Automation technologies can augment labour productivity but can also serve as a substitute for people by automating functions entirely. In the past, the geography of this impact has been stronger in semi-urban and rural areas. Generative AI, an emerging form of AI, may represent a new leap in technology and is impacting higher-skilled jobs located in urban areas. While the full extent of its impact is still uncertain, the effects on jobs or skills will likely be context- and place-specific. This chapter explores both the observed and anticipated impacts of AI technologies as they mature and achieve widespread adoption. It provides novel regional estimates to contrast previous forms of automation technologies with Generative AI and the geography of those impacts. Furthermore, it analyses which sectors and types of jobs might be most transformed, zooming in on specific occupations. Finally, it discusses possible actions going forward that could help regions seize new opportunities and manage the challenges AI poses.
Job Creation and Local Economic Development 2024
Abstract
Generative AI has the potential to affect a wide range of high-skilled workers in OECD regions, including many who have historically been less affected by technology.
AI has the potential to transform local labour markets by boosting productivity, creating or destroying jobs, and changing the very nature of some jobs, including job quality. This chapter examines exposure to automation technologies across regional labour markets in OECD regions, exploring the potential effects on job creation and productivity, as well as their distribution across different regions, industries, and workers.
Generative AI will likely have a much wider labour market impact than other recent digital technologies, affecting a broader group of people and places. Digital technologies that predate Generative AI had a narrower focus, being designed to excel in one or few specific tasks. Generative AI, on the other hand, excels at many tasks. Around a quarter of workers are exposed to Generative AI, meaning 20% (or more) of their job tasks could be done in half the time with the help of Generative AI, but this figure ranges from 13% to 48% across regions.
As Generative AI tools find greater adoption and integration with existing software, exposure to Generative AI could exceed 90% of workers in some regions. As new software is developed or integrated with Generative AI technologies, over 70% of OECD workers could be exposed, with almost 40% being highly exposed (meaning at least 50% of their job tasks could be done in half the time with the help of Generative AI). Metropolitan regions, such as Greater London (United Kingdom), Prague (Czechia) or Zurich (Switzerland), are expected to be significantly more exposed than less urban regions. Within countries, the most affected regions are around 60% more exposed than the least affected regions. Furthermore, in some countries such as Colombia, the top region is expected to be three times more exposed than the least affected region. These strong regional disparities can be explained in large part by differences in industrial composition, as regions with a focus on services, such as education, healthcare, Information and communication technology (ICT), and finance, are most exposed to Generative AI.
Regions previously considered to be at comparatively low risk of automation are the most exposed to Generative AI. While previous automation technologies, including other forms of AI, affected mostly non-metropolitan and manufacturing regions, Generative AI has the potential to alter a significantly higher share of jobs in metropolitan regions and cities. On average, 32% of workers in urban areas are exposed to Generative AI, compared to 21% in non-urban regions. Furthermore, this gap can exceed 17 percentage points in some countries such as Colombia, Greece, and Romania. A similar reversal becomes evident when analysing worker exposure by gender and education level. Exposure to Generative AI is greater for high-skilled workers and women, while previous technologies mainly affected low-skilled workers and men.
While the exact effects of Generative AI on the geography of job creation and displacement remain to be seen, little evidence exists on technology-led automation leading to mass job destruction. Instead, empirical results link automation technologies in regional labour markets with regional productivity growth. A small but significant number of regions across the OECD have experienced automation-led job displacement in the last decade. In these regions, job displacement was outpaced by the creation of new jobs. Nevertheless, there is no assurance that these new jobs went to the workers that lost their jobs and could instead have been taken up by, for example, new workers entering the labour force. There is no reason to believe that the full effects of automation technology adoption have been realised, which means regional labour markets may still see further impact.
AI technologies could offer OECD regions a strategic tool to address critical economic and labour market challenges, including labour shortages, stagnant labour productivity growth, or workforce inclusivity. AI technologies can be leveraged to supplement workers, helping to ease labour shortages and the effects of an ageing workforce. The same technologies that threaten some jobs can help regions access untapped talent in low-skilled workers or workers with disabilities for whom many jobs were previously out of reach. If used to augment or complement workers’ skills, AI could also help catalyse new labour productivity growth. However, training and upskilling will be required, with 61% of European workers believing that they will need new knowledge and skills to cope with the impact of AI tools on their work.
Although Generative AI holds potential to improve labour productivity and job satisfaction, it also poses risks for workers. On the one hand, Generative AI can allow workers to perform tasks quicker and with greater accuracy, automating routine tasks and freeing up time for more meaningful work. On the other hand, practices such as algorithmic management raise questions about AI’s impact on job quality, along with concerns about privacy and potential biases in AI systems. Collaboration with social partners and establishing clear, transparent guidelines for AI use will be important in protecting workers’ rights.
To harness the benefits of Generative AI and address its potential risks, regions should adopt data-driven, place-specific solutions that foster workforce adaptation. This includes identifying opportunities for AI-driven growth, updating skills provision systems, and developing comprehensive skills inventories to address mismatches and prepare workers for new demands. Supporting SMEs in AI adoption, fostering collaboration with local stakeholders, and providing tailored support for displaced workers are also key to sustainable and equitable regional development amid technological changes.
Technological progress has been an important driver of economic development. In many cases, such progress has had direct labour market consequences, with implications for firms, workers, and the skills required in many jobs. Generative Artificial Intelligence (Generative AI)1 is one of the latest waves of innovation in Artificial Intelligence (AI) and has captured the attention of people around the world. Progress in AI, with its easy access, practical use and showcased applications, has become highly tangible for a much larger proportion of the population. Yet, the current state of Generative AI might only be the start, with companies around the world working on further applications. While the adoption of Generative AI might provide tangible benefits for its users, the exact impact on workers, firms, and different communities are yet to fully materialise. One major concern is the potential for job displacement through task automation.
The discussion around the effects of automation encompasses both positive and negative aspects, including job losses, greater productivity, and overall economic growth. In broad terms, the concept of automation can be understood as the application of technology to achieve outcomes with minimal human input (Box 3.1). In assessing its impact, on the one hand, there is automation anxiety (Akst, 2013[1]) which is the real concern that technology will lead to the loss of jobs. On the other hand, automation can lead to productivity gains, thus possibly translating into higher incomes for workers.
In the past, new technologies have impacted regions and groups of people differently, and this uneven development is expected to continue. As new technologies spread throughout the economy, they often become concentrated in specific geographical areas, influenced by factors such as infrastructure, culture, natural resources, institutions, industrial composition, and workforce skill composition, among other regional characteristics.
One of the main motivations for adopting AI and automation technologies is to increase productivity while reducing labour costs. Firms using AI tend to be, on average, more productive than other firms (Calvino and Fontanelli, 2023[2]). For instance, studies show that robot adoption in Spanish firms can result in productivity gains of 20–25% within four years, a reduction of the labour cost share by 5–7 percentage points, and net job creation of 10% (Koch, Manuylov and Smolka, 2021[3]). One explanation for these findings is that automated firms become more productive and competitive, which allows them to lower product prices, gain market share, and ultimately increase labour demand (Banco de España, 2023[4]; Aghion et al., 2022[5])
However, productivity gains do not solely materialise through the adoption of new technology. Instead, these gains often come from a combination of factors, including how well the technology works alongside other digital technologies, a firm’s resources (including its workers’ skills), and overall economic policy. Firms with better access to key technical, managerial, and organisational skills have benefitted more from digital technologies than other firms. These varied impacts are relevant to other topics in this report, as technologies can address labour shortages, introduce new challenges related to digitalisation, drive the need for upskilling among workers, and influence overall employment levels, among others. (OECD, 2019[6]).
The actual impact of new technologies on regional labour markets and workers will likely be more nuanced than scenarios of mass job losses or significant productivity gains. Furthermore, labour market impacts may not be evenly distributed, may interact with each other, and potentially be felt disproportionately by specific groups of workers. For example, automation can lead to job polarisation and wage inequality (Acemoglu and Loebbing, 2022[7]), job losses concentrated within some sectors of the economy and particularly among workers with lower education (Georgieff and Milanez, 2021[8]), or reduce the quality of jobs even if the quantity is not affected (Autor, 2015[9]).
As technologies evolve, they often transform job roles and thus local labour markets, rendering some jobs unrecognisable from their previous iterations or giving rise to entirely new occupations. Bank tellers serve as a pertinent illustration of this phenomenon. As automated teller machines (ATMs) were introduced in the 1970s, their widespread adoption accelerated notably throughout the 1990s and 2000s. Nevertheless, human bank tellers as an occupation did not disappear (although their share in the US economy did decline). While the number of bank tellers per bank decreased, the number of banks increased in parallel due to lower operating costs (Bessen, 2015[10]). This allowed banks to expand their urban branches by 43% and focus on customer needs that machines could not address, particularly small business clients who may have previously been overlooked. Overall, the tasks carried out by bank tellers expanded to encompass soft skills more akin to those of salespersons.
Although the labour market impacts of technology and automation are a main topic of concern, knowledge regarding the effects of Generative AI remains limited. Given this recent wave of technology, research on the subject is sparse, with most existing studies predominantly focusing on a single country. Furthermore, empirical work has so far mainly neglected subnational differences, resulting in a gap in our understanding of variations across regional labour markets. This chapter presents new evidence on the geographic distribution of exposure to Generative AI, addressing a significant data gap and examining how the labour market impacts of this latest advancement differ from previous technologies.
The following sections explore and discuss the impact of modern digital technologies, at least some of which can be categorised as AI, on local labour markets and workers. These technologies can be grouped based on both the scope of tasks they perform and the ease of accessibility to the user. First, an overview of the different technologies involved in automation and AI is provided. Next, a set of empirical estimates on risk of automation and exposure to Generative AI in OECD regions is presented and discussed. The last section addresses policy-relevant issues aimed at helping regions harness the benefits of new technologies, mitigate associated risks, and leverage AI to both tackle labour market challenges and enhance the work of the public administration, including public employment services (PES).
AI developments over the last decades: Narrow-purpose technologies
AI encompasses various technologies – some of which have been around for decades – and has seen remarkable advancements in the 2010s. In fact, AI development dates back to the early days of computers, with foundational work such as the 1943 perceptron model (McCulloch and Pitts, 1943[11]) leading to today's neural network models a - centrepiece of modern AI technology. AI progress in the latter half of the 20th century and in the early 21st century has been marked by periods of advancement but also setbacks (Anyoha, 2017[12]). Over the last 15 years, this sector has boomed with deep learning, for example, being named one of 10 breakthrough technologies in 2013.2 These technical advancements have powered the gears behind major new technologies such as image recognition (with applications ranging from medical imaging to crop monitoring), text and speech recognition, or language translation among others.
Automation can be understood as the application of technology to achieve outcomes with minimal human input. Automation technologies encompass then a large set of technologies, including digital technologies – which include both conventional and AI powered software, which may or may not be Generative AI – and mechanical systems, among others (Figure 3.1).
The use of technology most commonly leads to the automation of specific work tasks and only rarely leads to the automation of entire jobs (Bonney et al., 2024[13]). A given job can be understood as a set of diverse tasks carried out by the worker. While technology adoption has led to the disappearance or creation of jobs, it typically replaces specific tasks within a job rather than eliminating the entire job. In other words, automation technologies can potentially displace tasks, not jobs. Nevertheless, technology-led task automation may still lead to job losses (typically productivity enhancing) as, with the use of technology, one worker can do the work that previously required several workers, which may prompt employers to scale down their labour force.
It is useful to classify automation-related technologies based on their format, scope of application, and accessibility. Figure 3.1 illustrates a non-exhaustive classification of technologies associated to task automation, along with examples. Mechanical technologies refer to physical technologies that provide engineering solutions for human tasks, with robotics and various manufacturing technologies being prominent examples. Digital technologies involve the use of computers and range from simple software, such as email clients and text editors, to conversational AI platforms such as ChatGPT. Although technology groups are visualised as separate, these are often used together as, for example, delivery robots might use image recognition software to navigate their environment. However, the technology groupings presented serve as an intuitive and broad illustrative classification, though they are not exhaustive.
Examples of digital and mechanical technologies
Note: Although categories are illustrated as separate, some technologies might fit into more than one group depending on their specific application.
Source: OECD elaboration.
AI technologies can in turn be classified based on the context of their development, their inputs and outputs, the specific underlying model or the economic context they are deployed in. The OECD Framework for the Classification of AI systems (OECD, 2022[14]) provides a structure to characterise the application of an AI system deployed in a specific project and context (Figure 3.2). The framework allows for a systematic characterisation of AI systems across various dimensions. In this context, it is both relevant and practical to classify Generative AI technologies as distinct from earlier AI technologies. This distinction arises, at least in part, from the varying degrees of generality among AI technologies.
Characteristics per classification dimension and key actor(s) involved
Note: Actors are illustrative, non-exhaustive and notably relevant to accountability.
Source: (OECD, 2022[14]).
The degree of generality of an AI system refers to its ability to perform several tasks including ones for which it was not initially trained. While there is no single indicator of generality, several criteria in this framework can indicate generality when combined (OECD, 2022[14]) such as the (1) scale of the AI system, (2) its model development/maintenance and (3) its ability to combine tasks and actions into multi-task, composite systems.
Generative AI significantly differs from previous AI systems in at least these three criteria, indicating a higher degree of generality. First, the training datasets behind Generative AI models are larger than most if not all previous AI models and are only getting larger (Epoch, 2024[15]). Second, these models can be universal, customisable or tailored depending on the needs and budget of its user, and in practice the most popular platforms have universal models which are free to use and widely available. Third, by themselves or when integrated with other technologies, Generative AI systems can combine many tasks, at least as many as previous technologies.
For the purpose of this chapter, it is useful then to separate Generative AI technologies from previous AI technologies. Following the OECD Framework for the Classification of AI systems, Generative AI can be considered as a separate group of AI systems that, given its more general purpose, is expected to have a different labour market impact relative to previous technologies. In addition, these Generative AI systems are more accessible than previous waves of AI. For practical purposes, this chapter classifies automation technologies that predate Generative AI as narrow-purpose technologies (left side of Figure 3.1), as under this framework, their development and use characterise them as less general purpose than Generative AI (bottom right of Figure 3.1).
Source: (OECD, 2022[14]).
Technology, either AI or non-AI, designed to solve specific problems or address specific work tasks can be grouped in the category of narrow-purpose technologies (Box 3.1). Such models and technologies have in common, at their core, their ability to perform strongly in one or few specific tasks. Therefore, their scope is limited, as one technology is unable to generalise its knowledge to different, unrelated tasks outside of its designed purpose. For example, an algorithm might recognise tumours in medical images better and faster than a qualified doctor, but it may not be able to recognise other symptoms, describe what a tumour is or suggest a treatment. Moreover, narrow-purpose technologies have not been accessible to a wide range of users, at least in part because they were not originally designed for a broad audience.
The impact of AI in the labour market is expected to shift as new frontier technologies are adopted. The labour market impacts of narrow-purpose technologies tend to align with the specific and limited scope of the technology, affecting mostly low-skilled workers within certain primary or resource-based industries. Nevertheless, recent advancements in AI have broadened the capabilities of digital technologies into a growing set of cognitive non-routine tasks which is not limited to the initial intended use of the technology. Therefore, the labour market impacts of these new technologies should be felt by a broader group of potentially different workers.
