretability may demand more substantial digital literacy, along with softer skills like resilience and adaptability. In economic terms, these internal effects can be understood as cross-derivative complementarities at the worker’s own “production function” level. Here, the productivity of the individual worker—the “output”—is determined by multiple skill inputs, and the marginal benefit of one skill (e.g., ethics) increases as another skill (e.g., AI technical proficiency) becomes more central. Thus, as the prevalence of AI-based tasks expands, the cross-partial derivatives of the worker-level production function become positive, indicating that the presence of AI skills enhances the return to these complementary non-technical skills.

External Effects at the Occupation, Industry, and Regional Levels

Beyond the core set of AI specialists, skill demands are influenced externally through spillover effects that affect other workers in the same occupation, competing firms within an industry, and broader economic regions. At the occupational level, even workers who do not interface directly with AI may need higher digital literacy or a fundamental understanding of algorithmic outputs. This ensures their ability to interpret, validate, and act upon AI-driven insights produced by colleagues in similar roles. At the industry level, when rival firms adopt AI technologies, competitive pressures may lead firms without a direct AI focus to invest in raising the overall technical proficiency of their workforce. Improving employees’ teamwork, resilience, and ethical understanding can further complement technology adoption, allowing these firms to respond nimbly and responsibly to AI-driven market changes. The positive cross-derivatives here operate at the firm’s production function level, where the presence of an AI-intensive competitor increases the marginal productivity of complementary skills within non-AI roles.

At the regional level, the spatial clustering of AI capabilities can produce broad-based skill shifts. In regions hosting AI hubs—think of areas specializing in autonomous driving or AI-driven recruitment practices—spillover effects transcend firm or industry boundaries. Local labor pools adapt as a whole; non-AI workers in the region may benefit from digital upskilling and an enhanced understanding of AI’s ethical and social implications. Conceptually, this can be modeled through a “regional production function,” where the complementarities extend beyond firm-level interactions and involve multiple employers, educational institutions, and related services. As AI permeates the regional economy, the marginal value of digital literacy, teamwork, resilience, and ethics skills rises across a variety of roles, reflecting positive cross-partial derivatives that span geographic boundaries.

To guide our investigation, we propose a set of hypotheses reflecting the “internal” and “external” effects of AI’s diffusion on skill demand.

Internal Effects (Within AI-Roles):

• H1a: Roles that require AI skills also exhibit a higher demand for skills that complement AI.

• H1b: Complementary skills within AI-intensive roles command higher wages.

• H1c: Roles that require AI skills show reduced demand for skills considered substitutes for AI.

External Effects (Spillover Effects Beyond AI-Roles):

• H2a: In regions, industries, or occupations characterized by high AI adoption, non-AI roles exhibit increased demand for complementary skills.

• H2b: In these same contexts, non-AI roles show a decrease in demand for substitute skills.

In sum, our approach integrates both “within-role” and “beyond-role” perspectives to capture the multifaceted nature of AI’s labor market impact. By examining how complementary and substitute skills evolve both inside and outside of AI-heavy contexts, this research contributes to a more nuanced understanding of the ways in which technological change influences not only immediate job requirements but also the broader ecosystems of talent and expertise.

Methods and Data

Data

Following prior work, we assume that changes in demand for and valuation of skills can be observed from the ‘footprints’ left by job vacancy postings as employers communicate their evolving skills demands and their willingness to pay for specific skills ([2]; [39]). This study leverages a dataset of around 12 million ($n=11,729,662$) online job vacancy postings in the United States from January 2018 to December 2023, provided by the Burning Glass Institute (BGI). The BGI is a global leader in labour market analytics and collects these postings by scraping over 65,000 online sources, including national and local job boards, company career pages, and job vacancy aggregators ([65]), to cover the “near-universe” of online vacancies ([2], p.1). Each posting includes information on e.g., title, occupation category, salary, and a detailed list of skills demanded for the position, which have been scraped from the text of the posting and organised according to around a thousand standardised labels.

Noteworthy prior research has made use of BGI data for understanding the effects of technology on the labour market (e.g., [11]; [2]; [43]; [56]; [19]; [58]). A key question raised (and addressed) in these works concerns the representativeness of online job vacancy data, leading me to conduct numerous tests to interrogate the representativeness of my data.

Acemoglu, Autor, Hazell and Restrepo ([2]) have demonstrated that BGI data closely mirror the trends in overall job vacancies in the United States, as reported by the Labor Turnover Survey (JOLTS) and the Bureau of Labor Statistics (BLS) Job Openings, and that its industry composition is closely aligned with industry-level vacancy statistics from JOLTS.

In addition, we further find that my dataset is closely aligned with key BLS statistics on job openings by state from the JOLTS and occupational composition from the Occupational Employment and Wage Statistics (OEWS) Survey (see Appendix Table 1-Table II). There is extremely little divergence in the state level-statistics and some divergence in the occupation category level statistics, such as overrepresentation of Computer and Mathematical Occupations and underrepresentation of Farming, Fishing, and Forestry Occupations. As highlighted by [35], this is characteristic of online vacancy postings data, which can overrepresent professional and technical occupations relative to blue collar occupations. However, overall, the BGI data is satisfactorily representative – the average difference compared to BLS statistics is only 3%. Moreover, due the focus of the research in comparing shares rather than absolute figures, and on knowledge work occupations, the impact of over- or under-representation is likely to be small.

Jobs that require AI skills are more likely to include complementary skill requirements, indicating that AI adoption correlates with greater emphasis on these supportive capabilities.
Regression Results: Demand for Complementary Skills
	Dependent variable: demand_complementary_skills
	Model 1 (individual controls)	Model 2 (job/structural controls)	Model 3 (full model)
	(1)	(2)	(3)
Is an AI job	-0.467*** (0.014)	0.482*** (0.016)	0.195*** (0.016)
Minimum years of experience	0.135*** (0.000)		0.096*** (0.001)
Minimum education level	0.370*** (0.001)		0.146*** (0.001)
Knowledge work occupation		0.208*** (0.003)	-0.014*** (0.003)
Full-time employment		0.405*** (0.004)	0.248*** (0.004)
Skill variety level		0.735*** (0.001)	0.718*** (0.001)
High GDP state		-0.009*** (0.003)	-0.011*** (0.003)
Year 2019		0.025*** (0.005)	0.038*** (0.005)
Year 2020		-0.116*** (0.005)	-0.074*** (0.005)
Year 2021		0.021*** (0.005)	0.060*** (0.005)
Year 2022		0.043*** (0.005)	0.074*** (0.005)
Year 2023		0.037*** (0.005)	0.075*** (0.005)
Observations	5292754	5292754	5292754
Pseudo R²	0.066	0.421	0.430
Note:			* p<0.1; ** p<0.05; *** p<0.01

State2	BGI Data	BLS Data
Alabama	1.4%	1.2%
Alaska	0.3%	0.2%
Arizona	2.2%	2.7%
Arkansas	0.9%	0.6%
California	9.9%	13.2%
Colorado	2.5%	2.8%
Connecticut	0.9%	1.1%
Delaware	0.4%	0.3%
Florida	6.5%	6.1%
Georgia	3.7%	3.2%
Hawaii	0.3%	0.4%
Idaho	0.6%	0.6%
Illinois	4.3%	3.9%
Indiana	2.0%	2.0%
Iowa	1.2%	1.0%
Kansas	1.0%	1.0%
Kentucky	1.4%	1.1%
Louisiana	1.4%	1.0%
Maine	0.4%	0.3%
Maryland	2.0%	1.8%
Massachusetts	2.8%	3.0%
Michigan	3.0%	2.9%
Minnesota	2.2%	2.1%
Mississippi	0.9%	0.5%
Missouri	1.9%	1.8%
Montana	0.4%	0.3%
Nebraska	0.7%	0.7%
Nevada	1.1%	1.1%
New Hampshire	0.4%	0.4%
New Jersey	2.1%	2.5%
New Mexico	0.7%	0.5%
New York	4.4%	4.5%
North Carolina	3.7%	3.3%
North Dakota	0.3%	0.2%
Ohio	4.0%	3.6%
Oklahoma	1.2%	1.1%
Oregon	1.1%	1.4%
Pennsylvania	3.8%	3.4%
Rhode Island	0.3%	0.3%
South Carolina	1.7%	1.1%
South Dakota	0.3%	0.2%
Tennessee	2.5%	2.0%
Texas	8.1%	8.6%
Utah	1.1%	1.0%
Vermont	0.2%	0.2%
Virginia	3.0%	3.0%
Washington	2.2%	3.3%
West Virginia	0.5%	0.2%
Wisconsin	2.0%	2.0%
Wyoming	0.2%	0.1%

Defining Skills

To define the complementary and substitutable skills for the analysis, we turned to literature. This review (e.g., Autor et al, 2008; [1]; [48]; [18]; Wirtz et al., 2018; Felten et al, 2019; Acemoglu & Restrepo, 2019; [46]; Colombo et al 2019) yielded five substitutable skills (clusters), which we labelled Basic Data Skills, Summary and Reporting, Language and Text Review, Customer Service and Office and Financial Administration, and seven complementary skills (clusters), which we labelled Analytical Thinking, Digital Literacy, Resilience and Agility, Technical Proficiency, Ethics, Self-Efficiency, and Working With Others. See Table III for literature per individual skills cluster.