Recent waves: Towards a more general use AI
Recent developments have supported the broadening use of AI technologies, leading to the emergence of Generative Artificial Intelligence (Generative AI). During the last three to five years, the technology industry has seen a shift away from narrow-purpose technologies to more broadly applicable AI technologies (Filippucci et al., 2024[16]). Generative AI, one of the latest advancements, is designed to create human-like outputs – usually text, images, or video – based on existing data. In practice, it can assist with a wide range of applications, from content creation to problem solving. Furthermore, this technology has been integrated into platforms that are easily accessible to a wide, non-technical audience through intuitive, natural language interfaces, often in the form of chat-based systems.
The broader scope and easy accessibility of Generative AI technologies suggest they will have a more widespread impact on the labour market compared to previous waves of AI. Although Generative AI is trained to excel in a single task — content generation — it does so effectively across many diverse contexts, themes, and formats (including not only text but also speech, images, and video), giving it a more general purpose than previous AI technologies. In addition, its use can extend beyond the original intentions of its creators, significantly broadening the range of tasks at which it excels. Consequently, many industries and occupations may find utility in its capabilities.
Although early studies identified low-skilled workers as the most exposed to narrow-purpose technologies, recent research suggests frontier AI may increasingly impact high-skilled workers as well (Nedelkoska and Quintini, 2018[17]; Autor, Levy and Murnane, 2003[18]). As AI technologies mature, it becomes possible for them to automate increasingly cognitive and non-routine tasks, such as analysing text, drafting documents, or searching for information. Therefore, the impact of this new wave of technology is expected to be concentrated in jobs associated with knowledge work in high and upper-middle income countries (Gmyrek, Berg and Bescond, 2023[19]). At the occupation level, recent research indicates that AI technologies are directed at high-skilled tasks (Webb, 2019[20]), correlate with higher wages (Felten, Raj and Seamans, 2023[21]) and increasingly impact women and highly educated workers (Pizzinelli, 2023[22]).
Generative AI technologies could contribute to the displacement of some jobs while others could instead be augmented, but in most cases the overall impact remains unclear. Job augmentation refers to technology enhancing or supporting human workers, therefore complementing human work. High-skilled workers with higher incomes are on average more exposed to AI, but they also exhibit high potential for complementarity (Pizzinelli, 2023[22]). Occupations that require a high level of cognitive engagement and advanced skills are better positioned to benefit from increased productivity while minimising the risk of job losses. Recent analysis that classifies occupations based on the potential for augmentation or displacement finds that more occupations could be augmented rather than automated by Generative AI. However, a significant share of occupations3 remain where both the potential for displacement and augmentation exist (Gmyrek, Berg and Bescond, 2023[19]).4 While AI is unlikely to fully replace many jobs, highly exposed jobs share certain characteristics such as belonging to occupations that can be done remotely (Hering, 2023[23]).
Within economies and local labour markets, AI technologies may contribute to increased income inequality and job polarisation. The benefits of new technologies not only tend to accrue to high-skilled labour and owners of capital (in the form of higher capital incomes and returns), but may also lead to wage stagnation in some workers, therefore increasing inequality (Moll, Rachel and Restrepo, 2022[24]; Manning, 2024[25]). Advanced economies may also face increased job polarisation in the face of AI adoption – more so than emerging economies – as their employment structure is better positioned to benefit from growth opportunities but also makes them more vulnerable to job displacements (Pizzinelli, 2023[22]). On the other hand, research indicates that there is no relationship, or even a negative one, between AI and overall inequality, although there is an indication that higher occupational exposure to AI may be associated with lower wage inequality within occupations (Georgieff, 2024[26]; Webb, 2019[20]) possibly because AI technology may act as a skill leveller (Box 3.14). The following sections explore this idea further and substantiate it with quantitative estimates.
Even before the emergence of Generative AI, the impact of automation technologies differed across local labour markets. While the adoption of technology can bring tangible benefits, it can also lead to job losses. This impact depends on the nature of the tasks performed by workers, resulting in varying benefits and risks across different regions and economic groups, and can be quantified by examining the skills and abilities required for each occupation and the composition of local labour markets (Box 3.2). This measure of risk of automation serves as a useful metric to examine the effects of narrow-purpose technologies. As the underlying survey considers all available technology in late 2021, this measure considers all advanced automation and AI technologies up to that point (Lassébie and Quintini, 2022[27]).5 The metric is then used to measure the share of employment at high risk of automation, with Figure 3.3 illustrating the results for OECD regions.
While slightly more than a tenth (12%) of the workforce is at high risk of automation, those risks differ widely across regions. On average, regions in Latin America, Eastern Europe, Asia, and the United States tend to have a higher share of jobs at high risk of automation, especially compared to regions in Central, Southern, and Western Europe, Oceania, and Canada. The share of jobs at high risk of automation ranges from under 1% (Greater London, United Kingdom) to almost 29% (La Guajira, Colombia), highlighting the great diversity in automation risks across regions in the OECD (Figure 3.3).
The O*NET database
The O*NET programme is a US-based effort to, among other things, collect and classify information on occupations, skills, abilities, knowledge, and work tasks. It is made up of several relational datasets which are updated on a regular basis and describe over 1 000 occupationsa. Skill and ability requirements of occupations are measured in terms of importance and level. The former indicates whether the particular skill or ability is important to perform the job, while the latter indicates the level of mastery or proficiency in that skill or ability needed for the job. This information can be combined with other datasets and surveys to explore different dimensions of work as well as labour market impacts of other phenomena such as the green transition (OECD, 2023[28]; Vona, Marin and Consoli, 2019[29]; OECD, 2024[30]), or remote work (OECD, 2020[31]; Dingel and Neiman, 2020[32]), among others.
Automation of skills and abilities
To examine the relationship technologies have had with local labour markets, a measure of risk of automation is developed. Drawing from expert surveys and detailed information on skills and abilities (O*NET), several metrics are developed to measure the extent to which occupations are automatable. (Lassébie and Quintini, 2022[27]). These metrics explore the risk of automation of occupations given available technologies, where the available technologies include all technologies which existed at the end of 2021 when the surveys were conducted. One measure considers that occupations are at high risk of automation if over 25% of its skills and abilities are highly automatable with available technologies. This measure serves as both an expansion and an update of the measure of risk of automation presented in the 2018 Job Creation and Local Economic Development report (OECD, 2018[33]).
The group of experts was comprised mostly of AI experts who were asked to rate the degree of automatability of skills and abilities in general, but they were not asked to distinguish between the specific technologies that might be driving automation. Therefore, even if the most recent advances have occurred in the field of AI, their ratings reflect the capabilities of older and newer automation technologies.b Furthermore, experts recognised that it would be best to focus on the capabilities of current technologies, since predictions regarding the distant future made in the past proved to be far from reality.
Measuring the automation potential of narrow-purpose technologies
The resulting dataset considers all technologies available in December 2021 and is therefore a practical metric to examine the impact of narrow-purpose technologies on labour markets. As experts were instructed to consider all available technologies at the time, this measure serves as a baseline estimate for that specific point in time. For the purpose of this chapter, the share of occupations at high risk of automation serves as proxy for labour market exposure to narrow-purpose technologies.
Note: a See full taxonomy in https://www.onetcenter.org/taxonomy.html. b The responses were averaged without removing outliers. Those skills and abilities that had a mean value larger than 3.5 were considered to be highly automatable.
Share of employment at high risk of automation in OECD regions, latest available year
Note: Estimates for TL-2 regions where available except for Slovenia which is TL-3. Last available year is 2024 for Canada and Korea, 2023 for Australia, Colombia, Costa Rica, Mexico, New Zealand, the United Kingdom, and the United States, 2022 for all others.
Source: OECD calculations based on (Lassébie and Quintini, 2022[27]), labour force survey and employment by occupations tables. See Annex 3.A for more details.
The share of employment at high risk of automation also varies significantly within countries. On average, the share of workers at high risk of automation in the most affected regions is almost four times larger than in the least affected regions within the same country (Figure 3.3). The largest subnational differences can be found in Czechia, where the top region is 9 times more exposed than the bottom region. Furthermore, such regional differences are also considerable in Spain and the United States where the top region is between 7 and 8 times more exposed than the bottom region.6 Even though the very low risk of automation in the District of Columbia (DC) affects the degree of dispersion in the United States, regional differences remain large (2.3 times) if DC is excluded.
In most countries, capital regions have a significantly lower share of workers at high risk of automation. In 22 out of 28 countries with data for multiple regions, the capital region has the lowest share of jobs at high risk of automation in the country. Australia, Canada, Colombia, Mexico, Portugal, and Spain are notable exceptions. However, even in those countries, capital regions or regions with very large metropolitan areas have a relatively low exposure to automation (e.g., New South Wales in Australia, which contains the city of Sydney). A subsequent section discusses in more detail how the impact of technology is distributed among urban and rural regions.
Regional dispersion is primarily driven by the significant variation across industries, with the manufacturing sector being the most affected over the last decade. Over 33% of employment in the manufacturing sector is (and has been) at high risk of automation, which is 5 times more than mining and quarrying, the next most exposed industry, and around 9 times more than the following four most exposed industries (Figure 3.4). Furthermore, 8 out of the 18 industries analysed have under 1% of employment at high risk of automation and only three industries (manufacturing, wholesale and retail, trade, and construction) account for almost 90% of employment at high risk. This shows the high specificity of most of these technologies, affecting only a small portion of the economy.
Share of employment at high risk of automation by industry – EU, Iceland, Norway, and Switzerland, 2022
Note: In addition to Iceland, Norway, Switzerland, and the United Kingdom this data includes the following EU countries: Austria, Belgium, Croatia, Czechia, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Luxembourg, Netherlands, Poland, Portugal, Romania, Slovak Republic, Spain, and Sweden. Numbers represent total employment, as measured by LFS surveys.
Source: OECD calculations based on (Lassébie and Quintini, 2022[27]) and EU-LFS. See Annex 3.A for more details.
The impact of technological progress on automation: Job losses, productivity gains, or job creation
A common concern among policy makers is that automation technologies will lead to the displacement of workers and the destruction of jobs. However, recent evidence points to more nuanced implications. Based on economic theory, task automation could lead to both job destruction and job creation, while stimulating productivity growth. It could result in falling employment, as employees are replaced by frontier technologies and displaced from the labour market, or in an increase in employment as labour demand is boosted by more productive workers (Acemoglu and Restrepo, 2019[34]). In addition, completely new jobs might be created by new technologies, as, for example, new machines need a specialised workforce to fix and maintain them. Furthermore, productivity gains from these technologies can lead to income growth, which may stimulate expansion in unrelated sectors, such as the leisure industries.
Empirical evidence on the labour market effects of automation has been mixed and suggests that both negative and positive consequences are unequally distributed. Recent analysis shows that automation of job-tasks is concentrated within routine tasks mainly held by low-skilled workers (Acemoglu and Restrepo, 2021[35]; Schwabe and Castellacci, 2020[36]). Between 50% and 70% of all changes in the wage structure in the United States (between 1980 and 2016) appear to be directly linked to automation-related changes in tasks. For instance, young men with no high school degree have experienced a 15% decline in real wages during this period, largely due to their heavy concentration in various routine occupations in manufacturing, mining, retail, and wholesale industries (Acemoglu and Restrepo, 2021[35]). Evidence from European countries suggests that between 2011 and 2019, employment has increased, overall, in occupations more exposed to AI-enabled automation, which also tend to have younger and more skilled workers. However, the results vary largely across countries (Albanesi et al., 2023[37]).
In the United States, a slowdown of wage and employment growth has been linked to an acceleration in the adoption of automation technologies. While the creation of new jobs due to technological advancement has been significant, it has been outpaced by the substitution of workers by new technologies (Acemoglu and Restrepo, 2022[38]). However, empirical work on this topic is so far limited and existing results may not translate outside of the United States (Acemoglu, Manera and Restrepo, 2020[39]) as taxation in the United States is much lower for capital than labour (especially capital involved in automation, such as equipment and software).
So far, there is little evidence of a significant fall in overall employment in local labour markets with a higher share of jobs at high risk of automation. On average, regions with a higher risk of automation have neither recorded significant job losses nor experienced slower job creation than other regions (Figure 3.5).7 However, this regional analysis may overlook more subtle impacts, such as the reduction in work hours rather than the complete loss of jobs. For instance, exposure to AI has been linked to a decline in average working hours in occupations with low levels of computer usage (Georgieff and Hyee, 2021[40]). Moreover, significant employment changes may still be observed in the future as technologies may still be adopted in lagging regions or industries.
Annualised increase in employment and share of employment at high risk of automation (2011–2019) in selected OECD regions
Note: Selected countries based on data availability at the regional level. Base year is 2012 for the regions of Bremen (DEU), Mecklenburg-Vorpommern (DEU), Saarland (DEU) and Ireland, 2013 for the regions Warsaw (POL), Mazowiecki (POL), Hungary, Korea, Lithuania, and Mexico, 2014 for the regions Corsica (FRA), Guadeloupe (FRA), Martinique (FRA), French Guiana (FRA) and La Réunion (FRA) and 2011 for all others. Size of bubbles represent labour market size (employment).
Source: OECD calculations based on (Lassébie and Quintini, 2022[27]), (OECD, 2024[41]), labour force surveys, employment by occupations table. See Annex 3.A for more details.
Nevertheless, this does not necessarily imply that jobs were not destroyed. Instead, evidence suggests that in many regions job losses were offset by job creation in other sectors of the economy. In fact, in most OECD regions (70%) which exhibited a drop in employment at high risk of automation, more than enough jobs were created to make up for this (top-left quadrant in Figure 3.6). On the other hand, in 70% of the regions where employment increased between 2011 and 2019, part of this growth can be attributed to occupations considered at high risk (top-right quadrant of Figure 3.6), highlighting the possible role of technology in facilitating labour expansion. However, this expansion of labour into high-risk jobs may prove to be a double-edged sword, as a potential future wave of job displacement may impact an even larger share of workers in these regions.
Regions by change in total employment (%) and change in high-risk employment (%), 2011-19
Note: Selected regions based on data reliability. 2011 or nearest available year. Regions marked with employment growth had growth in overall employment. Regions marked with employment had a decrease in overall employment. Regions marked with displacement had a reduction of total jobs at high risk of automation. Regions marked with creation had an increase in of total jobs at high risk of automation.
Source: OECD calculations based on (Lassébie and Quintini, 2022[27]), (OECD, 2024[41]) ,labour force surveys, employment by occupations table. See Annex 3.A for more details.
Overall job destruction across OECD regions can be attributed to automation in only a handful of cases. Figure 3.7 illustrates how the percentage change in employment is distributed among high-risk and non-high-risk occupations for regions that had overall negative employment growth. In only 11 out of the 42 (26%) regions that experienced a fall in overall employment between 2011 and 2019 was the decrease higher in occupations at high risk of automation. Furthermore, only six (14%) recorded employment losses exclusively for occupations at high risk of automation (bottom regions in Figure 3.7). For the latter group of regions, the fall in employment could be attributed, at least in part, to the job destruction of high-risk employment. In other regions, various economic factors may have been more influential than automation; however, automation-related job displacement could still occur in occupations deemed lower risk, albeit to a lesser extent.
Share of job change by occupation type in regions with overall job destruction, 2011-19
Note: Selected regions which had less employment in 2019 compared to the 2011 (or nearest year available). Value in both bars sum up to -100%, which is represent 100% of the negative change in employment. Background colours relate to quadrant colours in Figure 3.6.