	Model 1 (individual controls)	Model 2 (job/structural controls)	Model 3 (full model)
	(1)	(2)	(3)
Is an AI job	-0.766***	-0.454***	-0.427***
	(0.014)	(0.015)	(0.015)
Minimum years of experience	0.006***		-0.043***
	(0.000)		(0.000)
Minimum education level	0.222***		0.080***
	(0.001)		(0.001)
Knowledge work occupation		0.373***	0.349***
		(0.003)	(0.003)
Full-time employment		0.032***	0.059***
		(0.003)	(0.003)
Skill variety level		0.208***	0.208***
		(0.000)	(0.000)
High GDP state		0.013***	0.015***
		(0.002)	(0.002)
Year 2019		0.002	0.002
		(0.004)	(0.004)
Year 2020		-0.086***	-0.084***
		(0.004)	(0.004)
Year 2021		-0.063***	-0.063***
		(0.004)	(0.004)
Year 2022		-0.036***	-0.036***
		(0.004)	(0.004)
Year 2023		-0.043***	-0.042***
		(0.004)	(0.004)
Observations	5292754	5292754	5292754
Pseudo R²	0.016	0.158	0.160

Next, to ensure successful operationalisation of the research, we set out to verify which skills are covered by our BGI dataset, and identify skills that would be the closest match with those found in the literature review. To do so, we extracted a list of unique skills from the data and mapped these skills to those described in the literature. This yielded 18 individual skills from the BGI database to make up the 5 substitutable skills clusters (see Table IV) and 71 individual skills from the BGI database to make up the 7 complementary skills clusters (Table V). The final criteria for inclusion were (1) referenced in at least two high-quality publications, (2) have equivalent skills (same or sufficiently similar) in the data, (3) have a minimum sample size of $n=30$ job vacancies demanding that skill in the BGI dataset, a common threshold for statistical analysis ([59]).

	Model 1 (individual controls)	Model 2 (job/structural controls)	Model 3 (full model)
	(1)	(2)	(3)
Demand for Validated Substitutable Skills	-0.177{sup}***	-0.120{sup}***	-0.085{sup}***
	(0.013)	(0.015)	(0.014)
Minimum years of experience	0.047{sup}***		0.046{sup}***
	(0.002)		(0.002)
Minimum education level	0.036{sup}***		0.028{sup}***
	(0.004)		(0.003)
Knowledge work occupation		0.750{sup}***	0.606{sup}***
		(0.056)	(0.048)
Fortune 500 company		0.157{sup}***	0.119{sup}***
		(0.011)	(0.010)
Full-time employment		0.061{sup}***	0.047{sup}***
		(0.018)	(0.016)
Skill variety level		-0.009{sup}***	-0.014{sup}***
		(0.001)	(0.001)
High GDP state		0.097{sup}***	0.097{sup}***
		(0.010)	(0.009)
Year 2019		0.045	0.044
		(0.034)	(0.032)
Year 2020		0.060{sup}*	0.043
		(0.033)	(0.031)
Year 2021		0.027	0.008
		(0.029)	(0.027)
Year 2022		0.091{sup}***	0.040
		(0.029)	(0.027)
Year 2023		0.209{sup}***	0.172{sup}***
		(0.029)	(0.027)
Observations	5513	5513	5513
$R^2$	0.193	0.186	0.318
Adjusted $R^2$	0.192	0.184	0.316

	Model 1 (individual controls)	Model 2 (structural controls)	Model 3 (full model)
	(1)	(2)	(3)
Demand for Complementary Skills	-0.124*** (0.011)	-0.035*** (0.012)	-0.042*** (0.011)
Minimum years of experience	0.048*** (0.002)		0.047*** (0.002)
Minimum education level	0.039*** (0.004)		0.029*** (0.003)
Knowledge work occupation		0.744*** (0.056)	0.597*** (0.049)
Fortune 500 company		0.164*** (0.011)	0.123*** (0.010)
Full-time employment		0.071*** (0.018)	0.056*** (0.016)
Skill variety level		-0.012*** (0.001)	-0.016*** (0.001)
High GDP state		0.098*** (0.010)	0.098*** (0.009)
Year 2019		0.046 (0.034)	0.046 (0.032)
Year 2020		0.063* (0.034)	0.046 (0.031)
Year 2021		0.024 (0.030)	0.008 (0.027)
Year 2022		0.096*** (0.029)	0.045* (0.027)
Year 2023		0.213*** (0.029)	0.176*** (0.027)
Observations	5513	5513	5513
$R^2$	0.177	0.175	0.313
Adjusted $R^2$	0.177	0.173	0.312
Note:			* p<0.1; ** p<0.05; *** p<0.01

We label a posting as demanding complementary or substitutable skills if it demands at least one skill from the complementary or substitutable skills list. The justification for this twofold, both conceptual and practical. Firstly, the conceptual justification for this choice is clear – for example, job postings requiring “Critical Thinking” or “Digital Literacy” are both postings demanding complementary skills, irrespective of whether the job posting requires both those skills. Demand for at least one complementary or substitutable skill should be considered indication that that job is related to complementary or substitutable skills utilisation, and can be classified as such. In addition, some skills clusters contain only 1-2 equivalent BGI skills, such as Customer Service (“customer service”) or Ethics (“ethical standards and conduct, business ethics”), meaning these important skills categories would be eliminated from the analysis entirely were a higher threshold used. Secondly, as shown in Appendix Figures 1-2, the number of postings able to be included in the analysis would decline rapidly if a higher threshold of minimum two or more complementary or substitutable skills were to be applied.

Defining AI roles

To identify AI roles we use a skills-based approach, following the example of numerous researchers who apply this approach for understanding the effects of artificial intelligence jobs in the labour market (e.g. [1]; [11]; [43]). The approach uses skills required for a job for classification – thus a job vacancy posting that lists skill(s) to create or use AI as a requirement is considered to be an AI role. This avoids categorising vacancies in AI companies or the AI industry that are not truly AI roles (e.g. a Human Resources Manager at an AI company) as AI roles.

In line with the skills-based approach, in order to identify AI roles we first needed to identify AI skills. Not all AI roles are the same, and interesting variations are likely to exist across different categories of AI roles. To identify and classify subcategories of AI roles, we first leveraged BGI’s existing classification of AI roles into those using ‘AI User’ or ‘AI Creator’ skills ($n=115$ skills). Next, taking these 115 skills, we further classified each into a subcategory, consulting literature (e.g., [74], [76]) where our existing knowledge was not sufficient for classification. Any skill that was unclear or broadly phrased was verified to ensure its correct classification by rank-ordering and printing a list of occupations demanding that skill (at the highest specificity, SOC_5 level) to better understand its applications. For example, “ChatGPT”, initially classified as Natural Language Processing (NLP) in AI Creators was re-classified as NLP AI Users after the check showed that this skill was predominantly deployed by “writers and editors” and “sales managers” who are more likely to be users than creators of AI. Finally, we verified whether the dataset contained the skills in question, so that the AI role categories could be identified, which led to the dropping of one category, Computer Vision and Image Processing (AI Users). The final result is eight categories of AI roles, related to the creation (first five categories) or usage of AI (last three categories):

1. Machine learning creators

2. Natural language processing (NLP) creators

3. AI development and operations

4. Computer vision and image processing

5. Prediction and analytics

6. Machine learning users

7. Natural language processing (NLP) users

8. AI Applications and Tools

See Appendix Table VI for a full list of skills used to identify jobs in each category.

Across occupations (Model 1), industries (Model 3), and regions (Model 5), growth in AI jobs is associated with a statistically significant positive effect on the number of non-AI roles demanding complementary skills in that occupation, industry or region. Across regions (Model 6), growth in AI jobs is associated with a statistically significant negative effect on the number of non-AI roles demanding substitutable skills in that region. The same is observed in Models 2 and 4 for occupations and industries, though these findings are not statistically significant.