Source: OECD calculations based on (Lassébie and Quintini, 2022[27]), (OECD, 2024[41]), labour force surveys, employment by occupations table. See Annex 3.A for more details.
Even though new job creation outweighed job losses in most regions, newly created jobs might not have benefitted those workers who lost their jobs. In fact, regions that experienced job losses due to automation may have seen these workers enter long term unemployment or leave the labour force, potentially through early retiring, with new jobs taken up by workers either transitioning from other jobs or recently entering the workforce. Labour market policies that target affected workers could facilitate their reintegration into the workplace by, for example, providing them with the skills needed in their local labour market and which enable them to take advantage of new AI technologies (Box 3.3).
Across the OECD, regions with a higher share of employment at risk of automation saw a small but significant increase in productivity. The annualised increase in productivity8 from 2011 to 2019 tended to be significantly higher in those regions where industrial activities were at higher risk of automation (Figure 3.8). This positive impact on productivity holds even when one considers heterogeneity across countries as, when including country fixed effects, an increase of 10% in the share of jobs at risk of automation is related to an increase of 1.1% in annual productivity, which amounts to a 5.6% increase over five years. The limited magnitude of this effect may be explained by the fact that some technologies were still under development during the 2010s and even when available may not have been instantly adopted9. Therefore, some regions may still stand to benefit from the untapped potential of technology.
Annualised increase in productivity and share of employment at high risk of automation (2011 – 2019) in selected OECD regions
Note: Selected countries based on data availability at the regional level. Base year is 2012 for the regions of Bremen (DEU), Mecklenburg-Vorpommern (DEU), and Saarland (DEU), 2013 for Hungary, Lithuania, and Mexico, 2014 for the regions of Corsica (FRA), Guadeloupe (FRA), Martinique (FRA), French Guiana (FRA) and La Réunion (FRA) and 2011 for all others. Size of bubbles represent labour market size (employment).
Source: OECD calculations based on (Lassébie and Quintini, 2022[27]), (OECD, 2024[41]), labour force surveys, and employment by occupation Tables. See Annex 3.A for more details.
Upskilling in Australia
The five-year National Skills Agreement in Australia seeks deliver high-quality and responsive Vocational Education and Training (VET) to support the development of a skilled workforce for current and future needs. The government will work with states and territories to deliver on shared national priorities which includes building Australia’s digital and technology capability. The government also collaborates with industry leaders to provide apprenticeships and traineeships that combine formal education with on-the-job training. This approach aims to help learners acquire practical skills that are immediately applicable in the workplace.
Romania’s initiatives to address digital skill gaps
Through the National Employment Strategy (NES) 2021-2027 and the Education and Employment Programme (EEP) 2021-2027, the Ministry of Labour and Social Solidarity (MLSS) in Romania implements measures to enhance digital skills in the labour force. These include revising vocational training regulations, providing digital skills training for job seekers, and establishing funding mechanisms for employee professional development. Digital skills training, for instance, will be provided starting in 2024 for unemployed individuals through county employment agencies.
Romania is involved in the European Commission’s pilot project for the European Certificate of Digital Competences (EDSC), alongside countries such as Finland, Spain, Austria, and France. The EDSC aims to promote digital skills, increase recognition of these skills, and support citizens in understanding and improving their digital competence levels. This initiative is part of the European effort to address the digital skills gap and achieve the goal of 80% of adults having basic digital skills by 2030.
South Korea’s AI National Strategy
South Korea’s AI National Strategy includes significant investments in education and training to prepare the workforce for an AI-driven economy. The strategy supports the establishment of AI education centres at universities and research institutions, offering specialised programs in AI and machine learning. The South Korean government also provides subsidies for workers and students to enrol in AI courses and certifications. Additionally, the strategy includes initiatives to foster industry-academia collaboration, so that the training programmes align with the needs of the market.
Korea facilitates use of the acquired data by establishing an AI hub to provide companies and researchers with AI training data and cloud-based high-performance computing (an essential tool to process large amounts of data efficiently). The ecosystem will include big data platforms to produce and manage data, especially for sectors such as finance and healthcare. Korea also provides AI vouchers to SMEs and start-ups that need AI-powered products or services. Using the vouchers, the beneficiary companies can purchase necessary AI solutions from AI-solution suppliers.
Japan’s strategic support to the private sector for adopting AI
Japan, through the Ministry of Economy, Trade, and Industry (METI), launched the J-Startup programme in 2018 to foster innovation and help Japanese startups succeed. This public-private programme identifies and supports top startups and provides them with resources and opportunities to innovate in various sectors, including healthcare, with a strong focus on new technologies and AI. This initiative, beyond empowering the ecosystem in Japan, looks to foster better partnerships between the private and public sectors. J-Startup describes the support offered by the private sector to the government as working together to do experimental studies with robots, products, and infrastructure networks. The support by the government includes applying supportive frameworks for startups such as subsidies and simplifying procedures.
Regions in Japan have their own J-Startup programmes, aimed at supporting regional growth and development. For example, J-Startup Kansai identifies promising startup companies that will serve as a role model for the region and supports their growth within the region (J-Startup Kansai, 2024[42]). Similarly, Central Japan implemented J-Startup Central to give robust regional support for their development (Central Japan Startup Ecosystem, 2024[43]).
The rise of large language models (LLMs) might have an impact on a much larger share of workers than previous digital technologies. LLMs such as Generative Pre-trained Transformers (GPTs) have characteristics of a general-purpose technology (Eloundou et al., 2023[47]) and can generate high-quality text, images and other content based on large amounts of training data. While technology in the past focused on specific parts of the production of goods and services, Generative AI in the form of LLMs can intervene in manifold ways.
This section provides novel estimates on the exposure of different workers and local labour markets to Generative AI. It expands the work done by (Eloundou et al., 2023[47]), which examines how the tasks of individual jobs can be done significantly faster via the use of LLMs and associated user interfaces (Box 3.4). It constructs four related estimates that examine occupational exposure at two levels of intensity (exposed and highly exposed) and two points in time (now and in the near future).10 Table 3.1 provides example occupations and their respective exposure to Generative AI. Overall, the analysis covers regional labour markets in 36 OECD countries (further details on country coverage and data can be found in Annex 3.A).
Eloundou et al. (2023[47]) measure the potential implications of Generative AI, specifically frontier LLMs such as General Pre-trained Transformers (GPT’s), on the US labour market. They classify jobs according to their task content. This methodology is based on the idea that jobs can be decomposed into distinct tasks, which can then be analysed in detail (Box 3.2). Conclusions for jobs are then inferred from the tasks that make up these jobs.
The authors along with AI experts classify tasks based on how much faster they could be completed by a human using an LLM or an LLM-powered system. This optimisation is conditional on the fact that the quality of output is not compromised. If the time required for a human to perform a task can be reduced by at least 50% with the use of an LLM, then this task is classified as exposed. For example, online merchants commonly “deliver e-mail confirmation of completed transactions and shipment”, which could be optimised with the use of Generative AI models.
Although LLMs are advanced in their capabilities, their use in everyday work may remain limited due to poor software integration or restricted access to relevant data. For example, LLMs are commonly used to write or de-bug software code, but writing prompts and copying code can be done faster if the LLM is embedded directly in a developer’s software tool (such as their integrated development environment), as is the case with GitHub Copilot (Box 3.10). In addition, some Generative AI platforms have restrictions that could be relaxed if necessary. For example, prompts given to early LLM powered chatbots were limited to 2 000 words, but this limit has since expanded. In addition, some LLM platforms cannot access the internet to retrieve up-to-date facts, which is not necessarily a limitation. In fact, some LLM-powered systems could be more useful for certain occupations if they were trained on internal company data, thereby also enhancing safety by reducing exposure to misinformation from the internet. Some tasks could theoretically be optimised with the use of LLMs, but the technology is either unnecessarily restricted or has not yet been converted into tools easily adoptable by users. As a result, some tasks are not yet exposed but are likely to become exposed in the (near) future.
To address this, the authors create multiple measures of exposure for occupations. Two of them are the most relevant: exposure now and exposure now or in the near future. As implied, the second measure is an expanded version of the first. Exposure now is defined as the share of tasks within an occupation that can be completed in half the time by using LLMs in their current form, i.e., Chat-GPT 3.5 or similar. Exposure now or in the near future is defined as the share of tasks within an occupation that can be completed in half the time by using with LLMs in their current form (the same as the first scenario) plus those tasks where it is easy to imagine additional software that could be developed on top of the LLMs that would reduce the time taken to complete the task by half. It is important to note that this last definition does not presume major advances in technology, but it merely expects software tools to catch up to integrate frontier LLMs and specialise in certain applications.a Neither of these measures necessarily imply that highly exposed occupations will face displacement. It could also be the case that those occupations can be done more efficiently with the use of Generative AI tools.
This report expands on the work of Eloundou et al. (2023[47]) and applies a modified methodology for OECD regions. Occupations are considered to be exposed if these measures – the share of tasks that can be done with a set of technologies – exceed 20%. Furthermore, occupations are considered to be highly exposed if these measures exceed 50%. These thresholds were chosen to reflect two particular dimensions of labour market exposure. In short, a threshold of 20% measures those occupations that have a low but significant exposure while a threshold of 50% measures those occupations where most of the job can be optimised by the use of Generative AI. Table 3.1 summarises the four resulting measures and provides examples.
Note: a Annex 3.A contains more details on these measures.
Source: (Eloundou et al., 2023[47]).
Most occupations exhibit some degree of exposure to Generative AI, though the extent of this exposure varies significantly across different fields. Higher paying occupations tend to be more exposed to Generative AI, while occupations heavily reliant on science and critical thinking skills are less exposed on average. Similarly, jobs that require more education and/or training tend to, on average, be more exposed to Generative AI (Eloundou et al., 2023[47]).
A quarter of workers are currently exposed to Generative AI, but this share is expected to grow
Across the OECD, around 26% of workers are exposed to Generative AI, but only 1% are considered to be highly exposed. Nevertheless, as Generative AI technologies are integrated into the workplace, up to 70% of workers could be exposed to Generative AI in the near future, with 39% of these considered to be highly exposed. Table 3.1 provides examples of occupations within each category. The extent to which these estimates materialise in regions depends on the actual uptake by workers and firms, which in turn hinges on both investment and training.
Share of employment exposed to Gen-AI now, latest available year
Note: Estimates for TL-2 regions except for Slovenia which is TL-3. Last available year: 2024 for Canada and Korea, 2023 for Australia, Colombia, Costa Rica, Mexico, New Zealand, the United Kingdom, and the United States, 2022 for all others.
Source: OECD calculations based on (Eloundou et al., 2023[47]) , labour force survey and employment by occupations tables. See Annex 3.A for more details.
The exposure of jobs to Generative AI across OECD regions varies significantly. The share of workers highly exposed to Generative AI in the near future is expected to range from 16% in Guerrero (Mexico) to 77% in Greater London (United Kingdom) (Figure 3.10). The within-country dispersion for this measure averages around 14 percentage points, indicating that the top region in a country is, on average, 1.6 times more exposed to Generative AI compared to the bottom region. In Colombia, the country with the highest regional dispersion, the top region (Bogotá Capital District) is over 3 times as exposed as the bottom region (La Guajira). Capital regions tend to account for the high share in the within-country dispersion, pointing to an urban-rural divide in this dimension.
Share of employment highly exposed to Gen-AI now or in the near future, latest available year
Note: Estimates for TL-2 regions except for Slovenia which is TL-3. Last available year: 2024 for Canada and Korea, 2023 for Australia, Colombia, Costa Rica, Mexico, New Zealand, the United Kingdom, and the United States, 2022 for all others.
Source: OECD calculations based on (Eloundou et al., 2023[47]), labour force survey and employment by occupations tables. See Annex 3.A for more details.
Exposure to AI and automation technologies can be viewed through more than one lens. The results presented in this section examine labour market exposure to Generative AI when this is embedded in an accessible platform. At least three different datasets have been developed to examine this issue through other vantage points and present complementary results.
AI Occupational Exposure (AIOE)
This dataset individually links 10 AI application to 51 skills and abilities (O*NET). It does this by surveying online gig workers and combines their responses into an occupation level indicator which describes the extent to which a given occupation is exposed to each AI application, as well as AI in general. The main indicator is the Artificial Intelligence Occupational Exposure (AIOE) and has been used to examine the geographical and industrial distribution of AI exposure (Felten, Raj and Seamans, 2021[48]), the AI exposure of labour markets across developed and developing economies (Pizzinelli, 2023[22]) and a detailed analysis of labour market exposure in Canada (Mehdi and Morissette, 2024[49]), among others. National level estimates of the share of exposed workers ranges from around 25% (India) to 65% (United States), with 57% of employment considered exposed in Canada.
A global analysis of potential task exposure to Generative AI
This dataset evaluates ISCO occupations, including their description and task content, in regards to their potential exposure to Generative AI (Specifically GPTs). The study used ChatGPT-4 to score individual occupations and creates national and global estimates on the share of jobs exposed (Gmyrek, Berg and Bescond, 2023[19]). In addition, the study evaluates which occupations have augmentation potential (they are likely to be complemented by Generative AI) and which have automation potential (they are likely to be displaced by Generative AI). Results show that 15.3% of jobs across the globe are considered exposed, with 13% having augmentation potential and 2.3% having automation potential. Nevertheless, this figure can reach 18.5% in high income countries, with 13.4% and 5.1% of jobs having augmentation and automation potential respectively.
It may be too early to determine the full impact of Generative AI on local labour market composition. While certain labour markets are highly exposed to AI, this has not yet resulted in significant changes. It may take some time before we observe shifts in employment figures in response to the impact of Generative AI. An analysis of online job postings in the United States (Box 3.6) indicates no structural changes in hiring practices since Generative tools were launched. While these results may not fully represent the OECD, it is anticipated that the United States will act as an early indicator for the rest of the OECD, given its likelihood to be among the first to respond to the impact of Generative AI.
So far, labour demand has not reacted to exposure to Generative AI, as evidenced in the United States. The share of job postings of occupations expected to be exposed to Generative AI has not changed significantly in the last three years. Around 82% of US online jobs postings in 2023 are considered to be exposed to Generative AI either now or in the near future, of which 59% are highly exposed (). This share of employment was also around 82% in 20211, before Chat-GPT 3 was released to the public. Although there has been no observable change, it might still be too early as at least part of the impact of Generative AI may have not materialised yet. In other words, although the correlation is non-existent right now, stronger effects may be observed in the future.
Nevertheless, labour markets may change as they integrate Generative AI tools. These exposure estimates do not imply job displacement, but they should correlate with productivity as they measure the tasks that can be done faster with the help of Generative AI. Therefore, Generative AI may lead to both job destruction and creation. Nevertheless, employers may decide to postpone both hiring and layoffs until the practical uses of new technologies yield results.
Share of online job postings by exposure to Generative AI, United States
Source: (Eloundou et al., 2023[47]) and Lightcast data on online job postings.
1. This trend is described for the exposure now or in the near future, but it holds for exposure now as well.
Most sectors of the economy can benefit from Generative AI
Industrial composition is the main driver of differences in exposure to Generative AI across local labour markets. Across sectors in the EU, only 5% of workers in agriculture are considered exposed to Generative AI compared to 71% of workers in the information and communications industry (Figure 3.12). Among the latter, a small share of workers (5%) is considered to be highly exposed right now (i.e., half of their tasks could be significantly accelerated through the use of Generative AI), but this figure might reach almost 90% in the future. The share of highly exposed workers in the financial and insurance industry in the future could be even higher at almost 97%.