	Model 1 (Occupation, Complementary)	Model 2 (Occupation, Substitutable)	Model 3 (Industry, Complementary)	Model 4 (Industry, Substitutable)	Model 5 (Region, Complementary)	Model 6 (Region, Substitutable)
(1)		(2)	(3)	(4)	(5)	(6)
AI job postings	0.028*** (0.007)	-0.024 (0.023)	0.054*** (0.009)	-0.019 (0.029)	0.030*** (0.003)	-0.058*** (0.007)
Total job postings	0.839*** (0.023)	-1.007*** (0.197)	0.755*** (0.028)	-0.829*** (0.210)	0.708*** (0.012)	-0.374*** (0.065)
Non-AI substitutable job postings	0.203*** (0.020)		0.235*** (0.030)		0.256*** (0.012)
Non-AI complementary job postings		1.868*** (0.189)		1.726*** (0.218)		1.450*** (0.067)
Observations	165	165	96	96	805	805
R²	0.981	0.833	0.984	0.882	0.998	0.987
Adjusted R²	0.980	0.830	0.983	0.878	0.998	0.987
Note:						*p<0.1; **p<0.05; ***p<0.01

Similarly to the complementary and substitutable skills classification, we label a posting as an AI role posting if it demands at least one skill from the AI skills list. The justification for this twofold, both conceptual and practical. Firstly, the conceptual justification for this choice is clear – for example, job postings requiring “Generative Adversarial Networks” or “Large Language Modeling” are both AI role postings, irrespective of whether the job posting requires both those skills. Demand for at least one AI usage or creation skill should be considered indication that the job can be classified as an AI role. Secondly, as shown in Figure 3, the number of postings able to be included in the analysis would decline rapidly if a higher threshold of two or more AI skills were to be applied. The threshold of defining a job posting as an AI role posting when it leverages a minimum of one AI is consistent with other similar research (e.g., [43]; [76]).

Measures and variables

Demand for skills

The number of AI roles is far fewer ($n=47,871$) compared to non-AI roles ($n=11,681,791$). Therefore, to enable useful comparison between AI and non-AI roles, demand for skills must be computed as the proportion (share) of postings demanding skills (rather than using absolute values). A custom function is defined using Python to count the number of job postings that mention the skill ($S$) and dividing it by the total number of job postings in that category ($C$) (e.g., in a given AI role category), and communicate this as a share (%) out of 100.

Percentage Growth / Percentage Difference

To compare relative changes in demand or salary between two groups or time periods, we create a custom percentage growth (also used for comparative percentage difference) function. The percentage growth/difference is calculated using the following formula:

where $V1$ and $V2$ are the initial and final values, respectively.

Log-transformed salary

Salary data can be skewed, which could affect the results of statistical analyses. Therefore, before undertaking our salary analyses, we conducted some tests to validate if the salary variable requires transformation before its use in the analysis. A visual inspection of a histogram and Q-Q plot (see Appendix Figure 4) showed that the distribution of salary is right-skewed and the Q-Q plot is deviated from the line. Thus, to stabilise the variance, making it more suitable for my analysis, we log-transform (natural logarithm) the salary variable.

Methods

Time series analysis

In what follows, we first conduct a time series analysis to display trends and compare demand and salary patterns over time. This is an important step in setting the scene for the analysis, before answering the main research questions. First, we compare demand for complementary and for substitutable skills in AI roles compared to non-AI roles over the years 2018-2023. We begin by filtering data to include only knowledge work occupations as this is the focus of this paper, and because it provides a fairer comparison point than blue-collar occupations such as manufacturing or agriculture which are significantly less likely to demand skills such as office administration or technical proficiency. The share of postings demanding complementary and substitutable skills was calculated as a percentage of the total postings (for AI/non-AI), and visually presented over the years 2018-2023 using a line chart. In the second part of the analysis, we repeat this process for salary, comparing log salary for AI roles demanding complementary and for substitutable skills and non-AI roles with these skills over time.

Regression models

Regression as a method of analysis has an established precedent of being used to understand the relationships between artificial intelligence and patterns in demand and wages in the labour market (e.g. [2]; [78]). In this paper, we employ logistic regression to understand the associations between AI roles and demand for complementary and substitutable skills, and linear regression to understand if demand for complementary and substitutable skills is associated with higher or lower salaries in AI occupations (internal effects), as well as in the external effects analysis. A range of control variables are accounted for, which are presented in the table below, along with their justification based on literature review.

Control variable	Description/Operationalisation	Justification
Minimum years of experience (internal effects)	Minimum years of experience required, as specified by job posting (discrete variable)	Salary differs by experience level and different experience levels are associated with different skills demands [41, 67, 68]
Minimum education level (internal effects)	Minimum education level, as specified by job posting (None listed, High School or GED, Associate’s Degree, Bachelor’s Degree, Master’s Degree, PhD or professional degree) (ordinal variable)	Salaries differ by education level and skills demand is related to education levels [33, 43, 51, 81]
Knowledge work occupation (internal effects)	Sales, Management, Business and Financial Operations, Computer and Mathematics, Office and Administrative Support, Community and Social Service,	Labour/skills demand and salaries differ by occupation [30]
	Healthcare Practitioners and Technical Occupations, Architecture and Engineering, Educational Instruction and Libraries, Life, Physical, and Social Sciences, Arts, Design, Entertainment, Sports, and Media, Unclassified3, or Legal Occupations; or not (binary variable)
Full-time employment (internal effects)	Derived from “Employment type” column, job postings specified as “Full-time ($> 32$ hours)” or not (binary variable)	Salary differs between full-time and part-time or freelance employment ([45]; [61]), part-time or freelance work may require a focused set of skills ([78]), and supply demand trends differ between full-time and part-time or freelance jobs ([32]; [60])
Fortune 500 company (internal effects,	Job vacancy posted by a Fortune 500 company (company name matched using	Larger companies can pay higher salaries ([25])
salary regressions only)	fuzzymatch) or not (binary variable)
Skill variety level (internal effects)	Quantity of skills requested in a job posting (discrete variable)	Salary is related to skills level and skills variety / combinations (Acemoglu & Autor, 2011; Goldin & Katz, 2007a; Stephany & Teutloff, 2024) and skills are demanded at different levels and jobs can be more or less skills-intensive (Acemoglu & Autor, 2011; Green et al., 2003; Handel, 2020)01/08/2024 05:14:00
High GDP state (internal effects)	Job posted in any of ten states with largest Gross Domestic Product (GDP), using 2023 figures from the U.S. Bureau of Economic Analysis (binary variable)	Salaries and labour/skills demand differ by state (Bureau of Labor Statistics, 2023b)
Year (internal effects)	Year (control variables for 2019-2023 included, reference year is the omitted category 2018)	Labour/skills demand and salary fluctuate and differ year to year (Acemoglu & Restrepo, 2018c; Bureau of Labor Statistics, 2024b; Goldin & Katz, 2007a)
Total postings (external effects)	Total number of job postings per year per observation category (occupation, industry or region)	Captures overall labour market activity, ensuring that the observed effects of AI role postings are not confounded by broader
		variations in labour demand (e.g., market size fluctuations)
Converse skill category postings, i.e. complementary skills in substitutable regression, substitutable skills in complementary regression (external effects)	Number of non-AI complementary or substitutable job postings per year per observation category (occupation, industry or region)	Controls for the structural dynamics of skill ecosystems/interactions within the labour market – complementary and substitutable jobs exist in interdependent ecosystems where the presence or absence of one may affect the demand for the other.

In this study, logistic regression (Maximum Likelihood Estimation, MLE method) is used to understand the association between demand for complementary or substitutable skills (dependent variables) and AI roles (independent variable). Motivated by the theoretical precedent detailed in the section above, various controls are introduced. Model 1 controls for individual factors (education and experience levels) and Model 2 controls for job-related and structural factors (full-time employment, knowledge work occupation, skill variety level, high GDP state, year). Model 3 includes all control variables along with the independent variable of interest. Models 1-3 are fitted both for the complementary skills model and the substitute skills model. The regression formula is captured as follows:

Where:

• $p$ is the probability of demand for skills

• $\beta_{0}$ is the intercept

• $\beta_{1}$ represents the effect of a job being an AI role

• “Controls" represents a collective term for all other control variables included to adjust for confounding factors, which for Model 3 include minimum years of experience, minimum education level, knowledge work occupation, full-time employment, skill variety level, high GDP state, and year indicators (2019, 2020, 2021, 2022, 2023)

To complete our analysis of internal effects, we use linear regression (Ordinary Least Squares, OLS method) to understand the association between (log-transformed) salary for AI roles (dependent variable) and demand for complementary or substitutable skills (independent variables). Motivated by the theoretical precedent detailed in the section above, various controls are introduced. Model 1 controls for individual factors (education and experience levels) and Model 2 controls for job-related and structural factors (full-time employment, knowledge work occupation, company size, skill variety level, high GDP state, year). Model 3 includes all control variables along with the independent variable of interest. Models 1-3 are fitted both for the complementary skills model and the substitute skills model. The regression formula is captured as follows:

For substitutable skills:

For complementary skills:

Where:

• $\operatorname{log\_salary}$ is the natural logarithm of the salary for AI roles

• $\beta_{0}$ is the intercept

• $\beta_{1}$ represents the coefficient for "Demand for Substitutable Skills" or "Demand for Complementary Skills"

• “Controls” is a collective term for all other control variables in the model, which for Model 3 include minimum years of experience, minimum education level, knowledge work occupation, company size (Fortune 500 company or not), full-time employment, skill variety level, high GDP state, and year indicators (2019, 2020, 2021, 2022, 2023)

• $\varepsilon$ (epsilon) is the error term, which captures the variability in $\operatorname{log\_salary}$ not explained by the model

An additional control variable is included in the salary regressions (Fortune 500 company) owing to the need to adjust for the fact that the largest companies are likely to pay the highest salaries [25]. Company names in the job postings data set were first directly matched to Fortune 500 company names. However, company names are not always consistently written. For example, management consulting company McKinsey & Company appears in the data as “mckinsey”, “mckinsey & co inc united states”, and “mckinsey & company, inc. united states”. Consequently, we employed a fuzzy match using the FuzzyWuzzy package, which uses Levenshtein Distance to calculate the differences between sequences [37].⁴

Finally, we turn our attention to measuring external effects. We use linear regression (Ordinary Least Squares, OLS method) to understand the association between the penetration of AI (as measured by prevalence of AI role postings) and their external effects on the demand for complementary and substitutable skills also in non-AI roles, at the occupation, industry, and regional levels. For each of our six regressions, we aggregate data by the relevant category – occupation, industry, or region – and year, excluding those with fewer than ten AI role postings per year to ensure robust results. To linearise relationships, log transformations were applied to all variables. We control for the total number of postings and postings for the converse skill category (i.e. complementary skills in substitutable regression, substitutable skills in complementary regression) to ensure that the observed effects of AI role postings are not confounded by broader variations in labour demand (e.g., market size fluctuations) and account for structural dynamics of skill ecosystems/interactions within the labour market.

The regression equation is expressed as:

Where:

• $y_{i,k}$ is the dependent variable (Complementary Non-AI roles or Substitutable Non-AI roles) by year $i$ and level $k$ (occupation/industry/region)

• $X_{1,i,k}$ is total Postings (log-transformed)

• $X_{2,i,k}$ is AI Postings (log-transformed)

• $X_{3,i,k}$ is Substitutable or Complementary Non-AI Postings (log-transformed)

• $\beta_{0}$ is the intercept

• $\beta_{2}$ (AI postings coefficient) quantifies the elasticity of complementary or substitutable skill demand with respect to AI role postings.

• $\beta_{1}$ and $\beta_{3}$ control for

• $\epsilon_{i,k}$ is the error term

Lastly, we estimate the real-world impact of these findings, measuring external effects with the following formula:

Where:

• Coefficient refers to the relevant regression coefficient, as follows:

  • Occupation Level: $\beta_{\mathrm{complementary}} = 0.0282$, $\beta_{\mathrm{substitutable}} = -0.0237$

  • Industry Level: $\beta_{\mathrm{complementary}} = 0.0541$, $\beta_{\mathrm{substitutable}} = -0.0191$

  • Regional Level: $\beta_{\mathrm{complementary}} = 0.042$, $\beta_{\mathrm{substitutable}} = -0.0447$

• The Multiplier adjusts the effects computed using our BGI data to a real-word estimate using BLS data:

Comparative descriptive statistics analysis:

First, we employ comparative descriptive statistics to examine differences in demand for specific complementary skills in AI compared to non-AI roles, and for complementary skills in AI user jobs compared to AI creator jobs, visualised in simple line charts. Second, motivated by these nuanced findings, we further examine the differential demand for specific complementary or substitutable skills clusters between specific AI role categories compared to non-AI roles, with results visualised using a heatmap. In addition, motivated by recent findings (e.g., [66]; [82]) suggesting a recent major step change in AI adoption, specifically the use of (generative) AI to substitute and complement knowledge work, we further investigate the change in demand for complementary and substitutable skills clusters across AI role categories over time, from 2018-2022 to 2023.

In the second part of the analysis, we investigate the compensation of complementary or substitutable skills within various AI role categories. We begin by transforming the salary variable by taking its natural logarithm, resulting in a new variable, 'log_salary,' to address skewness. Next, we calculate the median log salary for each AI role category, and compare this to the median log salary for each AI role category with each complementary or substitutable skills cluster. The results were again visualised using a heatmap to easily depict the magnitude and direction of findings.

Finally, for the comparisons detailed above, we run statistical significance tests to ensure our observed results are statistically significant. For demand comparisons we employ a two-sided two-proportion $Z$-test for those comparisons fulfilling the test condition that each group should have at least 10 successes and 10 failures. The hypotheses are as follows:

$H_0$: $P_{\mathrm{AI}} = P_{\mathrm{non-AI}}$, the proportion of AI roles that require a specific skill is equal to the proportion of non-AI roles that require that skill

$H_A$: $P_{\mathrm{AI}} \ne P_{\mathrm{non-AI}}$, the proportions are not equal

For a small number of comparisons ($n=2$ for the complementary skills analysis and $n=5$ for the substitutable skills analysis) not fulfilling the test conditions, we use Fisher’s Exact test, which is particularly useful for small sample sizes and is used to test the equivalence or difference between two proportions [59].

For the salary analyses we employ a two-sided Mann-Whitney U test to check for statistical significance in differences, with the following hypotheses:

$H_0$: The distributions of both groups are equal, there is no difference in the medians of the two groups

$H_A$: The distributions of the two groups are not the same, there is a difference in the medians of the two groups

For a number of comparisons ($n=20$ for the complementary skills analysis and $n=21$ for the substitutable skills analysis) not fulfilling the condition of $n \ge30$, we use Mood’s median test, which has good efficiency for smaller sample sizes [26]. For all tests, results are considered statistically significant at $p < 0.05$.

Findings

Part A: Internal Effects

In this section, we analyse the prevalence and value of complementary and substitute skills within AI roles compared to non-AI roles. Complementary skills refer to those skills that are complementary to the creation and/or usage of artificial intelligence. These include skills that are foundational for creating or working with AI systems, such as digital literacy or technical proficiency, or skills that augment or are augmented by AI systems and tools, bettering their impact or enabling their more effective operation in combination with human workers, such as analytical thinking or ethics. Substitutable skills (also referred to as substitute skills), refer to tasks or abilities that are substitutable by AI – that is, they can be performed by AI systems and tools with no human intervention. These skills comprise performance of routine or rule-based tasks, such as basic data processing skills or text synthesis and grammar skills that can be standardised and programmed, and thus need not be performed by human workers. Prior work has also referred to such tasks and skills as “automatable” or “computerisable” (Acemoglu & Restrepo, 2018a; Autor, 2015; Frey & Osborne, 2017). The United States Bureau of Labour Statistics (BLS) has concluded that the adoption of AI will reduce labour demand in office and administrative, sales, design, and legal occupations due to the ability of AI to substitute for humans in these occupations, and projected that computer and mathematical occupations - where the bulk of AI roles lie - will experience the second fastest job growth of all occupations until 2033, significantly outstripping the growth of the average market (BLS, 2024c).

Against this backdrop, in what follows, we reveal a significant increase in demand for skills identified as complementary to AI technologies, alongside a rising premium for these skills within AI roles. Conversely, skills that are considered substitutes for AI have declined both in popularity and value within AI-related positions. Finally, we take a deeper look at complementary skills and show that for some skills, both the demand and the premium are more pronounced, that certain combinations of AI skills and complementary skills are more popular, and that there exists a compounding effect of complementary skills where combinations of specific complementary skills may be both more popular and profitable. We further illustrate this with specific AI roles, such as data scientists, to exemplify the compounding effect of complementary skills.

Prevalence of Complementary and Substitute Skills

To begin, we establish a baseline understanding of demand for Complementary and Substitutable skills in AI and non-AI roles, to contextualise interpretation of results. As shown in Figure I, demand for complementary skills is proportionally higher for AI roles compared to non-AI roles and demand for substitutable skills is proportionally lower for AI roles compared to non-AI roles, in line with our assumptions expressed in the prior section, that vacancies that ask for AI technologies will have proportionally lower requesting of the substitutable knowledge worker skills and proportionally higher requesting of the complementary knowledge worker skills compared to non-AI roles. Additionally, we note a widening gap between AI and non-AI in complementary skills demand and a narrowing gap in demand for substitutable jobs from 2020-2023.

We then introduce a regression model to quantify the likelihood of observing complementary and substitute skills in AI postings (tbl:II and tbl:III). In the full model, which controls for individual-level as well as job and structural factors (Model 3, pseudo $R^2$ of 0.43), “Is an AI role” has a coefficient of 0.195 ($p<0.01$). Thus, being an AI role is associated with approximately 22%⁵ higher likelihood of demand for complementary skills, when controlling for other factors. The LLR p-value confirms the model is statistically significant, as are all control variables ($p<0.01$). Models 1 and 2 display mixed sign coefficients of -0.467 and 0.482, however these models contain less controls, underscoring the importance of considering a wide range of explanatory factors to capture the complexity of salary determinants in the job market. For substitutable skills, a given job vacancy posting for an AI role has a coefficient of -0.427 ($p<0.01$) , thus being an AI role is associated with approximately 34.8% lower odds of demand for substitutable skills, when controlling for other factors. The LLR p-value confirms that the overall model is statistically significant.