Almost half of all sectors could see the majority of their workers highly exposed to Generative AI. In eight out of the eighteen sectors analysed in the European Union, over 50% of employment could be highly exposed in the near future. In four industries, real estate activities, information and communication, professional and scientific activities, and financial and insurance services, the share of exposed workers could exceed 80%.
Labour market exposure to Gen-AI by industry for EU countries, 2022
Source: OECD calculations based on (Eloundou et al., 2023[47]), EU-LFS and employment by occupations tables. See Annex 3.A for more details.
Only a few sectors, such as construction, accommodation, and agriculture appear to not face significant changes due to Generative AI. In those three sectors, less than a quarter of workers could be highly exposed to AI in the future. The common factor across these sectors is the more limited use of Information technology (IT) than elsewhere. In fact, the agriculture sector is expected to have only 7% of its workers highly exposed to Generative AI.
Generative AI exposure is expected to intensify, with industries that currently have a high share of exposed workers being those with a high share of highly exposed workers in the near future as well. It is reasonable to conclude that occupations that make use of Generative AI now will continue to do so in the future. Furthermore, as this technology evolves it is expected that its future use cases will align with its current use cases, therefore deepening exposure levels within the same occupations. Industry level estimates (Figure 3.12) support this conclusion as the most and least exposed industries, in terms of exposure to Generative AI, are expected to remain relatively unchanged.
Measuring complementarity with AI technologies across occupations
It is unclear whether these measures of exposure to Generative AI correspond to the replacement or complementarity of workers’ tasks. Some occupations involve some aspects that are necessarily human, such as physical presence, responsibility, or face-to-face communication. This methodology (Pizzinelli, 2023[22]) leverages O*NET data on work context and job zones to propose a framework that conceives complementarity as driven by a set of factors – social, legal, technical – that are independent of exposure itself.
Workers significantly exposed to Generative AI are expected to be impacted differently depending on the extent to which this technology complements or substitutes their work. This methodology attempts to shed some light on this issue by grouping workers into three categories: (1) high exposure-high complementarity, (2) high exposure-low complementarity and (3) low exposure (Figure 3.13).
Exposure to Generative AI now or in the near future and potential complementarity to AI
Contrasting exposure to Generative AI and potential complementarity
The relative positioning of occupations in terms of their exposure to Generative AI and potential complementarity provides insights into the likelihood of job displacement and opportunities for productivity enhancement. Occupations on the right-hand side of Figure 3.13 are relatively more exposed while occupations on the top-half present relatively higher levels of complementarity. Occupations on the top right are better positioned to be complemented by Generative AI while occupations on the bottom-right are better posed to be displaced. Nevertheless, as this is a relative measure it is not perfect and can only shed light on the potential for productivity or displacement relative to other occupations. Therefore, it can still not be determined which occupations will be displaced or have their productivity enhanced.
On average, occupations that require more education tend to have higher levels of complementarity with AI technologies. On the other hand, the opposite is true for occupations that only require little or some preparationa. The level of preparation refers to a composite measure which combines the levels of education, experience, and training necessary to perform the occupation (O*NET[50]).
Exposure to Generative AI now or in the near future and potential complementarity to AI by job family
Note: Y-axis categories corresponds to job family as defined by O*NET (O*NET, 2024[51]).
In most job families, more than half of their relatively highly exposed occupations are complemented by AI technologies. Nevertheless, there are 5 job families that contain mostly occupations that are relatively highly exposed with low levels of complementarity, illustrated at the bottom of Figure 3.14. For example, over 70% of occupations in Management and Educational Instruction and Library have relatively high levels of exposure with low complementarity which indicates they are more susceptible to job displacement. At the opposite end, Healthcare Support; Farming, Fishing, and Forestry; and Building and Grounds Cleaning and Maintenance present low levels of exposure.
Note: aNote that the complementarity measure considers job zones as a measure of complementarity as occupations with longer periods of required professional development would have a greater ability to integrate AI knowledge into their training programs and thus equip future workers with complementary skills. Therefore, some of this dispersion across skills is a mechanical result given the way the complementary measure is constructed. Nevertheless, job zones make up only a small part of the complementarity measure with work contexts making up largest part.
Metropolitan workers experience higher exposure to Generative AI than non-metropolitan workers
The concentration of industries within or outside cities drives disparities in Generative AI exposure between urban and non-urban labour markets. Certain industries, such as financial services or technology development, often concentrate around metropolitan areas while non-metropolitan or rural areas tend to rely on industries with a different production structure, such as agriculture or manufacturing. Similarly, workers are also spatially concentrated, with highly skilled workers often being more present in clusters in or around a few metropolitan areas.
Workers in urban and metropolitan areas are significantly more exposed to Generative AI than workers in the rest of the country, by every measure (Figure 3.15). The exact gap varies by country but, on average, labour markets in urban areas are over twice as exposed than non-urban labour markets. This average is primarily driven by Colombia where urban regions are 2.6 times more exposed than non-urban regions. Nevertheless, if we Colombia is not included, urban regions are 74% times more exposed than their non-urban counterparts, on average.
Share of workers highly exposed to Gen-AI now or in the near future by area type – 2023
Note: The definition of urban and rural may differ across countries so estimates are not directly comparable. Country selection is based on data availability. The use of different category names is done to align with each country’s classification. In the US, Metropolitan and non-metropolitan areas are groups of US counties and are defined and named accordingly the US-BLS.
Source: OECD calculations based on (Eloundou et al., 2023[47]), metropolitan level Occupational Employment and Wage Statistics (OEWS) and labour force survey data (LFS). See Annex 3.A for more details.
In European Union countries, workers in cities are significantly more exposed than elsewhere. Examining the exposure of workers to Generative AI by the degree of urbanisation (DEGURBA11), shows that cities are significantly more exposed than rural areas (Figure 3.16). Across the European Union, over 36% of jobs in cities are exposed to Generative AI. In contrast, in rural areas, only 21% of jobs are considered exposed to Generative AI. In some countries, such as Poland, Hungary or Greece, the share of employment exposed to Generative AI is at least twice as high in cities compared to rural areas.
Ratio of share of employment exposed to Generative AI in EU countries - 2022
Note: The degree of urbanisation (DEGURBA) is a classification that indicates the character of an area. It classifies the territory of a country on an urban-rural continuum.
Source: OECD calculations based on (Eloundou et al., 2023[47]) and EU-LFS. See Annex 3.A for more details.
The gap between rural areas and cities in exposure to Generative AI is expected to vary significantly by country. In the near future, the rural-urban gap in exposure could vary from just under 8% in Belgium to close to 35% in Romania (Figure 3.17). Luxembourg is the only country where rural areas are more exposed than towns and/or semi-dense areas. In addition, Luxembourg is expected to have the most exposed urban population in the European Union with 84% of employment in cities expected to be exposed to Generative AI in the near future.
Aside from a few outliers, cities across EU countries are expected to have a similar average exposure to Generative AI. Setting aside Luxembourg and Spain, which have the most and least exposed cities respectively, the share of exposed workers in EU cities is expected to differ by up to 11 percentage points, ranging from 53.1% in Romania to 64.2% in Sweden. However, the difference between exposure (now or in the near future) in rural areas differs by 31 percentage points across countries in the European Union, with the Netherlands having the most exposed and Romania the least exposed rural areas (Figure 3.17).
Share of workers highly exposed to Gen-AI now or in the near future by the degree of urbanisation in EU countries,- 2022
Note: The degree of urbanisation (DEGURBA) is a classification that indicates the character of an area. It classifies the territory of a country on an urban-rural continuum.
Source: OECD calculations based on (Eloundou et al., 2023[47]) and EU-LFS. See Annex 3.A for more details.
Shifting landscapes: The impact of AI on people and places
The latest wave of AI technologies represents a departure from the development of narrow-purpose digital technologies towards more general-purposed solutions. This section explores how labour market impacts differ between narrow-purpose technologies and Generative AI technologies, examining in detail which economic groups are most exposed and how relative exposures are distributed across regions. Results indicate that the impact of these technologies differs according to a worker’s level of education and gender. Furthermore, a massive shift is observed in the distribution of this impact across OECD regions.
OECD regions previously only mildly at risk of automation are now significantly exposed to Generative AI and vice versa. There is a clear negative correlation between the expected share of highly exposed workers to Generative AI in the near future, and a region’s share of workers at high risk of automation (Figure 3.18). This trend is statistically significant despite the presence of a few regions that display low values in both estimates, such as Andalusia (Spain), Central Greece, and Sicily (Italy), among others.
Share of employment highly exposed to Gen-AI now or in the near future and at high risk of automation, latest available year
Note: Horizontal and vertical lines represent unweighted regional averages. Size of bubble represents labour market size (employment).
Source: OECD calculations based on (Eloundou et al., 2023[47]), (Lassébie and Quintini, 2022[27]), labour force survey and employment by occupations tables. See Annex 3.A for more details. Estimates for TL-2 regions except for Slovenia which is TL-3. Last available year is 2024 for Canada and Korea, 2023 for Australia, Colombia, Costa Rica, Mexico, New Zealand, the United Kingdom, and the United States, 2022 for all others.
Workers with a higher level of education, who were previously at lower risk of automation, are now significantly more exposed to Generative AI than their less educated peers. Generative AI is causing a large shift in the impact of technology on local labour markets, as previous waves of innovation – not only involving digital technologies but also other technologies – had a stronger impact on less educated workers who were more exposed (Figure 3.19). Less educated workers tended to be between 60% and 70% more at risk of automation than their highly educated counterparts. Workers with higher levels of education are now more than twice as exposed relative to less educated workers, and this gap is expected to be maintained as both groups of workers become more exposed to Generative AI over time. The expected gap in the exposure to Generative AI between workers with high and low levels of education can be attributed to the large overlap between Generative AI and the tasks carried out by highly educated workers. This technology is not only useful for a larger set of tasks but it also is increasingly helpful with both cognitive and non-routine tasks, which are significantly more common among highly skilled workers.
Average exposure to Gen-AI and risk of automation by level of education, last available year
Note: Level of education according to ISCED 2011 (UNESCO, 2012[52]). Low corresponds to first digits 0-2, medium corresponds to first digits 3-4, high corresponds to first digits 5-8. Figure represents the weighted average across all OECD countries for which there is data. The estimates are weighted averages based on employment by level of education.
Source: OECD calculations based on (Eloundou et al., 2023[47]), (Lassébie and Quintini, 2022[27]), labour force survey and employment by occupations tables. See Annex 3.A for more details.
The large differences in exposure and risk of automation across levels of education is consistent across all regions (Figure 3.20). High-skilled workers are expected to be more exposed to Generative AI than the rest of the workforce in all regions. In just over 10% of regions, mostly non-urban regions in Latin America and Romania, high-skilled workers are expected to be at least twice as exposed to Generative AI as the average non-high skilled workers. As shown in Figure 3.20, the relative gap is much lower regarding narrow-purpose technologies as, across regions, the risk of automation for high-skilled workers ranges from 50% to 90% of the risk faced by low- and medium-skilled workers.
Relative gap in exposure estimates between highly educated workers and low- and medium-educated workers across regions, last available year
Note: Figure represents the un-weighted regional average for all OECD for which there is reliable data. Relative gap represents the average exposure of highly educated workers relative to the average exposure of medium low educated workers. Level of education according to ISCED 2011 (UNESCO, 2012[52]). Low corresponds to first digits 0-2, medium corresponds to first digits 3-4, high corresponds to first digits 5-8.
Source: OECD calculations based on (Eloundou et al., 2023[47]), (Lassébie and Quintini, 2022[27]), labour force survey and employment by occupations tables. See Annex 3.A for more details.
Women are somewhat more exposed to Generative AI than their male counterparts, which is again a reversal of previous trends. On average, the share of tasks done by women that can be sped up by the use of LLMs is currently 4 percentage points higher than that of men, and this gap is expected to slightly increase to over 6 percentage points in the near future (Figure 3.21).
Average exposure to Gen-AI and risk of automation by sex, last available year
Note: Figure represent the weighted average across all OECD countries for which there is reliable data. The estimates are weighted averages based on employment by gender.
Source: OECD calculations based on (Eloundou et al., 2023[47]), (Lassébie and Quintini, 2022[27]), labour force survey and employment by occupations tables. See Annex 3.A for more details.
The disparity in exposure between men and women is generally consistent across regions except for a few regions where men are significantly more exposed to Generative AI than women (Figure 3.22). Men tend to take on jobs that are at higher risk of automation in all regions studied. On the other hand, women tend to perform jobs that are expected to be more exposed to Generative AI in most but not all regions. Notable exceptions to this are La Guajira (Colombia), Oslo and Viken (Norway), Zurich (Switzerland), Yorkshire and The Humber (United Kingdom), Greater London (United Kingdom) and South East England (United Kingdom). In addition to differences in occupational uptake, these regional disparities also reflect variations in sectoral composition across regions.
Relative gap in exposure estimates between men and women across regions, last available year
Note: Figure represents the un-weighted regional average for all OECD for which there is reliable data. Relative gap represents the average exposure of women relative to the average exposure of men.
Source: OECD calculations based on (Eloundou et al., 2023[47]), (Lassébie and Quintini, 2022[27]), labour force survey and employment by occupations tables. See Annex 3.A for more details.
Examining the gender composition of industries sheds some light on the determinants of gender differences in Generative AI exposure. This further highlights the ripple effects of varying industrial compositions on the exposure of different economic groups. The share of men and women in each industry (as opposed to the share of employment in an industry that are men or women) varies greatly across industries (Figure 3.23). For example, in the real estate sector, approximately 2% of men and 3% women are exposed. This similar share is understandable, as the industry is relatively small in terms of employment, and its gender composition is balanced. On the other hand, manufacturing and construction (less exposed) concentrate a significantly higher share of men while health (more exposed) concentrates a significantly higher share of women.
Education, human health and social work are the primary industries driving high exposure to Generative AI for women. These industries account for over 30% of female employment but only for around 9% of male employment. In addition, these industries are expected to have a relatively high level of exposure to Generative AI. Education, which concentrates close to 12% of female employment, is expected to have an exposure to Generative AI that is 4 percentage points larger than average (44.4%), while human health and social work is close to the mean. Nevertheless, the occupations that make up these industries are quite broad as, for example, human health and social work includes both office jobs (often clerical) and physical jobs such as doctors or physicians. This simple analysis does not reflect potential gender differences within these industries that might further explain the gender gap in Generative AI exposure.
Conversely, industries with mainly male workers are less exposed to Generative AI. Construction and manufacturing, which together account for nearly 32% of male employment, have a below-average exposure to Generative AI, with construction approximately 18 percentage points and manufacturing 9 percentage points below the average. Furthermore, the gender gap in these industries is substantial as they account for only 12% of female employment. Other industries with a higher exposure to Generative AI employ a significant share of both men and women, but the distribution of employment between genders in these industries is more balanced, therefore contributing less to the gender gap in Generative AI exposure (Figure 3.23).
Share of employed men and women in each industry for EU countries, 2022
Notes: Highlighted sectors are the industries where the employment gap between men and women is most relevant, this is further discussed in the text.
Source: OECD calculations based on (Eloundou et al., 2023[47]), labour force survey and employment by occupations tables. See Annex 3.A for more details.
Zooming in on specific occupations
The latest wave of AI will not affect all groups of occupations to the same extent. While most occupations are, or will be, affected in some way or another by the latest wave of AI, some occupations may face more considerable change. The rest of this section zooms in on possible implications for jobs and workers in cultural and creative occupations (OECD, 2022[53]), healthcare12, and software-related occupations13.
At least some cultural and creative occupations are extensively using Generative AI, which has already sparked tensions in the industry. Generative AI tools can help creative workers come up with ideas, write scripts and text, and generate audiovisual content, among others. This has raised concerns about copyright, as the way Generative AI models handle copyrighted content remains unclear, and existing legal frameworks often fall short in addressing these issues. Workers in the industry have expressed concerns over the use of this technology, with some escalating their grievances to the point of strike action (Box 3.8).
Cultural and creative sectors (CCS) is a term used to describe a range of activity which has its basis in creativity and can typically be exploited through intellectual property rights. Broadly, CCS includes the following sectors: advertising, architecture, book, newspaper and magazine publishing, dance, design, fashion, film and television, libraries and archives, museums, art galleries and heritage sites, music, radio, theatre, video games, and visual arts. Each of these subsectors require different skills and have different dynamics and business models. Many jobs in CCS are non-cultural and creative jobs (e.g., an accountant working for a film company) and therefore these sectors are exposed to AI risks associated with many occupations beyond just cultural and creative ones. There are several examples of the impact of AI in these sectors.
Film and television
AI has had a significant impact on the film and television industry in multiple areas. Machine learning has been used in the visual effects industry (VFX) for over a decade in tasks such as Rotoscoping (tracing the outline of an element of an image to move or replace it), motion tracking for computer generated images (CGI), and picture or colour adjustments. Recently, AI use in VFX includes extensive moving picture generation and digital asset manipulation, as well as helping to generate code for some of the underlying programming of VFX engines.
Generative AI also has applications in idea generation, script writing and editing. While there is yet to be a blockbuster hit with a screenplay written by AI, Generative AI is being used to help generate ideas, edit, and even provide first drafts of scripts. Concerns over the implication of Generative AI for the writing industry in the United States contributed to the Writers Guild of America strike action in 2023. This large-scale industrial dispute was primarily due to differences in contract terms for writers working on live broadcast vs streaming content. However, the issue of Generative AI was also raised as a significant concern during the action. Resolution of the strike included agreements on new regulations for the use of Generative AI in script writing.
Music
AI has been used in music production for a number of years, but advances in generative AI models could lead to greater use of AI-generated vocals, instrumentals, and lyrics. Free AI music generators are already on the market, allowing people to create new musical pieces with ease. Similar to concerns with the use of LLMs in the publishing industry, this type of Generative AI raises significant issues in relation to copyright and the ownership of AI-generated music. Machine learning models are also playing an increasing role in shaping musical tastes, with major streaming platforms such as Spotify using machine learning algorithms to recommend music to listeners. In addition, AI is being used in less obvious ways in the music industry by, for example, helping rights management companies to identify use of copyrighted material in film, television, and the internet to establish royalty payments and protect copyrights.
Source: (OECD, 2023[54]).
Similarly, software related occupations are also extensively using Generative AI and qualitative evidence suggest they have benefited from increased productivity and job satisfaction. These occupations, which include programmers, database architects and software developers, for example, already have multiple tools at their disposal which help them write and correct code, organise and clean data, and produce digital designs, among others (Box 3.10).
The health sector faces significant labour shortages in certain regions and occupations (Box 3.9). Although this sector has a lower labour market exposure to Generative AI (Figure 3.24) than average, this technology could still help address labour shortages as there is a relevant share of occupations that are highly exposed and therefore offer scope for Generative AI adoption (Figure 3.26). Health professionals in charge of records, such as medical record specialist, are the most exposed within this group. In regard to doctors, physicians, dietitians, and general physicians count among the most exposed. On the opposite end of exposure are medical professionals whose jobs are not limited to diagnosis and who conduct physical procedures, such as surgical assistants, dental hygienists, and paramedics.
In addition, applications of AI in the health sector are diverse, including development of pharmaceuticals, diagnostics, and behavioural interventions through chatbots. For example, Japan has been at the forefront of AI innovation and strives to actively integrate AI into its healthcare system to enhance efficiency and address challenges, such as labour shortages (Cabinet Secretariat, 2024[55]; Ministry of Health, 2023[56]). The Japan Agency for Medical Research and Development (AMED) is working to support the development of medical devices that utilize AI (AMED, 2024[57]). These efforts aim to improve patient outcomes, streamline hospital operations, and mitigate the effects of an ageing population and workforce shortages (Box 3.9).
Average exposure to Gen-AI now or in the near future by occupation type
Note: Figure represents the weighted average across all OECD for which there is data, see Annex 3.A for more details. Top and bottom lines indicate the 95% confidence interval. Source: (Eloundou et al., 2023[47])
Nursing care industry in Japan
Japan’s nursing care industry is facing a serious labour shortage. The number of people requiring nursing care is increasing due to the declining birthrate and ageing population. According to estimates done by the Ministry of Health, Labor and Welfare, 570 000 new nursing care workers will be needed by 2040. Under these circumstances, providing high-quality nursing care services efficiently to those requiring care by actively utilizing digital technology has become an important issue in Japan.
AI has attracted a great deal of attention to solve these issues in the nursing care and welfare industry. There are solutions already available, including DRIVEBOSS, which helps to plan transportation routes and schedules to take facility users to and from nursing care facilities. AI Sakura-san, a voice conversation customer service system, is also attracting attention as a solution to labour shortages. AI Sakura-san takes over the work of in-house help desks, call centres, and inbound customer service. Given the prevalence of night shifts in the nursing care and welfare industry, there are high hopes for the system to be able to communicate with the elderly on behalf of staff.
AI innovation for the health sector in Japan
Japan's J-Startup program (Box 3.3) links the public and private sector to foster AI in various industries. Startups are selected among a pool of over 10 000 projects, with examples in the healthcare sector including:
Aillis: Uses AI for real-time diagnostic assistance and aims to alleviate the shortage of specialists by providing an emergency medical coordination system to share medical information remotely. This includes remote support through live video streaming, surgery support, and linking doctors to depopulated and remote areas to provide medical support.
Ubie: Optimises the patient experience by streamlining in-hospital operations through AI online interviews and matches patients with the relevant medical institution. This shortens not only the process of patient uptake but also their stay in the hospital. It is currently used by over 1700 hospitals and clinics.
AIM: Specialises in AI-driven endoscopic diagnostic support for GI cancers, improving early detection and treatment planning. There is a shortage of endoscopists in Japan and the world, and 20% of cancers are overlooked due to limitations of human observation.
AI to address shortages of medical professionals in the United Kingdom
The UK is experiencing a significant shortage of medical professionals and is turning to AI to help address this challenge. By 2036, England is expected to face a shortfall of 260 000 to 360 000 doctors, nurses, and other health professionals, despite increasing recruitment of foreign-trained staff. Administrative demands and documentation requirements are further compounding this problem, decreasing the time available for patient care, and leading to greater workloads and burnout.
The National Health Service (NHS) in the UK has convened an expert group to explore ways in which AI can support healthcare by automating routine tasks and augmenting the abilities of clinical professionals. Technologies under consideration include speech recognition and natural language processing (NLP) for clinical documentation, which could help free up time for medical professionals to focus on patient care, as well as automated image interpretation, robotics for interventions and rehabilitation, and predictive analytics using AI.
Planning and commissioning of local services will differ across regions. This approach aims to enable each region to tailor technologies to their unique needs and challenges, addressing regional priorities, and supporting the best use of available workforce and talent.
COVID-19 vaccines in El Chaco, Argentina
Argentina took a different approach on AI and healthcare, aiming to use behavioural nudges through a Whatsapp chatbot to increase vaccination rates. In 2022, in a randomised controlled trial (RCT), the country launched a WhatsApp chatbot, integrated with a booking system, aimed at increasing vaccination rates in the province with the lowest COVID-19 vaccination rates, El Chaco. The Ministry of Health of El Chaco collaborated with the private sector through a social impact company to increase vaccination rates by helping people to schedule their appointments and sending reminders.
The programme was highly effective as the treatment group’s vaccination rate was three times higher than that of the control group. Following the success of the RCT, the province of Tucuman launched a chatbot in 2024 with the aim to increase routine vaccinations such as influenza and COVID‑19 boosters.
The use of AI in healthcare presents significant ethical and privacy concerns. One of the main risks involves the handling of large volumes of personal health data, raising concerns about how this information is collected, stored, and used, and whether it could lead to breaches of patient confidentiality (Khan et al., 2023[66]). AI algorithms also risk perpetuating biases if they are trained on biased data, potentially resulting in unequal treatment for certain patient groups. Additionally, concerns about accountability arise when AI makes errors in diagnosis or treatment, as assigning responsibility can become challenging in such cases (Nicholson Price II, 2019[67]). Setting standards to protect patient data, maintaining confidentiality, and avoiding exacerbation of inequalities are key ethical considerations that need to be tackled as the use of AI solutions in the healthcare sector become more common. Changes in medical education may also be required to prepare healthcare professionals for evolving roles and responsibilities.
Generative AI could lead to a greater transformation of cultural, creative and software occupations than other types of occupations. On average, the potential of Generative AI to support or take over specific work tasks are expected to lead to a 2 percentage points higher share of exposure in cultural and creative occupations than for all occupations on average. In software occupations, the already existing integration of AI tools and the progress in development of new tools that facilitate programming by creating or optimising software code, such as GitHub Co-pilot (Box 3.10), results in even greater exposure to Generative AI, which could reach up to 87% of workers (42 percentage points above the average).
GitHub Copilot is an AI-powered coding assistant developed by GitHub in collaboration with OpenAI. It leverages machine learning models to assist developers by providing code suggestions and autocompletions directly within their integrated development environment. Copilot can help generate whole lines or blocks of code, making it easier and faster to write software, reducing repetitive tasks, and enhancing productivity.
The tool has a significant positive impact in the production of software developers. A study quantifying GitHub Copilot’s impact on developer productivity demonstrates that its users had a higher rate of completing tasks and did so 55% faster than developers who did not use the tool. Letting GitHub Copilot shoulder the routine and repetitive work of development reduced the cognitive load of developers and allowed for better productivity. This tool is being used by 90% of the Fortune 100 companies and by more than 100 million developers.
Between 60–75% developers reported feeling more fulfilled in their job, less frustrated when coding, and able to focus on more satisfying work when using the tool. It also helped with conserving mental energy, as developers reported that GitHub Copilot helped them stay in the flow (73%) and preserve mental effort during repetitive tasks (87%). That is associated with developer happiness, since previous research shows that certain types of work are mentally draining, and that context switches and interruptions can ruin a developer’s day.
Source: (Kalliamvakou, 2022[68]).
Some cultural and creative occupations are expected to be up to 90% exposed now or in the near future, unlike many others that are much less exposed. The more cultural occupations such as dancers, choreographers, glassmakers, or potters, mainly consist of tasks that are least likely to be significantly accelerated through the use of Generative AI (Figure 3.25). These occupations generally require more physical skill or presence, meaning they generally involve a greater share of low-exposure tasks. Conversely, occupations such as journalists, translators, or graphic designers, involve more computer-related tasks and do not have a substantial physical aspect, so these occupations are expected to be highly exposed. Moreover, other types of Generative AI technologies beyond LLMs (such as image and sound generation technologies) are already beginning to make an impact in cultural occupations that may have been thought to be low risk for example, generative image technology can create and manipulate so-called ‘digital doubles’, replicating an actor’s likeness for film or television. Similarly, in music, generative sound technology can automate the generation and production of music.
Average exposure now or in the near future, cultural and creative occupations
Note: Cultural and creative occupations are defined in ISCO at the 4-digit level according to the Eurostat definition used by the OECD.
Source: (Eloundou et al., 2023[47]), (OECD, 2022[53]).
Average exposure now or in the near future, health occupations
Note: Selected health occupations.
Source: (Eloundou et al., 2023[47]).
AI holds potential to improve productivity and support regional growth
AI, particularly new general-purpose models which can work autonomously, hold potential to address stagnating productivity gains and enhance innovation across OECD economies (see Chapter 1). This can be particularly relevant in tackling the demographic challenges of an ageing society. As the working-age population shrinks, AI can help maintain productivity by enabling fewer workers to achieve a higher output, thereby sustaining economic growth despite demographic decline. Micro-econometric studies find that the productivity gains from non-generative AI on firms are comparable to those from previous digital technologies (up to 10%). However, performance benefits from using Generative AI in tasks such as writing, programming, or handling customer services requests can range from 10% to 56% (Filippucci et al., 2024[16]). These gains are especially impactful for workers with less experience.
As AI adoption is still low, the long-term, and broad economic impact of this technology remains uncertain. In the US, less than 5% of firms report using Generative AI, making it difficult to observe macroeconomic gains at present (Filippucci et al., 2024[69]). Although the user-friendly nature of new AI tools may accelerate their adoption, realising their full potential still requires substantial complementary investments in data, skills, and software. Estimated aggregate gains suggest AI could contribute between 0.25 and 0.6 percentage points to annual total factor productivity (TFP) growth over the next ten years, boosting annual labour productivity by 0.4 to 0.9 percentage points. However, these estimates carry significant uncertainty (OECD, 2024[70]).
AI research and development, as well as adoption, is uneven across places, potentially deepening existing societal divides. AI activities could bring significant benefits to cities and regions, with the sector projected to grow from USD 184 million in 2024 to USD 826 billion by 2030 (Statista, 2024[71]). Some places, however, are better positioned to benefit from the opportunities in AI than others. In Europe, Germany, France, and Spain have traditionally been the strongest AI economic players (Righi et al., 2022[72]). In the US, AI firms are highly concentrated in the San Francisco Bay Area, with other important hubs across the country such as New York, Boston and Seattle, where large companies are already driving early adoption (Muro and Liu, 2021[73]). Factors such as regional skills, leading research institutions, and a strong innovation ecosystem, have proven to be some of the main aspects driving AI research, development, commercialisation, and adoption in these places.
There are concerns about market competition, as AI development and adoption is concentrated in a few dominant players. Adoption trends indicate a divide: larger firms and those in ICT sectors are integrating AI faster, while smaller and older firms are lagging behind (Calvino and Fontanelli, 2023[2]). This raises concerns about potential market distortions, particularly if early adopters gain substantial market power, making it challenging for other firms to compete. It remains uncertain whether the productivity gains achieved by early adopters will diffuse to other firms, potentially deepening existing economic divides (Filippucci et al., 2024[16]). Furthermore, the success of AI adoption often relies on complementary assets, including ICT skills and training, firm-level digital capabilities, robust digital infrastructure, access to quality data and computer power, as well as general workforce skills (Calvino and Fontanelli, 2023[2]). Not all regions or firms have equal access to these resources, which may exacerbate existing disparities and leave some regions struggling to keep pace. Additionally, the demand for AI talent is more widespread across sectors, which could provide evidence of the general-purpose nature of AI technologies.