Additionally, we note that the complementary skills intensity (share of skills in a job posting that are complementary) is on average around 25% higher for AI roles compared to non-AI roles (17.8% compared to 14.2%), and a given job posting being for a job that requires creation or use of AI is associated with around 1 percentage point (0.91) higher complementary skills intensity.

Value of Complementary and Substitute Skills

In what follows, we delve into how the wages of AI roles are influenced by the presence of complementary and substitute skills. Descriptive evidence (Figure II) suggests that complementary skills are associated with higher salaries and substitutable skills with lower salaries, across both AI and non-AI roles. To further investigate this observation, we introduce a regression model to quantify the impact of these skills on wages, considering them as individual dummies (Tables IV and V).

Table V: Even though complementary skills are in higher demand for AI jobs, their presence does not consistently translate into wage premiums, suggesting a complex relationship between skill composition and compensation.

Regression Results: Log Salary for AI jobs

Demand for Substitutable Skills exhibits a statistically significant and consistent association with lower salary levels for AI roles across all model specifications. In Model 3, which controls for individual-level as well as job and structural factors and has the highest adjusted $R^2$ of 0.32 (explaining 32% of the variability in the dependent variable log salary, a moderate level of explanatory power), substitutable skills demand has a coefficient of -0.085 (p<0.01). This suggests that AI roles demanding these types of skills are associated with an approximately 8.2%⁶ lower salary than roles without such skills, controlling for other factors. The p-value associated with the $F$-statistic ($F=156.3$, $p=0.0$) confirms that the overall model is statistically significant, and all control variables are considered statistically significant at $p<0.01$ except for the years 2019-2022.

Interestingly, demand for complementary skills also exhibits a statistically significant and consistent association with slightly lower salary levels for AI roles across all model specifications. I will further discuss the implications and potential explanations for this finding in the Discussion chapter. In Model 3 ($F=153.2$, $p=0.0$, $R^{2}=0.31$), complementary skills demand has a coefficient of -0.042 ($p<0.01$). While this negative association is much smaller than for the substitutable skills, this suggests that AI roles demanding complementary skills are associated with an approximately 4.1%⁷ lower salary than roles without such demand, controlling for other factors.

Granular Analysis

Motivated by understanding this surprising finding, we turn our focus to a granular analysis of complementary skills. First, we compare the odds of finding demand for specific complementary skills in AI roles compared to non-AI roles (Figure IIIa). We make three observations. First, the comparatively higher importance of Analytical Thinking, Technical Proficiency and Resilience and Agility in AI roles; second, the growing technicality of AI roles, as reflected in the increasing odds ratio for Technical Proficiency skills since 2020; and finally, the thus-far neglected but growing comparative importance of Ethics skills in AI. Next, we turn to examine whether comparative demand for complementary skills (compared to non-AI roles) differs between different kinds of AI roles, namely between AI user jobs (jobs with skills like e.g., ‘ChatGPT’, ‘Sentiment Analysis’, ‘Applications Of Artificial Intelligence’) and AI creator jobs (jobs with skills like e.g., ‘Large Language Modeling’, ‘Topic Modeling’, ‘Word2Vec Models’) (Figure IIIb). We observe, interestingly, that while comparative demand for complementary skills is higher for AI creator jobs at the beginning of our observation period, AI user jobs have caught up to the level of AI creator jobs, even exceeding these levels in 2020 and 2022.

Additionally, comparative demand for complementary skills in both AI user and creator jobs compared to non-AI roles has grown over time until a peak 2022 before dipping in 2023, and comparative demand levels for AI creators is generally less volatile compared to for AI users.

Motivated by these discovered nuances, we turn to further investigate specific skill complementarities and find that certain combinations of AI skills and complementary skills are more popular (Figure IV). In this analysis, we compare AI role category–complementary skills cluster combinations demand against a baseline of demand for that complementary skills cluster in non-AI roles⁸. These differences are presented in the heatmap below (Figure IV). The results suggest that some combinations are more interesting to labour markets than others. The heatmap shows consistently higher demand for most complementary skills in AI roles – analytical thinking, resilience and agility, technical proficiency, self-efficiency and working with others. The exceptions to this are ethics and digital literacy, perhaps reflecting the thus far neglected topic of ethics in AI and unspoken assumption of digital literacy in AI professions.

Finally, while there is mostly high consistency across AI role categories for each skill cluster, there is less consistency within AI roles categories, which suggests stronger and weaker skills complementarities for each kind of AI role. For example, for machines learning jobs, technical proficiency shows high complementarity while ethics does not. Two-proportion z-tests (or Fisher’s exact test, if the conditions for the two-proportion z-test are not fulfilled) confirm that most of the observed differences are statistically significant ($n=41$, of 56)⁹. Finally, although qualitative text analysis is not within the scope of this study, a pull of AI role postings that explicitly mention the word “complement(ing)” in the job vacancy description (body text) ($n=263$) provides a quick validation of the findings and sheds further light on the way employers communicate demands for complementary skills in AI roles. Employers describe the need for skills to work and communicate in cross-functional teams and navigate the translation process between technical solutions and business requirements, communicate demand for “growth mindset and continuous learning” or “agility” in a rapidly evolving tech context, specify the need for “foundational” digital literacy and technical proficiency skills, and express demand workers that can “use[s] intuition and experience to complement data” (as quoted by one job posting) in AI roles.

Similarly, the results for substitutable skills suggest that some combinations are less interesting to employers. This is especially visible for basic data skills, summary and reporting, customer service, and office and financial administration skills, which display significantly lower demand for substitutable skills in AI roles compared to non-AI roles. Patterns are mostly consistent across AI role types, suggesting similar demand effects for substitutable skills regardless of AI roles category – for example, there does not look to be significantly different results between AI creator and AI user jobs.

To test if the observed differences are statistically significant, we conduct a two-proportion $z$-test (or Fisher’s exact test, if the conditions for the two-proportion $z$-test are not fulfilled). Only a small number ($n=5$)¹⁰ of the observed 56 differences are not statistically significant. Notably, all the differences discussed above are statistically significant. We also test a prevailing view in the literature and popular media that information management and prediction and decision-making are skills substitutable by AI (e.g., [7]; [8]; [38]; [62]), and our results suggest that this is not the case. While these results cannot speak to causation, they show that for all AI role categories, these skills are demanded at proportionally higher levels compared to non-AI roles.

Finally, we turn to seek a more granular understanding of worker compensation in specific AI role category–substitutable skills cluster combinations. In this analysis, for each AI role category (e.g., machine learning creators), we compare median log salary for vacancy postings in that job category to median log salary to vacancy postings in that job category which also demand specific substitutable skills (e.g. machine learning creator jobs that demand basic data skills).¹¹ To ensure a meaningful comparison and meaningful median calculation, cells are marked n/a if they do not have a minimum sample size of $n=30$, a common threshold for statistical analysis [59]. Overall, the results suggest that AI roles demanding substitutable skills have lower median salary levels. This pattern is strongest for computer vision and image processing jobs. There are three exceptions, AI development and operations jobs that demand customer service skills (+1.7%), NLP (creator) jobs that demand customer service skills (+0.4%), and AI prediction and analytics jobs that require office and financial administration skills (0.0%). To determine which of these differences are statistically significant, we conduct a Mann-Whitney U test or Mood’s median test and, notably, we find the three exceptions to the hypothesis detailed above are not statistically significant.

For specific AI role category–complementary skills cluster combinations, the results suggest that select combinations of complementary skills clusters in AI role categories command a small salary premium. As above, to determine which of these observed differences are statistically significant, we conduct a Mann-Whitney U test. Notably, only two positive (prediction and analytics jobs requiring technical proficiency and machine learning user jobs requiring resilience and agility) salary differentials are statistically significant.

Second, we find that for some skills, both the demand and the premium are more pronounced, and that there exists a compounding effect of complementary skills—combinations of complementary skills may be both more popular and profitable (Figure V). Below we present a selection of the most significant combinations.

The case of Data Scientists

Finally, to shed light on a real-life occupation, we turn to a detailed analysis of Data Scientists, including coefficient plots to assess the compounding impact of complementary skills on wages. We find that for Data Scientists, we find higher odds of observing analytical thinking (89% higher), technical proficiency (41% higher), resilience and agility (36% higher), digital literacy (18% higher), self-efficiency (13% higher) and teamworking skills (7% higher), and lower odds of demand for ethics skills (44% lower), echoing our findings for AI jobs overall (Figure VIa). We further note that for working with others and digital literacy this effect is greater than for other AI jobs, perhaps reflecting the broader stakeholder interactions common to data scientist roles. Moreover, we also find that these resilience and agility skills and ethics skills are associated with 4% and 8% higher salary, respectively (Figure VIb). Notably, we observe that for resilience and agility skills, this salary premium is higher than for both other AI roles and non-AI roles, underscoring the high perceived value of this skill in the rapidly-evolving data science profession.