The widespread use of AI also raises concerns about its impact on employment. While AI technologies have the potential to boost productivity, augment human abilities, and create new job opportunities, they also bring the risk of displacing workers. Generative AI, for instance, poses risks even for jobs traditionally viewed as secure from automation. However, recent evidence indicates that instead of reducing their reliance on human labour, companies across various sectors are responding to AI adoption by restructuring internally, reallocating workers to tasks where humans have a comparative advantage rather than displacing them (Filippucci et al., 2024[16]).
Effective public policy can maximise the benefits of AI technologies for workers and communities, while also addressing its potential negative effects. A key concern is promoting widespread AI adoption while preventing excessive market concentration. This means reducing barriers to adoption, supporting smaller firms, and ensuring fair competition without stifling innovation. Human complementarity with AI is also important, implying that AI should augment, rather than replace, human work. To achieve this, policies can encourage retraining and labour reallocation, helping workers transition into new roles and adapt to the use of AI technologies (Box 3.11). Preparing for rapid adaptation is equally important. AI-driven transformations are unpredictable, and responsive policy measures that adjust to these shifts can help technological progress benefit workers, communities, and industries, rather than contributing to inequality or causing disruption (OECD, 2024[70]). Furthermore, public policy can foster acceptance and trust by developing mechanisms to keep AI technologies in line with strict ethical standards, thereby improving transparency and accountability.
The Future Skills Centre is promoting various initiatives to enhance professional development and equip workers with the skills to work with AI. The Future Skills Centre is a national organisation dedicated to preparing workers across Canada for the changing demands in the workforce by investing in skills development initiatives. Three examples of projects related to AI issues are:
Accelerating the appropriate adoption of Artificial Intelligence in healthcare: This pan-Canadian project completed in 2024 aimed to overcome healthcare professionals’ reluctance to adopt AI by offering tailored education and training. The project developed interventions such as certificate programmes and mentorships to equip healthcare professionals with the skills and confidence to integrate AI into their practice. As a result, the project saw improved attitudes toward AI adoption, enhanced patient care, more efficient workflows, and increased confidence in AI systems.
From data to decision: AI training and professional certification: This pan-Canadian project addresses the growing demand for AI skills by offering a short-duration, online training programme developed by IVADO and Université de Montréal. The project aims to train 1 000 mid-career professionals and leaders across Canada on how to integrate AI into their organisations and make better use of the data they generate. The initiative includes self-assessment tools to identify skill gaps, a training path with up to seven courses, and professional certification for successful participants.
Facing the challenge of digital transformation in the insurance sector: Women at work: This project in the Quebec region addresses the impact of digital transformation on female workers in the insurance sector, where AI and automation threatens low-skill positions typically held by women. Led by a consortium at Laval University, the project assesses the skills development needs of female insurance workers and creates training pathways to reskill and support them in transitioning to future-facing roles, helping them remain competitive in the changing labour market.
Evidence suggests that policy attention could also benefit from considering the shifting geography and socio-economic effects of Generative AI compared to previous technological advancements. Unlike earlier waves of automation, AI’s impact increasingly falls on urban labour markets, which could drive growth in cities but also risks widening disparities between urban and rural regions (Figure 3.18). Moreover, Generative AI is likely to impact a different set of workers compared to previous technologies. This changing landscape suggests the need for adaptive policies for a different and emerging group of workers and regions.
Regions with a high proportion of jobs at risk of automation could benefit from monitoring job displacement trends to facilitate timely responses. Tracking employment in specific sectors and occupations can help regions identify local needs and react with appropriate place-based policies. This approach allows for the timely development of reskilling programmes, support for affected workers, and strategies to foster new employment opportunities. A proactive policy response can help mitigate potential negative effects, while enabling regions to harness the benefits of new technologies.
Assessing the existing skills base in regions with high labour market exposure to Generative AI could help develop targeted programmes to strengthen AI-related competencies. Conducting a comprehensive inventory of available skills could help policy makers and businesses identify gaps and opportunities, enabling them to better prepare the workforce to adapt to the evolving skill demands. By implementing tailored training and upskilling initiatives, these regions can mitigate potential disruptions, but also capitalise on the transformative potential of Generative AI. These efforts can include training programmes, grants, tax incentives and establishing innovation networks, among others.
By fostering AI local capabilities, regions can modernise traditional industries, attract investments, and leverage emerging technologies for sustainable growth. This is not only relevant for regions that are heavily impacted by automation, but also for those dealing with demographic challenges, low job creation, outmigration, and outdated industrial bases. For these regions, adopting AI strategies presents a dual opportunity: revitalising their economies while mitigating the potential adverse effects of automation (Box 3.12). Effective AI integration, paired with policies to support skills development and job transitions, can lead to more resilient regional economies and balanced growth across various sectors.
La Rioja, Spain
La Rioja’s industrial composition makes it particularly vulnerable to automation. La Rioja is a small region in Northern Spain with a strong industrial and agriculture sector. This industrial composition is particularly well suited for jobs at high risk of automation, as the region’s industrial sector is mostly manufacturing, and in fact the region has the highest share of jobs at risk of automation within Spain (Figure 3.3). The regions’ decline in employment in the last 10 years can in large part be attributed to a drop in employment in these jobs (Figure 3.7).
The region is advancing initiatives to promote research and business development focused on the application of AI and big data. This includes initiatives like the Parque Científico Tecnológico de La Rioja, an initiative for which the first stone has already been set in the form of the Tech FabLab which aims to promote technological business initiatives based on AI and other disruptive technologies (Gobierno de La Rioja, 2024[77]). This aligns the regions path with the stated national goal of having at least 25% of companies use AI and big data in five years.
Piedmont, Italy
Piedmont is one of the most industrialised regions in the OECD, with around a fifth of its jobs in the industrial sector. The automotive industry has been historically a big part of this sector but has driven a 17% decline in industrial jobs between 2004 to 2018 (OECD, 2021[78]). A large part of this drop in employment can be attributed to automation-led job displacement (Figure 3.7) and most likely happened in waves, as this region has and above average number of mass layoff events in the last 10 years (Figure 1.27). Within the automation sector, the main actor is Fiat, an Italian automaker that has scaled down production in the region since it became part of Stellantis, a multinational automotive manufacturing company.
The region is aiming to shift its development strategy towards one of innovation by providing support for existing businesses and start-ups. These incentives take the form of grants, direct financing and financing guarantees, vouchers, and tax benefits, among others. Although the new development strategy is still ongoing, Piedmont has managed to leverage its existing industry and talent to establish itself as an R&D hub which concentrates over 20% of Italy’s venture capital (VC) funding (EY, 2022[79]).
Georgia, United States
The state of Georgia has a high percentage of jobs at risk of automation (13.5%) compared to the OECD average (8.8%) (OECD, 2024[80]). Manufacturing is the second-largest contributor to Georgia's GDP, primarily in machinery, electrical equipment, and fabricated metals (IBISWorld, 2024[81]). However, Georgia's manufacturing industries face challenges that stem from a combination of global competition, reliance on traditional manufacturing, and a need for modernisation.
The Georgia AI Manufacturing (GA-AIM) coalition, supported by the U.S. Department of Commerce’s Economic Development Administration, received a grant of USD 65 million through the American Rescue Plan Regional Challenge (Georgia AIM, 2024[82]). The initiative focuses on accelerating AI technology integration in key sectors like semiconductors, batteries, electrification, food production, aerospace, and defence. It aims to build regional resilience, catalyse local industries, and create quality jobs, while fostering economic growth, and serving as a national example for how to accelerate automation in manufacturing. The coalition will also establish an AI Manufacturing Pilot Facility at Georgia Tech and seeks to expand job training for underserved communities so that automation benefits workers rather than replace them (EDA, 2022[83]).
AI can be a tool to alleviate current labour market challenges
With over 30% of OECD regions losing population and over 90% ageing,14 a major challenge for policy makers is addressing workforce needs with a shrinking working-age population (OECD, 2024[86]). At the same time, labour shortages are pervasive across OECD regions (see Chapter 2) and may widen given current demographic trends.
AI and related technologies can be leveraged to address current region-specific labour market challenges and provide support to marginalised workers. Regions experiencing workforce shortages may benefit from increased use of AI – either through upskilling workers with AI tools or by directly implementing AI technologies – to fill positions left vacant due to demographic decline or other factors. This is particularly true for regions with a concentration of industries facing challenges in workforce availability (Box 3.13). In such cases, capital investment in AI becomes a viable alternative, with its applications broadening thanks to technological advancements. In addition, low- and medium-skilled workers can also benefit as AI may be used as a tool to close skill gaps.
These technologies can also help mitigate labour shortages by offering career guidance for critical occupations and enhancing labour market matching across regions. Labour markets shortages often arise in seemingly unattractive careers, but career guidance can help address misconceptions and attract talent. In the education sector, for example, Generative AI can be leveraged through chatbots aimed at answering questions and addressing doubts of high school students who are interested in becoming teachers (Box 3.13). There are already examples of AI technologies being used to better match skilled workers with open positions, enhancing regional mobility and, in turn, improving labour market efficiency.
Manufacturing
In the United States, the manufacturing sector has critical labour shortages. It is estimated that even if every skilled worker in the US was employed, there would still be 35% more unfilled job openings in the durable goods manufacturing sector than skilled workers capable of filling them (Gow, 2022[87]). Technology presents itself as a solution, as AI – with the support of robotic automation – can save at least 75% of the labour costs of using humans alone, enable 24-hour continuous production, and help avoid injuries. In fact, the use of technology in tight labour markets helps maximise the productivity gains from any potential automation of work (OECD, 2023[88]).
Generative AI provides novel, more advanced solutions for this sector. Additive manufacturing, such as 3D printing, requires highly skilled designers and engineers to draw on years of experience and a “best guess” approach to arrive at the best design solution. AI now empowers a rapid, generative approach to developing complex and highly optimised design solutions that can be produced quickly through 3D printing. Machine vision and product optimisation are also technologies that can help. For example, the use of cameras and AI that recognise the shape, orientation, and condition of assembly line products under various lighting conditions eliminate the need for human eyes and hands in the questions and answers process.
Education
The education sector is critical to a region’s development, but filling essential positions can be challenging, leading to some of the most relevant examples of labour shortages. In the United States, for example, there were an estimated 55 000 vacant teaching positions at the start of the school year in 2023, 51% more than in 2022, and 270 000 teachers were working without meeting state qualifications. Chile is another example of a country facing a significant teacher shortage, with an estimated deficit of 26 000 qualified teachers for 2025, representing 19% of the required workforce. Regions far from the capital can have a much larger shortage, reaching 40% of the required workforce in regions such as Atacama, Magallanes, and the Chilean Antarctica. This deficit is caused by a low level of attractiveness of the profession and a high desertion rate as 19% of new teachers leave after their first year.
AI tools have emerged to produce learning materials, attract and retain talent, and provide mentoring and support for new teachers. Bookbaker, for example, is a Generative AI tool that allows teachers to create custom learning materials personalised for their students, and aligned with their curricula, therefore freeing up time and resources from already heavy teaching workloads. In Chile, the non-profit organisation Elige Educar has been pioneering the use of AI with two initiatives: Quiero Ser Profe and Somos Profes, Somos Educadores. The former uses chatbots alongside human tutors to provide personalised information and support to students interested in pursuing a teaching career, helping them make informed decisions. The latter employs AI tools to provide mentoring and support to new teachers and early childhood educators during their initial years in the profession.
And other sectors
The practical solutions offered by AI extend to several more sectors, from professional services to agriculture and life sciences. For example, in the legal sector, Latham & Watkins deployed a model that scans documents and produces accurate legal briefs. This tool can reduce contract review time by 60% (Virtasant, 2024[89]). In agriculture, AI-powered drones facilitate precision farming, while other technologies are being used to automate irrigation, pest control, and enhance crop yields. In addition, “smart farms” have emerged in San Francisco, where agricultural activities are conducted in highly controlled environments monitored and managed by digital systems. In life sciences, AI-powered tools designed to assist in patient diagnosis are expected to improve accuracy and free up medical professionals to spend more time on patient care, while AI technologies for drug development are expected to speed up the research process.
The integration of AI in the workplace holds the potential to address skill gaps among workers. Workers with lower skill levels typically face more difficulties in the workplace, such as limited job opportunities, lower wages, and reduced job security. AI-powered tools can be used to enhance workers’ capabilities, thereby levelling the playing field and providing access to better job opportunities. For example, AI-powered tools can assist workers in performing complex tasks that would otherwise be beyond their skillsets, such as coding, and data analysis.
AI can improve the participation in the workforce of people with disabilities who remain significantly underrepresented in the labour market. As of 2019, people with disabilities in OECD countries were 2.3 times more likely to be unemployed compared to their non-disabled counterparts, with 27 percentage points lower employment rate (OECD, 2023[92]). Equipping workplaces with AI tools and teaching workers with disabilities the skills to use them could increase the labour force, closing employment gaps and promoting greater inclusion.
AI-powered tools, such as automated vehicles, can enhance mobility, and improve accessibility in work environments. To harness these benefits, there is a need for pro-active government policies that promote inclusive AI development, and sustainable funding for AI research and accessibility solutions. Mobilising AI-solutions for workforce participation for people with disabilities could be particularly relevant in regions with poor communications or limited accessibility, as well as regions struggling with labour shortages. One example is the Minnesota’s Autonomous Rural Transit Initiative (goMARTI) project in Grand Rapids, Minnesota, United States. This project offers a free self-driving shuttle service designed to help residents, particularly those with mobility challenges, access different locations under rural and winter conditions (goMARTI, 2024[93]).
Policies aimed at closing the AI skills gap could facilitate a more equitable access to technological advancements, benefiting workers of all skill levels. Demand for advanced AI skills - those needed to develop AI systems - such as natural language processing (NLP) or machine learning (ML) is growing rapidly, but they only account for around 1% of jobs postings (Borgonovi et al., 2023[94]). The main policy challenge lies in boosting AI skills which are useful for a larger segment of the population. Recent research shows that 61% of European workers agree that it is fairly or very likely that they will need new knowledge and skills to cope with the impact of AI tools on their work in the next five years, but 44% think it is unlikely that their organisation will provide the training (Cedefop, 2024[95]).
Reliable internet access is essential for workers to fully leverage AI tools, but limited or slow connectivity remains a challenge in certain OECD regions. On average, 84% of households have internet access across the OECD, but this figure can be as low as 50% in some OECD regions. Furthermore, cities experience 13% faster internet on average than the rest of the country (OECD, forthcoming[96]). Consequently, the internet is faster and more accessible in urban areas, which are also the regions with higher levels of exposure to Generative AI. This fact deepens the regional divide in Generative AI exposure as the fewer workers exposed may, in practice, experience even lower levels of adoption due to barriers posed by slow or limited internet access in their region. This underscores the importance of ensuring that, alongside training programmes and software tools, physical infrastructure also receives policy attention.
Impact of AI within the workplace
New roles are emerging to develop and deploy AI solutions, increasing the demand for AI-related skills. In the US, for instance, firms providing AI solutions have increased by 4 percentage points between 2010 and 2017, going from less than 1% to more than 5% of all technology firms (Muro and Liu, 2021[73]). Examples of roles that fine-tune AI tools include data annotators, who label and organise data — such as images, text, or audio — enabling AI models to learn and make accurate predictions; AI operations specialists, who oversee the performance and integration of AI systems in real-world settings; and AI trainers, who improve AI models by providing feedback and adjusting algorithms to enhance performance. Machine learning and natural language processing are some of the most sought-after skills in AI-related job vacancies. Despite the rapid growth of AI, however, less than 1% of all job postings are AI-related (Borgonovi et al., 2023[94]).