Part B: External Effects

In the second half of this paper, we turn to expanding our analysis to measure the external effects across the broader economy, showing that AI impacts complementary and substitute skills even outside AI roles. Here, we hypothesise that the increasing AI adoption (increasing prevalence of AI roles) within an occupation, industry or region affects the demand for complementary skills and decreases the need for substitute skills, even outside of AI roles. For example, as AI becomes more common in autonomous vehicle technology, there may be a growing demand for urban planners who understand how to design cities optimized for self-driving systems. Conversely, traditional roles like traffic signal maintenance workers may decline in demand as AI systems manage traffic flow dynamically without relying on physical signal infrastructure. In what follows, we find evidence to support our hypothesis at the occupational, industry, and regional level (tbl:VI) and translate these findings into tangible labour market impact using real BLS statistics for our observation period (BLS, 2024) (fig:VII). Finally, we replicate these findings for the United Kingdom and Australia (Figure VIII).

We find that on an industry level and over the time period of our data (2018-2023), growth in AI roles is associated with a positive effect on the number of roles demanding complementary skills in an industry. A 1% increase in AI role postings is associated with a 0.054% increase in complementary non-AI role postings, holding other variables constant ($p<0.05$). Thus, a doubling of AI role postings – as it has happened between 2018 and 2023 – is associated with a 5.4% increase in demand for complementary skills in external (non-AI) postings, at the industry level. Furthermore, growth in AI roles is associated with a negative effect on the number of roles demanding substitutable roles. A 1% increase in AI role postings is associated with a 0.019% decrease in substitutable non-AI role postings (or a 1.9% decrease for every doubling of AI role postings), holding other variables constant, though this result is not statistically significant ($p=0.51$).

Similarly, we find that on an occupation level and over time, growth in AI roles is associated with a positive effect on the number of roles demanding complementary skills and a negative effect on the number of roles requiring substitute skills in an occupation. A doubling of AI role postings is associated with a 2.8% increase in complementary non-AI role postings ($p<0.05$) and associated with a 2.4% decrease in substitutable non-AI role postings ($p=0.30$), holding other variables constant.

Finally, on a regional level and over time, we observe similar effects. A doubling of AI role postings is associated with a 3.0% increase in complementary non-AI role postings ($p<0.05$) and associated with a 5.77% decrease in substitutable non-AI role postings ($p<0.05$), holding other variables constant. Furthermore, we repeat these analyses for similar data from the United Kingdom ($n=10{,}003{,}732$) and Australia ($n=5{,}001{,}726$) and observe similar findings (Figure VIII). To translate this into tangible labour market impact, apply these findings to real BLS statistics on job vacancies for the period covered by our data set, 2018-2023 (BLS, 2024). With an estimated gain of 51,113 jobs and decline of 34,612 jobs per year of our observation period (2018-2023), we can thus link the rise of AI roles to a net job growth of 16,501 per year in our observation period for non-computing occupations (fig:VII) (see later in chapter for discussion of computer and mathematical occupations). Moreover, we observe three relevant findings.

First, the complementary effect is larger than the substitution effect. This is true for all industries, but especially Management, Business and Financial, Office and Administration, Arts and Entertainment, Architecture and Engineering and Sales occupational categories where for every lost substitutable role there are around 1.5 times the number of gained complementary roles. This is at odds with popular media discourse around risk of worker displacement and a literature similarly preoccupied with these risks (e.g., [6]; [48]), and has meaningful implications for individuals, companies and policymakers, which we address in the Discussion chapter.

Second, the industry-level and regional effects are more significant than occupation-level effects. This is intuitive, as it is more credible to assume that, to use the same example as above, as AI penetration increases in the automotive industry, also marketing, maintenance and other non-AI workers in the industry must better understand e.g., AI use cases (technical proficiency, analytical thinking) and its ethical implications (ethics). Similarly, at the regional level, an increase in AI role postings within Silicon Valley may lead to a rise in complementary non-AI role postings, such as roles in legal compliance (e.g., drafting AI-related regulations) or education (e.g., training programs for AI skills). It is less likely that as AI penetration increases in the HR Assistants occupation, that other HR Assistants in other companies would see increased demand for complementary skills. We discuss this further in the Discussion chapter.

Third, the effect size varies across industries. The strongest effects can be observed in Business and Financial Operations, followed by Management, Office and Administration, and Sales. Notably, we exclude the Computer and Mathematical occupational category from fig:VII for improved legibility and because of the difficulty of separating external and internal effects in this industry. For this industry, we observe an estimated increase of 192,364 non-AI complementary jobs associated with AI adoption – 44,776 jobs in occupation-level spillover effects, 99,476 jobs in industry-level spillover effects, and 48,112 in regional spillover effects. Conversely, we estimate a decrease of 111,025 non-AI substitutable jobs associated with AI adoption – 28,574 jobs in occupation-level spillover effects, 23,573 jobs in industry-level spillover effects, and 58,878 in regional spillover effects – the most net-positive external effect, with 1.73 complementary jobs gained for every substitutable job lost.

Overall, we link the rise of AI roles to a net job growth of 287,593 per year over our observation period (2018-2023), accounting for growth in non-AI complementary roles, decline in non-AI substitutable roles, and growth of AI roles. Across all occupations, we estimate an increase of 243,477 non-AI complementary roles and a decrease of 145,637 non-AI substitutable roles (see fig:VII and in-text discussion of external effects in Computer and Mathematical occupations). We provide evidence to link these changes to the growth of AI roles in the above section. In addition, we estimate¹² 189,753 new AI role vacancies per year over our observation period, bringing the total net labour market effect that can be linked to AI adoption to 287,593 jobs per year, or approximately 0.17% of the US labour market in 2023 [34].

Discussion

In this paper, we found empirical evidence for co-occurring substitution and complementation effects, reflected in both demand for and compensation of skills. Notably, we find evidence that these effects occur both internally (within jobs affected by AI) and externally (beyond jobs affected by AI, at occupation, industry and regional levels). These findings mark a contribution to the literature by examining the effects of AI on the granular level of specific skills, using real-world job postings data. This combination provides a relevant and more precise inquiry on the actual job market effects of AI compared the more general occupation or task-based research or to forecasting studies that rely on expert opinions or theoretical models.

Internal Effects

Demand

Our regression results suggest a positive relationship between AI roles and likelihood of demand for complementary skills and a negative relationship between AI roles and likelihood of demand for substitutable skills. This is significant, because it provides preliminary evidence that AI adoption is associated with both complementation of related knowledge worker skills, increasing demand for these skills, and substitution of other skills, reducing demand for these skills. While these findings cannot infer causality, they can nevertheless inform directions for further research, and underscore the strategic importance of workforce planning and skills development for policymakers, organisations and individuals. Furthermore, our findings presented evidence that vacancies that ask for AI technologies will have proportionally higher requesting of complementary knowledge worker skills compared to the non-AI data and lower requesting of substitutable skills. This is particularly noteworthy for complementary skills, which already have a high baseline of demand in non-AI roles (close to 70%). Together, the finding that AI creation/usage is associated with co-occurring complementation and substitution (as measured by likelihood of demanding skills and proportionally higher or lower demand for skills) reaffirms the importance of targeted policy interventions for e.g., facilitating structured reskilling initiatives or transitional programmes and potential adjustments in social security systems to address impacts of wage compression or job loss for displaced workers. In addition, they highlight a growing imperative for curricular reforms for embedding analytical and digital competencies and interpersonal skills in order to prepare future workers for AI-augmented roles, especially in sectors where effective human-AI collaboration can enhance productivity and innovation.

In addition, our findings provide evidence for three further contributions. First, our results also suggest that particularly powerful complementarities exist. To these authors’ knowledge, this is among the first studies to do so on this granular level. These include Technical Proficiency in technically intense jobs in machine learning creation, NLP creation, and computer vision and image processing (84-247% higher demand, $p<0.05$), or Digital Literacy in data-intensive AI development and operations (33% higher demand, $p<0.05$), or Resilience and Agility in the rapidly-evolving realm of machine learning user jobs (92% higher demand, $p<0.05$). Conversely, they also suggest that particularly strong substitution effects exist, such as Basic Data Skills in jobs using AI applications and tools (-86% lower, $p<0.05$), or Summary and Reporting in jobs using NLP, which includes generative tools such as ChatGPT (-78% lower, $p<0.05$). The identification of specific strong complementarities in specific technical skills across AI role functions highlights areas for strategic investment by firms, while specific strong substitution effects pinpoints automation susceptibility of skills in specific contexts.

Secondly, these findings also spoke to recent shifts in demand trends, in particular highlighting the growth of demand for Ethics skills across many AI role categories, which may reflect a broader societal and regulatory shift towards more responsible AI development and deployment.

Third, interestingly, we observed that while comparative demand for complementary skills has always been higher for AI creator jobs, AI user jobs have lagged behind at the beginning of our observation period, before catching up to the level of non-AI roles and finally by 2022 to the level of AI creator jobs. This finding signals a maturation of AI technologies, as they become more embedded in everyday workflows, the demand for complementary skills in AI user roles has now surpassed that of non-AI roles and matched the levels seen in AI creator roles. These shifts imply that workers in non-technical roles must increasingly develop complementary-to-AI skills, even when they are not directly involved in the creation of such technologies, while organisations and policymakers should focus on investments to democratise access to education and training initiatives to meet this expanding demand across the labour market.

Valuation of skills

Furthermore, assuming that complementation and substitution effects can be observed through how much skills are being valued as well as how much they are in demand, we turned to examine worker compensation. Overall, we found evidence to support the assumption that vacancies for jobs that create/use AI and demand substitutable skills will have lower salaries than AI role vacancies that do not but no evidence to support that complementary skills are associated with a wage premium. The regression results show that demand for complementary and substitutable skills are associated with small negative effects on salary (around 8% and 4% respectively). Such findings underscore the multifaceted nature of worker compensation determination in the context of technological change.

Yet, we also discovered granular nuances in these findings for substitutable skills, for example that the wage ‘penalty’ is highest for Basic Data Skills, Customer Service, and Office and Financial Administration Skills in computer vision and image processing AI roles (around 7-8%, $p<0.05$). This perhaps reflects a broader phenomenon of wage compression exacerbated by technological advancements [3]. As some employers increasingly rely on AI for the automation of routine tasks such as basic data processing, summary and reporting or office and financial administration, occupations reliant on substitutable skills may experience downward pressure on wages due to lower demand and reduced bargaining power.

Similarly, we found granular nuances in complementary skills, for example that a very small wage ‘premium’ exists in prediction and analytics jobs requiring Technical Proficiency and in machine learning user jobs requiring Resilience and Agility (around 1%, $p<0.01$). However, overall, our evidence suggests that no wage premium exists for complementary skills in AI roles. This is at odds with, for example, the claim that the most valuable tech workers are those with social skills and strong abilities to work well with others (provided higher salary can be assumed to reflect higher perceived value by employers) (e.g., [42]). The observed wage results suggest that while complementary skills are in high demand in AI roles, their direct impact on salary levels may be moderated by broader economic conditions and employer-specific factors. Future research could explore this nuanced relationship between skill complementarity and wage outcomes to inform strategies promoting fair compensation in AI-affected occupations.

External Effects

Finally, our findings reveal that AI adoption generates significant external effects, at the occupation, industry, and regional level. Specifically, growth in AI roles is associated with increased demand for non-AI complementary roles and a decline in demand for non-AI substitutable roles. This pattern is evident across all our three levels of analysis, though most pronounced at the regional level, where a doubling in AI roles is associated with a statistically significant increase of 3.0% in non-AI complementary roles and statistically significant decrease of 5.8% in non-AI substitutable roles, and the industry level, where a doubling in AI roles is associated with a statistically significant increase of 5.4% in non-AI complementary roles. Importantly, we find that the complementarity effect is larger than the substitution effect, and effects vary across industries. For example, Management and Business Operations and Computer and Mathematical Occupations exhibit the strongest complementary effects, with over 1.5 and 1.7 times the number complementary roles gained for every substitutable role lost, respectively.

To these authors’ knowledge, these results are among the first to quantifying the external effects associated with artificial intelligence adoption. Our findings indicate a subtle reshaping of labour markets in the triggering of a shift in demand toward complementary skills also externally to AI roles, challenging prevailing narratives of net displacement and emphasising the nuanced nature of AI's impact on the workforce.

The implications of these findings are multifaceted. For individuals, the results underscore the importance of acquiring skills complementary to AI creation or usage – such as technical proficiency and analytical thinking skills that enable understanding of AI use cases, and knowledge of ethical frameworks to evaluate their myriad implications – to capitalise on evolving labour market demands. For companies, the findings highlight the necessity of proactive workforce planning, including investment in employee reskilling. For policymakers, the observed regional spillover effects in particular suggest the need for targeted investments in education and workforce development in areas with high AI adoption, alongside safety nets and incentives for displaced workers to mitigate localised disparities.

Limitations

Finally, we draw attention to the limitations to this work. First, as highlighted by prior work (e.g., [2, 35]) online vacancy postings data tend to overrepresent professional and technical occupations relative to blue collar occupations. In addition, the data does not capture other channels of talent acquisition (e.g. headhunting, offline job vacancy advertisements), or satisfying skills demand needs (e.g., external contractors or in-house skills development for current employees) ([43]).

Second, the skills requirements communicated in the postings may not reflect skills demand realities beyond the recruitment stage. For example, an employee, once hired, may end up using a slightly different set of skills than those originally requested due to market conditions or changes in the organisation. However, we underscore that the focus of this research is to observe demand and offered compensation at the point of recruitment, but is useful to bear in mind when interpreting the implications of results. A robust literature on organisational behaviour exists for those interested in these specific dynamics (e.g., [69]).

Third, we assume that higher salaries reflect the perceived value of demanded skills in the marketplace. However, salaries communicated in job vacancy postings may not always reflect the true remuneration paid to hired individuals; although they do indicate employers’ willingness to compensate for particular skills at the recruitment stage ([43]), and it is this valuation of skills that is of interest in the scope of this research.

Fourth, a classic pitfall of working with big data is that even small effects can appear statistically significant. We conduct a repeated sampling robustness check by running 1000 demand¹³ regressions on 1000 random samples of less than 0.1% of our data ($n=10{,}000$). In this check, coefficients range from approximately -1.0 to 1.5 (mean = 0.217, which is very close to our reported regression result of 0.195) for complements and approximately -2.5 to 0.5 (mean = -0.492, which is very close to my result of -0.427) for substitutes, and p-values range from 0.0 to 1.0 (mostly evenly distributed for complements and many statistically significant at $p<0.05$ for substitutes) (See Appendix Figure 5 for full outputs). However, it is crucial to note that the variables of interest, particularly demand for AI roles, are very thinly scattered across the data – recall that the number of AI roles is small ($n=47{,}871$) compared to the sample size ($n=11{,}729{,}662$). With taking a smaller sample, even with 10,000 observations, only a handful of AI roles may be captured, and it is equally important to be aware of this issue. Consequently, while the range of $p$-values could suggest that statistically significant results may not be found when using a smaller sample, the scarcity of AI roles in the data mean that a smaller sample may not be able to capture the hypothesised effects, motivating a ‘big data’ approach. In combining regressions with other methods such as descriptive statistics and analyses on smaller sub-clusters of data, this thesis has sought to mitigate against this limitation by answering each research question using a variety of methods leveraging a variety of data quantities. In addition, we have sought to sanity-check the observed results and interpret implications cautiously, with a focus on real-world relevance.

Fifth, these findings do not allow causal claims. However, in identifying and describing associations and relationships between variables, we hope these findings can speak to the broader debate on the nature of ongoing changes in the labour market, helping generate hypotheses for future research and suggesting potential causal relationships that could be explored through experimental or longitudinal studies.

Finally, the dataset predominantly used for this study comprises job vacancy posting data from the United States. As labour market outcomes associated with technology vary across different countries [13] the degree to which these findings can be generalised to other contexts remains unknown, and further research is needed to understand regional variations.

Conclusion

This research sought to extend our understanding of the internal effects and external effects of AI adoption – complementation of knowledge worker skills (e.g., technology literacy, ethics, analytical skills), as measured by higher demand or higher valuation, as well as potential substitution of knowledge worker skills (e.g., basic data tasks, office and financial administration) as measured by lower demand or lower valuation. Following [2] and [39], we assumed that changes in demand for and valuation of skills can be observed from the “footprints” left by job vacancy postings as employers communicate their evolving skills demands and their willingness to pay for specific skills, and leveraged a dataset of around 12 million online job vacancy postings to investigate these questions.

In this work, we shed empirical light on the relationship between AI creation/usage and complementation and substitution of knowledge worker skills. We make five main contributions. First, we show evidence for a positive association between AI roles and likelihood of demand for complementary skills and a negative association between AI roles and likelihood of demand for substitutable skills. Second, observe proportionally higher requesting of complementary knowledge worker skills in vacancies that ask for AI technologies compared to the non-AI data, and proportionally lower requesting of substitutable knowledge worker skills in vacancies that ask for AI technologies compared to the non-AI data. Third, we put forward evidence to suggest particularly powerful complementarities exist between specific types of AI roles and specific clusters of complementary skills, and conversely, that particularly strong substitution effects exist between specific types of AI roles and specific clusters of substitutable skills. Furthermore, we found evidence suggesting recent shifts in demand trends, in particular the growth of demand for ethics skills across many AI role categories. Fourth, we found evidence in support of the assumption that vacancies for jobs that create or use AI and demand substitutable skills will have lower salaries than AI role vacancies that do not, and show that the wage ‘penalty’ is highest for basic data skills, customer service, and office and financial administration skills in computer vision and image processing AI roles. We found no evidence to support the assumption that complementary skills are associated with a wage premium. Finally, we found external complementation effects – on industry, occupation and regional levels, growth in AI roles is associated with a positive effect on the number of non-AI roles demanding complementary skills. Importantly, we find that the complementarity effect is larger than the substitution effect.