AI is also creating new roles to work alongside emerging technologies. Many of these new roles are focused on monitoring, maintaining, and improving the systems that integrate AI into existing operations. For instance, in the mining industry, robotic automation has led to new roles managing machines through digital twins, which are virtual replicas of equipment. These digital twins optimise mining operations and improve worker safety by allowing remote monitoring and control, reducing exposure to hazardous conditions, and increasing operational efficiency through predictive maintenance (Saes, 2024[97]). This example highlights how AI and automation can render some traditional jobs obsolete (e.g., conventional mining) while simultaneously creating new opportunities to work alongside AI technologies.
The labour market impact of AI extends beyond job displacement, job creation, and productivity, as recent evidence suggests that the most significant impact of AI is concentrated on tasks rather than jobs. Instead of fully displacing jobs, most European workers declare that AI has an impact on their job tasks. In fact, 30% of European workers who use AI technologies and tools to do their job experienced a reduction or disappearance of some tasks, while 41% reported new tasks in their jobs. In addition, for 68% of workers, the main effect of AI technologies, so far, has been to enable them to do their job tasks faster (Cedefop, 2024[95])15. This suggests that jobs will not only be created or destroyed but also transformed in their execution and processes. Therefore, issues are expected to arise within the workplace as new skills become essential (or obsolete), knowledge of AI systems becomes necessary, and the use of technology to monitor workers increases.
People are likely to experience AI's impacts within their existing roles, emphasising a growing need for widespread digital skills. To introduce AI solutions in the workplace, worker’s will need to recognise how these tools can augment their capabilities at work and develop specialised skills to create a symbiotic relationship with AI-technologies (Zirar, Ali and Islam, 2023[98]). Digital skills, such as IT literacy or basic knowledge of machine learning models, will be important to work and interact with AI. These skills help workers comprehend how AI solutions operate, and understand AI system's capabilities, limitations, and underlying logic. Other high-level cognitive skills will also be important to understand how AI can fit within worker’s specific tasks, and to make informed decisions based on AI-generated outputs (OECD, 2023[88]). To adapt to AI systems and implement them in their job, workers may need upskilling and reskilling, which could in turn increase trust in AI adoption. Furthermore, in an OECD survey, workers indicated that AI has increased the importance of human and interpersonal skills, such as creativity and communication, more than of AI specialised skills (Lane, Williams and Broecke, 2023[99]).
As AI is incorporated into everyday work life, workers may benefit from increased productivity and reduced time spent on tedious tasks, potentially leading to higher job satisfaction. Software programmers who have integrated an AI tool in their everyday work, for instance, are able to finish more tasks, more quickly, and reported feeling more fulfilled in their job (Box 3.14). In addition, a recent survey of the manufacturing and finance sector indicated that nearly two-thirds (63%) of workers reported AI had improved their enjoyment at work (OECD, 2023[88]). Experimental evidence in the consulting industry shows that real-world tasks can be done better and faster with the use of Generative AI tools. Research also shows that AI can act as a skill-leveller, boosting the performance of low-skilled workers with a smaller effect on high-skilled workers (Box 3.14). Regions that struggle to attract high-skilled workers may therefore benefit from AI as it could enable low- and middle-skilled workers to fill positions that were previously beyond their reach. These examples highlight the importance of AI as a tool to augment productivity rather than replace jobs.
In a recent paper, Dell’Acqua et al. (2023[100]) experimentally tested the performance implications of AI on realistic, complex, and knowledge-intensive tasks. Consultants from the Boston Consulting Group (BCG) were evaluated on their ability to carry out creative, analytical, writing, marketing, and persuasiveness tasks. A total of 7% of the firm’s workforce participated in the experiment. This effort was multidisciplinary, involving multiple types of experiments and hundreds of interviews.
Participants were split into groups and asked to carry out several fictional tasks common to the consulting profession. The test group was allowed to use Chat-GPT 4, a widely utilised Generative AI platform, while the control group was not. All participants underwent a pre-test without any Generative AI technology to establish a baseline for measuring their improvement.
AI had a significant positive impact on all measures. Consultants using AI finished 12.2% more tasks on average, completed tasks 25.1% more quickly, and produced 40% higher quality results than those without AI. The quality of the output was rated by both human and AI graders. Introducing participants to the AI tools and giving them an overview of how they work had no significant effect.
Results suggests that AI is a skill leveller as poor performers were heavily boosted by AI while top performers were moderately boosted. The consultants who scored the worst when initially assessed experienced the largest jump in their performance, 43%, when they were able to use AI. Top consultants also received a boost, but a significantly smaller one.
In contrast, when faced with a particularly hard task, participants who did not have access to Generative AI came out on top. The authors constructed a specific task that was designed to be unsolvable by the Generative AI tool, or at least not correctly. Participants not using AI were successful 84% of the time while the treated group was successful only between 60% and 70% of the time. This result suggests that too much AI can potentially be harmful as users are over-confident of the platform’s ability and fail to check its output. Similar research in a different work environment arrived at a similar conclusion (Dell’Acqua, 2021[101]).
On the other hand, the adoption of AI in some work processes can have negative effects on both job satisfaction and the quality of produced work. For instance, modern practices such as algorithmic management can lead to heightened stress and overall job dissatisfaction (Box 3.15). In addition, although the automation of routine tasks can free up time for more engaging work, it can also have negative effects as it can, if implemented poorly, increase work intensity, and reduce a worker’s agency. The impact of technology on job satisfaction heavily depends on workers' ability to participate in its design and implementation (Gmyrek, Berg and Bescond, 2023[19]). Finally, using Generative AI can result in a lower quality of outputs if it is applied to tasks or activities that it is ill-suited to address (Box 3.14).
Organisational change management will play an important role in determining how AI technologies are embraced in the workplace. While AI has the potential to enhance productivity, improve decision-making, and increase job satisfaction, these benefits may remain unrealised without a proper approach to AI implementation. An OECD survey of employers and workers in the manufacturing and finance sectors showed that, although both employers and workers have generally positive attitudes about the use of AI, concerns over job loss, stability and potential pressures on wages also prevail (Lane, Williams and Broecke, 2023[99]). Early involvement of employees in the adoption process can help them understand the technology's value and potential benefits, reducing resistance to change (Lane, Williams and Broecke, 2023[99]; Hechler, Oberhofer and Schaeck, 2020[102]). Training is also relevant to help employees adapt to new tasks, improve worker’s trust, and facilitate inclusivity in the use of AI systems (Lane, Williams and Broecke, 2023[99]). Companies such as IBM strive to improve technology skills and usage among employees by mandating 40 hours of annual learning. The training includes courses on AI, allowing employees to test IBM's in-house AI tools, which has increased both technology adoption and trust in AI systems within the company.
The effects of algorithmic management on job satisfaction and working conditions emphasises the importance of worker agency and participation in technological adoption. Algorithmic management, which uses data-driven algorithms to manage work tasks and evaluate performance, often reduces worker autonomy and feedback opportunities, while increasing surveillance. This can lead to heightened stress and job dissatisfaction, particularly in environments where workers have little control over their tasks and limited interaction with management (Baiocco et al., 2022[103]).
Amazon warehouse workers in the United Kingdom, for instance, have expressed feeling pressured to work at the fast pace set by automated systems to meet performance targets (GPAI, 2024[104]). Amazon employs a system in their warehouses called Associate Development and Performance Tracker (ADAPT) to monitor performance and provide feedback across a range of dimensions including productivity, quality, safety, and behaviour. The ADAPT system is able to track how many tasks each worker completes, such as how quickly items are packed or processed, and uses this data to assess performance. In some cases, it can generate automatic warnings if workers fail to meet set targets.
Algorithmic management systems for task assignment and monitoring are also common in call centres and in the gig economy (McKensey & Company, 2021[105]). In call centres, AI systems are used to monitor call quality, response times, and employee interactions. In the gig economy, platforms such as Uber and DoorDash rely on AI to track worker performance and manage job allocation. These systems use predictive analytics to match workers with tasks, monitor productivity, and assess customer ratings. While AI's role in managing both customer interactions and gig workers can improve operational efficiency, it also raises concerns about the balance between worker autonomy and automated oversight.
However, technology can also have positive effects on job satisfaction. A recent report shows that when workers are actively involved in the design and implementation of new technologies, the outcomes can be more positive (Gmyrek, Berg and Bescond, 2023[19]). Countries with strong frameworks for workplace consultation and decentralised decision-making, such as those in the Nordic region and Germany, demonstrate higher worker acceptance and better job satisfaction. Thus, the key to maintaining job quality in the face of technological advancements lies in robust worker participation and dialogue, which can help integrate technology in ways that enhance, rather than detract from, job satisfaction.
Platform cooperatives have emerged as alternatives to conventional platforms to help tackle challenges such as job quality, employer status and asymmetry in bargaining power. Unlike traditional platforms, platform cooperatives are owned and democratically managed by their workers that use websites and mobile apps to sell goods and/or services (OECD, 2023[106]), In the context of AI, workers in platform cooperatives could actively participate in decisions about how algorithmic systems are designed and implemented, such as determining the criteria for task allocation. By involving workers in these decisions, platform cooperatives offer a model that could mitigate the risks of low job quality often associated with traditional platform work, where workers have limited agency.
AI's integration into the workplace raises important questions regarding workers' rights. Algorithmic management, where algorithms determine work assignments and evaluate performance, can result in heightened surveillance and work intensity as AI tracks worker behaviours, productivity, and movements in real-time (Box 3.15). This can lead to increased work pressure, stress, and risks to both mental and physical health (OECD, 2024[107]). This type of management can also reduce feedback and complicate collective bargaining, as it is challenging for workers to organise around a management system they cannot directly see or interact with (CLJE LAB, 2024[108]). Furthermore, algorithmic decision-making in hiring, promotions, or terminations is opaque and can lead to biases, potentially discriminating against certain groups. Extensive data collection without appropriate regulations may also compromise workers' privacy by exposing personal information to misuse or unauthorised access.
Putting transparency and responsibility at the core of AI use and giving workers a voice through strengthened social dialogue is critical to safeguarding worker’s rights. Following OECD AI Principles, having trustworthy AI means that AI development and use are safe and respectful of fundamental rights such as privacy, fairness, and labour rights. Additionally, employment-related decisions made by AI should be transparent and understandable by humans (OECD, 2023[109]). It is important for employers, workers, and job seekers to be informed about AI usage and for accountability mechanisms to be clear in case issues arise. Collaboration with social partners can also be critical in shaping how AI is integrated into the workplace and to protect workers' rights (Kinder et al., 2024[110]). Robust policy measures and ethical guidelines are required to maintain fair treatment, protect worker’s data, and provide retraining opportunities to mitigate the risks of displacement. The US Department of Labor (2024[111]), for instance, has published a set of AI Principles and Best Practices for integrating AI responsibly, emphasising worker engagement and training to prevent displacement and facilitate equitable benefits. In the EU, the AI Act (European Parliament, 2024[112]) addresses risks related to safety, health, and fundamental rights, with specific provisions for high-risk applications in the workplace. Although many OECD countries are developing similar initiatives, most AI-specific measures remain non-binding and depend largely on organisations’ ability to self-regulate (OECD, 2023[109]).
Opportunities and challenges of AI adoption for SMEs
Small and medium size enterprises (SMEs) can leverage digital tools, such a Generative AI, to optimise their operations, but support for technology adoption is limited. Recent evidence suggests that SMEs tend to view the opportunities of Generative AI as higher than the risks. These enterprises face challenges in adopting digital technologies including higher cost, a lack of skills, or cybersecurity risks. Initiatives that support AI adoption for SMEs can help improve their operations, allow them to fully realise the potential of these technologies, and minimise risks, enabling SMEs to innovate without falling behind in the rapidly evolving digital landscape (examples in Box 3.16).
SMEs-Digital (Mittlestand-Digital), Germany
The SMEs-Digital initiative is designed to help small and medium-sized enterprises (SMEs), crafts, and start-ups in Germany to navigate the complexities of digital transformation. With 163 local and thematic centres across Germany, its primary goal is to provide these businesses with free orientation and resources to understand and leverage the opportunities presented by digitalisation while addressing the associated challenges.
To achieve this, the initiative offers a comprehensive range of support services specifically tailored to the needs of SMEs. Activities within SMEs-Digital include workshops and training sessions on various technologies, such as AI, which are designed for different levels of expertise. Additionally, SMEs can benefit from personalised consultations with experts, who provide tailored advice on how to implement digital solutions. The initiative also facilitates networking and collaboration opportunities, along with the chance to participate in demonstration projects, allowing SMEs to better understand how digital technologies can be integrated into their processes.
Starting in 2024, the initiative will place a heightened focus on AI and AI readiness. This shift aims to enhance the availability and preparation of high-quality data, which is crucial for the effective implementation of AI applications in SMEs. By fostering a deeper understanding of AI and its potential applications, SMEs-Digital seeks to equip businesses with the tools and knowledge necessary to integrate AI into their operations.
SCALE-AI, Canada
SCALE-AI is a Canadian innovation supercluster programme focused on accelerating the adoption and commercialisation of AI across supply chains. Supported by the Canadian government and managed from its headquarters in Montreal, the programme is a public-private partnership bringing together businesses, research institutions, and universities. The main goal of SCALE AI is to strengthen Canada’s position as a global leader in AI by fostering innovation and collaboration among SMEs and large companies.
SCALE-AI provides a variety of tailored services and funding opportunities to help SMEs adopt AI technologies and improve their operational processes. Activities include hands-on workshops, training sessions, and access to mentorship programmes that guide SMEs through the process of implementing AI solutions. The programme also helps SMEs connect with AI experts and larger organisations and offers funding for AI-based research and development projects, helping companies experiment with AI applications in real-world scenarios. This includes support for data analytics, machine learning, and automation projects, all aimed at improving logistics, optimising inventory management, and enhancing decision-making processes. By providing resources and expertise, SCALE-AI strives to help smaller companies harness the power of AI so they can remain competitive in a rapidly evolving digital economy.
In contrast to earlier technological advancements that required significant infrastructure, Generative AI is more easily accessible. Businesses, including SMEs, can use generative AI through cloud-based software and applications, making it available at a low cost. These benefits, combined with the fact that anyone can use it with minimal knowledge through simple queries, make this application of AI very more accessible for businesses regardless of the limitations related to their size (OECD, 2024[115]).
Nearly 1 in 5 SMEs surveyed in the 2024 OECD D4SME Survey reported experimenting with generative AI less than a year after LLMs became available to the public in late 2022 (OECD, 2024[115]). This is in stark contrast with other applications of AI, which appear to be considerably less widespread among surveyed SMEs. For example, only 6% of SMEs reported they created or acquired tailored machine learning algorithms (produced by either internal or external experts) to be applied in their business functions (OECD, 2024[115]). However, it must be highlighted that many SMEs still lack a proper understanding of AI possible applications or uses in the tools they are already using. Policy should aim to provide tailored support for SMEs in adopting digital technologies, addressing their specific needs, as many SMEs report a lack of awareness of available assistance that meets their requirements (Box 3.17).
The widening gap across enterprises
SME’s can adopt digital tools to optimise their operations, but recent data suggest that they have not kept pace with large enterprises. Digital tools can help streamline operations, operate flexibly, and diversify revenue streams, which can help withstand external shocks. Recent data, however, shows that large enterprises continue to outpace SMEs in the adoption of software technologies, despite progress in SMEs digitalisation.