Directions for further research include extension to further geographies such as Europe and Asia to understand regional variations, complementing these findings with qualitative investigation such as interviewing recruitment managers on perceived demand for and valuation of skills, or testing for causal relationships through experimental or longitudinal studies. It is clear that in order to balance the benefits and concerns associated with “AI at work” [27], researchers, policymakers, employers and job-seekers alike require empirically-grounded understanding of these important effects.

Appendix

Appendix Tables

Occupational category composition – BGI data compared to BLS data

Leveraging the closest publicly available comparative data for job openings by occupational category: Employment by occupational category (2023).

Occupation Category (SOC 2){sup}13`	BGI Data	BLS Data
Architecture and Engineering Occupations	2.4%	1.6%
Arts, Design, Entertainment, Sports, and Media Occupations	2.3%	1.4%
Building and Grounds Cleaning and Maintenance Occupations	2.1%	3.0%
Business and Financial Operations Occupations	7.8%	6.5%
Community and Social Service Occupations	1.9%	1.7%
Computer and Mathematical Occupations	9.2%	3.3%
Construction and Extraction Occupations	1.7%	4.0%
Educational Instruction and Library Occupations	2.7%	7.2%
Farming, Fishing, and Forestry Occupations	0.1%	0.3%
Food Preparation and Serving Related Occupations	4.8%	8.6%
Healthcare Practitioners and Technical Occupations	12.1%	6.6%
Healthcare Support Occupations	3.2%	4.7%
Installation, Maintenance, and Repair Occupations	4.0%	3.8%
Legal Occupations	0.6%	0.8%
Life, Physical, and Social Science Occupations	1.3%	1.0%
Management Occupations	10.5%	6.8%
Office and Administrative Support Occupations	10.3%	12.1%
Personal Care and Service Occupations	1.4%	2.0%
Production Occupations	3.4%	5.4%
Protective Service Occupations	1.3%	2.4%
Sales and Related Occupations	10.0%	8.3%
Transportation and Material Moving Occupations	6.7%	8.6%

Appendix Table III. Additional literature on skills

Substitutable skills

Skill	Literature review (non-exhaustive)
Basic data skills	([48]; [82])
Summary and reporting	([?]; [?])
Language and text review	([?]; [?]; [84])
Customer service	([?]; [?]; [?]; [?])
Office and financial administration	([?]; [?]; [82])

Complementary skills

Skill	Literature review (non-exhaustive)
Analytical thinking	(Bansal et al., 2021; Dell’Acqua et al., 2023; Jarrahi, 2018; World Economic Forum, 2023)
Digital literacy	(Schwab, 2017; World Economic Forum, 2023)
Resilience	(Jarrahi, 2018; Kane, 2019; World Economic Forum, 2023)
Technical proficiency	(Felten et al., 2019; Hopgood, 2021; World Economic Forum, 2023)
Ethics	(Borenstein & Howard, 2021; Jobin et al., 2019; Müller, 2023; Whittlestone et al., 2019)
Working with others	(Bansal et al., 2021; Daugherty & Wilson, 2018; World Economic Forum, 2023; Zhang et al., 2023)
Self-efficiency	(Daugherty & Wilson, 2018; World Economic Forum, 2023)

Substitutable skills

Skill	Closest match skills from dataset
Information management	Information Gathering, Information Processing, Information Organization, Research, Information Synthesis
Basic data skills	Data Compilation, Record Keeping, Filing
Summary and reporting	Memos, Data Reporting, Report Writing
Prediction and decision-making	Forecasting, Decision Making
Language and text review	Multilingualism, Editing, Grammar, Basic Writing, Writing, Written English
Customer service	Customer Service
Office and financial administration	Clerical Works, Administrative Functions, Discount Calculation, Spreadsheets, Budgeting

Complementary skills

Skill	Closest match skills from dataset
Analytical thinking	Analytical Skills, Analytical Thinking, Complex Problem Solving, Critical Thinking, Logical Reasoning, Problem Solving
Digital literacy	Digital Literacy, Information Literacy, Computer Literacy, Basic Internet Skills, Computer Terminals, Desktop Computing, Computer Keyboards, Personal Computers
Resilience	Active Learning, Adaptability, Adaptive Leadership, Adaptive Reasoning, Calmness Under Pressure, Change Agility, Cognitive Flexibility, Dealing With Ambiguity, Dynamic Personality, Growth Mindedness, Intellectual Curiosity, Learning Agility, Lifelong Learning, Open Mindset, Proactivity, Resilience, Stress Management, Tenacity, Willingness To Learn
Technical proficiency	Technical Acumen, Technical Curiosity, Digitization, Information Technology
Ethics	Ethical Standards And Conduct, Business Ethics
Working with others	Collaboration, Collaborative Communications, Communication, Cooperation, Coordinating, Delegation Skills, Management, Social Collaboration, Social Intelligence, Social Perceptiveness, Social Skills, Socialization, Team Building, Team Effectiveness, Team Leadership, Team Management, Team Motivation, Team Oriented, Team Performance Management, Teamwork, Verbal Communication Skills, Virtual Collaboration
Self-efficiency	Driven Personality, Independent Thinking, Self Evaluation, Self-Awareness, Self-Confidence, Self-Control, Self-Discipline, Self-Motivation, Self-Sufficiency, Strong Work Ethic

AI role Type

AI role Category

AI Skills

AI Creators

Machine learning

AWS SageMaker Apache MXNet Azure Machine Learning Keras (Neural Network Library) MLflow TensorFlow Torch (Machine Learning) Reinforcement Learning Artificial Neural Networks Autoencoders Convolutional Neural Networks Deep Learning Deep Learning Methods Generative Adversarial Networks Generative Artificial Intelligence Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) Transformer (Machine Learning Model) Variational Autoencoders Quantization Ensemble Methods Gradient Boosting Random Forest Algorithm Machine Learning Machine Learning Algorithms Machine Learning Methods Xgboost Feature Engineering Feature Extraction

		Feature Learning Feature Selection
	Natural language processing (NLP)	BERT (NLP Model)
		Conversational AI Dialogflow (Google Service)
		GPT-3 (NLP Model)
		Hugging Face (NLP Framework)
		Hugging Face Transformers
		Large Language Modeling
		Natural Language Generation
		Natural Language Processing (NLP)
		Natural Language Understanding
		SpaCy (NLP Software)
		Text Classification
		Text Mining
		Topic Modeling
		Word2Vec Models
		Machine Translation
		Speech Processing
	AI Development and Operations	Artificial Intelligence Development
		Artificial Intelligence Systems
		Automated Machine Learning
		Explainable AI (XAI)
		Hyperparameter Optimization
		Machine Learning Model Monitoring And Evaluation
		Machine Learning Model Training
		ModelOps

		Dimensionality Reduction Supervised Learning Artificial Intelligence Risk Azure Cognitive Services
	Computer Vision and Image Processing	Machine Vision OpenCV Optical Character Recognition (OCR) Pose Estimation
	Prediction and Analytics	Advanced Analytics Predictive Modeling Uncertainty Quantification Apache Spark
AI Users	Natural language processing (NLP)	NLTK (NLP Analysis) ChatGPT Naive Bayes Classifier Question Answering Semantic Search Sentiment Analysis Text-To-Speech
AI Users	Machine learning	AI/ML Inference Anomaly Detection Collaborative Filtering Decision Tree Learning Document Classification K-Means Clustering Pattern Recognition Principal Component Analysis Recommender Systems Support Vector Machine

		Transfer Learning Unsupervised Learning Scikit-Learn (Python Package)
	AI Applications and Tools	AIOps (Artificial Intelligence For IT Operations) Amazon Lex Artificial Intelligence Chatbot Ethical AI Expert Systems Intelligent Virtual Assistant Language Model MLOps (Machine Learning Operations) PyTorch (Machine Learning Library) PyTorch Lightning Pyspark Speech Synthesis Voice Assistant Technology Applications Of Artificial Intelligence Predictive Analytics Prescriptive Analytics

Appendix Figures

Appendix Figure 8

Appendix Figure 9

There is a positive correlation between the share of AI roles in an occupation and the level of complementary skills demand in an occupation. In addition, growth in share of AI roles is positively correlated with growth in the level of complementary skills demand in an occupation.

Appendix fig:10

NB: “Regional” refers to a city-level analysis in the United States and Australia, but territorial units (NUTS1) for the United Kingdom.