The adoption of digital technologies by SMEs depends largely on its ability to increase sales, drive efficiency, and boost resilience. A recent survey1 shows that the primary objective of SMEs when adopting digital tools is increasing domestic sales (47%) and expanding their customer base (41%). Automation, which can significantly enhance operational efficiency, is an objective for 40% of business. This objective is more important in the professional services sector (48%), and among more digitally mature business (48%). Reducing operational costs is also an objective, especially for those businesses that operate exclusively online.
However, SMEs face significant bottlenecks in adopting digital tools. Over 1 in 4 SME’s point to bottlenecks to digitalisation, such as costs, skills shortages, and lack of time for training. Digital security practices are also a main concern as 1 in 3 business across surveyed European countries were aware of having been the target of cyber-attacks.
Generative AI has been rapidly embraced by SMEs, which generally view this technological advancement positively. Over half of the respondents (57%) reported a view that opportunities associated with Generative AI features surpassed the risks. Additionally, managers are more optimistic about generative AI than employees, with 62% believing the benefits of the technology outweigh the risks, compared to 48% of employees.
Despite this positive view of generative AI, many SMEs still lack a comprehensive understanding of its potential uses or how to embed it in the tools they already use. Overall, only 26% of respondent SMEs acknowledged their use of AI but, as the survey was distributed to SMEs active on large digital platforms, by design, all respondents in the sample were at minimum using it passively (through the machine learning algorithms embedded in platforms they were using). Overall, only about 5% of respondents recognised they were using AI passively in this sense, highlighting the need for continued support and understanding of AI among SMEs.
Support for SMEs in the adoption of digital tools remains limited. Overall, less than 1 in 5 SME’s are aware of government support for the adoption of digital tools. Furthermore, respondents who were aware of support cited concerns about the suitability of programmes for their digital needs, red tape in accessing the programmes and capacity challenges to sustain digital improvements after the completion of the support programme.
1. The D4SME Survey gathered responses from 1 005 SMEs from Japan (561), Korea (249), Europe-4 [113, combining France (44), Germany (15), Italy (33), and Spain (21)], and the United States (82). The sample was drawn from a distinct set of digital platforms and so cannot be considered as being representative of the entire universe of SMEs in those countries or indeed the OECD.
Source: (OECD, 2024[115]).
AI in the public administration
AI also holds potential to improve productivity and optimise processes in the public administration. By leveraging AI, governments can automate routine tasks, streamline workflows, and enhance decision-making capabilities, thereby freeing up public servants to focus on more complex, citizen-centric activities. AI-driven tools, such as chatbots, predictive analytics, and automated document processing, can contribute to more efficient public service delivery, reduce bureaucratic bottlenecks, and ultimately improve transparency and responsiveness in government operations.
One-in-two OECD public employment services (PES) have already implemented AI solutions in 2024 (Brioscú et al., 2024[116]). Specifically, PES have used AI to understand the needs on jobs seekers and target support, improve their labour market matching and employment services and optimise their own back-office processes and knowledge generation. However, PES should be careful in managing the risks linked to AI, such as privacy concerns, and potential biases, while also supporting PES staff in developing the skills to effectively use AI tools (Box 3.18).
The adoption of AI presents significant opportunities for PES, but it will also require proactive measures to address and minimise the associated risks. The OECD paper “A new dawn for public employment services” (2024[116]) explores the impact of widespread AI adoption in PES. Some of the recommendations in the paper include:
Making accountability, transparency and explainability central in PES AI use. Ensuring that those who design, deploy and operate AI systems are accountable for their proper functioning, creating well-design structures that foster transparency and explainability in AI outcomes, ultimately enhancing trust.
Implementing AI models that strive for data privacy and quality. PES handle sensitive administrative data and must remain vigilant about their data privacy obligations. They should also verify that the data fed into AI systems is accurate and appropriate for the intended use.
Mitigating the risk of bias within AI systems. PES undertake sensitive work that can have implications for the lives of citizens. Ensuring that biases and discrimination are minimised, while prioritising fairness and equality, should be a key focus for PES.
Developing skills among PES staff to work alongside AI and address resistance. Employees may resist AI due to fears of job security or discomfort with new practices. Proactive steps should be taken to close the skills gap and build trust in AI tools, facilitating greater adoption among staff and minimising potential concerns.
Conducting continuous monitoring and evaluation of AI systems. Positive results during development do not guarantee success of the tool post-implementation, making ongoing quality checks necessary to avoid unintended impacts.
Source: (Brioscú et al., 2024[116]).
Tools to match job seekers to employers and tools to design vacancy postings are among the most common uses of AI in PES, with over 20% of PES across the OECD implementing at least one of these (Brioscú et al., 2024[116]). AI-powered job matching solutions have been implemented in counties such as Japan, Canada and Mexico. Supporting the improvement of jobseekers’ CVs and employer’s job postings is another common use of AI in PES. Both the PES of France and Flanders, Belgium use tools that analyses CVs and job ads to identify implicit skills not listed and provide suggestions that can refine job descriptions and candidate profiles (Broecke, 2023[117]). Other common uses include profiling tools that asses a jobseeker’s job finding prospects, tools to provides PES clients with information (usually in the form of chatbots), and tools to help with career management and job search orientation (Box 3.19).
PES can harness AI to better tailor their services to the specific needs of regional labour markets. The services can then be scaled more efficiently while maintaining a strong focus on regional priorities, addressing the unique economic and employment challenges that vary between regions. Examples of this already exist, where counsellors work with AI tools to provide nuanced, region-specific solutions, or AI-powered chatbots that can boost the quality and accessibility of services (Box 3.19).
AI for jobseeker profiling in the Basque employment service
The Basque Country is an Autonomous Community in northern Spain known for its strong manufacturing base and increasingly educated workforce. The region benefits from a robust Vocational Education and Training (VET) system and a variety of active labour market programmes tailoring training and labour market incentives to job demand. Low job quality and the demand for skills, however, are contributing to overqualification in the region (OECD, 2020[118]). Lanbide’s, the region’s PES established in 2011, has a range of responsibilities that include managing active labour market programmes and administering the region’s main income support scheme, the Renta Garantia de Ingresos (RGI).
Recently, Lanbide has developed an AI tool to aid in their profiling process, which improves the management of active employment policies, and supports Lanbide’s career advisors. A profiling process involves classifying or defining jobseekers according to their employability, functioning as a diagnostics tool that identifies the risk levels of individuals regarding their chances of re-entering employment. Lanbide developed a methodology that leverages big data and automates the analytical component of jobseeker profiling.
The main objective of this tool is to classify job seekers to better know them and address four key issues in running the PES. The tool is meant to provide a (1) more personalised and (2) reactive service that helped PES professionals (3) better match job seekers to vacancies and (4) address chronic long-term unemployment.
Machine learning is utilised to profile jobseekers based on their employability, with the resulting analysis presented to counsellors through a comprehensive dashboard. Counsellors can leverage this tool during interviews with jobseekers to guide them towards the most suitable opportunities in the labour market. The methodology can be applied in other organisations and can be tailored to the local needs by, for example, training it on region-specific datasets, as it has been in the Basque region. The tool is currently in an experimental phase, with plans to integrate additional variables, ESCO skills, and macroeconomic trends into the analysis to further enhance the profiling model.
AI-powered career information chatbot in the Austrian employment services
In Austria, the labour market is changing due to technological advances, evolving job roles, and a growing emphasis on digital skills. The Austrian Public Employment Service (AMS) is addressing these challenges through multiple initiatives aimed at facilitating career guidance and continuous education for both new entrants to the workforce, and those looking to reskill or upskill to access new roles.
The AMS Career Information Assistant (Berufsinfomat) is a Generative AI-powered chatbot to assist with career-related inquiries. This chatbot, based on OpenAI technology, provides accessible information about careers, education, and training opportunities. The tool is freely accessible and allows users to ask career-related questions through an interactive chat interface serving a broad audience, including young people, parents, teachers, job seekers, career changers, and those returning to work. It is also used by AMS advisors to help them research career-related questions.
The chatbot offers information on a variety of career-related subjects. These include (1) descriptions of over 2 500 professions, including typical tasks, required skills, and qualifications, (2) suggestions for courses, seminars, and other education pathways such as apprenticeships, schools, and universities, and (3) information on starting salaries, apprentice wages, and wages according to collective agreements.
The interactive chat format provides real-time, personalised responses. The chatbot is designed to be barrier-free, meaning users do not need to register or provide personal data, maintaining users’ anonymity. It can direct users to local AMS offices for more specialised and in-person support. Supporting 90 languages, the Career Information Assistant facilitates accessibility to a broad audience, making the tool easy to use for individuals from different backgrounds and needs.
The use of AI can transform the work of public servants in various ways beyond optimising PES. One common application has been using AI for information processing and speeding up writing tasks. The Danish Environmental Protection Agency (EPA) and the Ministry of Digital Government and Gender Equality, in collaboration with the company cBrain, developed an AI-powered solution that assists with drafting briefing notes, speeches, translations, and summaries, which is being used, for instance, in the Copenhagen and Aarhus city governments. The tool works as a virtual assistant for public servants, using an on-premises LLM to ensure data security, and which is trained with thousands of government documents (cBrain, 2024[121]). Public servants and caseworkers can also use the assistant to ask questions and receive information on specific procedures, thereby improving productivity.
AI can be used for document processing and to sped up time-consuming bureaucratic tasks in public administration. The Danish EPA, for instance, is using an AI tool to accelerate environmental and building permitting processes. The tool is trained on local and national environmental and building permit documents, enabling it to handle case processing and filing, and improving the overall quality of decisions. This automation significantly reduces the workload for public servants and citizens by streamlining application self-service. The tool is currently to assist project developers, generate draft reports, and support civil servants who review and finalise decisions (Knudsen and Søndergaard-Gudmandsen, 2023[122]). cBrain is also piloting an AI chatbot in California, United States, specifically trained in environmental regulations, which can be used during the development of construction projects. A user might, for example, ask the chatbot, "What should I consider to improve the safety of birds in the development of my wind farm?", and the chatbot provides relevant information on applicable regulations and other pertinent resources Beyond permitting, this type of AI application has vast potential, as public administrations frequently manage complex and time-consuming bureaucratic tasks that could be automated with AI (Georgieff, 2024[26]).
Public administrations can leverage AI technologies to improve cybersecurity and address evolving security threats. In terms of system security, AI can accelerate the detection and mitigation of cyber threats, improving response times and safeguarding sensitive data, including user identities and key government datasets. These AI-driven systems can automatically identify vulnerabilities and neutralise cyber risks, offering greater protection against breaches. Additionally, AI is being used to track misinformation, monitor potential radicalisation, and detect other security dangers. For example, the AI tools developed by the company Cybara are being used in regional and local governments to track risks on social media, helping to combat misinformation, and identify early signs of foreign influence and coordinated disinformation campaigns. The company often collaborates with government intelligence units and police forces, and has worked on cases such as local elections, and to combat disinformation related to vaccines (Cyabra, 2024[123]).
The application of AI in public administration has great potential to improve productivity, but raises concerns similar to those seen with PES and that might impact their adoption (Box 3.18). One of the main challenges is the need for extensive training for personnel to effectively use AI systems, which often require new technical skills. There are also concerns about potential biases embedded in AI algorithms, which could perpetuate inequality or lead to unfair outcomes. Additionally, attitudes toward change play a significant role, as both employees and citizens may be resistant to the implementation of new technologies. Public acceptance is crucial, as citizens may be wary of AI’s role in decision-making processes, fearing a lack of transparency or loss of human oversight. Striving for transparency, deploying AI ethically, and with careful consideration of its social impact will be central in maintaining public trust.
In conclusion, by tailoring strategies to regional needs, policies can support economic growth, workforce development, and fair integration of AI technologies in the workplace. The following place-based recommendations aim to help local labour markets harness the benefits of Generative AI, while addressing potential risks and promoting equitable use.
1. Identify opportunities where AI can drive regional growth: By building local AI capabilities, regions can modernise traditional industries, attract new investments, and harness emerging technologies for sustainable development. This approach can be particularly valuable for regions facing demographic pressures, low job creation, or outdated industrial structures, and to close urban-rural divides.
2. Assess regional labour market exposure to different forms of AI: Monitoring the varied effects of AI and automation allows for targeted interventions that address specific needs in different regions. Tracking high-risk occupations can further enable regions to anticipate job displacement trends, address skill mismatches, and implement place-based policies to foster resilience and workforce adaptation.
3. Use data to develop comprehensive regional skills inventories: Developing a skills inventory involves assessing workforce’s competencies and identifying gaps, providing a picture of available skills and areas for development. In regions with high exposure to Generative AI, this inventory can be particularly helpful in guiding efforts to adapt the workforce, address potential skill shortages, and capitalise on the opportunities AI offers.
4. Foster collaboration with local stakeholders to strengthen policy intelligence: Engaging local stakeholders, such as employers, educational institutions, and community organisations, provides valuable, real-time insights that can enhance the effectiveness of workforce policies. Such collaboration promotes better-informed, and adaptable policies, and can be particularly important for the integration of Generative AI, as it can improve its uptake, foster trust among stakeholders, and tailor policies to regional economic conditions.
5. Build awareness of AI's benefits for workers and employers: Emphasising how AI can enhance productivity by automating routine tasks and freeing up time for more meaningful work can help workers and firms see its value. Engaging employees early in discussions about implementing AI in the workplace, along with providing training, can build further acceptance and make adoption smoother.
6. Improve the uptake of AI tools across businesses, with a special focus on SMEs: Businesses may need additional resources to take full advantage of AI tools, such as targeted training programmes, guidance on implementing AI technologies, and workforce retraining. Special attention should be given to SMEs, which often lag behind in technology adoption, by providing the support they need to remain competitive in an AI-driven economy.
7. Leverage AI in the public administration, including public employment services (PES): Using AI in the public administration can automate routine tasks, enhance accuracy, and improve service delivery through tools like multilingual chatbots and AI-powered document processing. In PES, AI tools can improve job matching, connecting job seekers and employers more effectively.
8. Establish frameworks to control AI risks: AI presents risks, including privacy concerns, biases, increased pressures on workers, and potential job displacement. Addressing these issues requires clear guidelines and ethical standards to enhance transparency and accountability, and safeguard workers’ rights. Collaboration with social partners can further facilitate responsible AI implementation.
9. Provide tailored support for displaced workers: Technological advancements, such as AI and automation, lead to shifts in the labour market and may cause worker displacement. Targeted support for these workers, including retraining programs and re-employment assistance, can help mitigate long-term economic losses and prevent vulnerable groups and regions from being left behind amid technological changes.
10. Revise regional skill provisions to address workforce needs: Addressing changing skill needs, such as those driven by the adoption of Generative AI in the workplace, requires updating the skills provision system to better align with current demands. This includes revising vocational education programmes, expanding access to adult education, and introducing targeted programmes for critical roles in emerging industries.
References
[7] Acemoglu, D. and J. Loebbing (2022), “Automation and Polarization”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.4238255.
[39] Acemoglu, D., A. Manera and P. Restrepo (2020), “Does the us tax code favor automation?”, Brookings Papers on Economic Activity, Vol. 2020/1, https://doi.org/10.1353/ECA.2020.0003.
[38] Acemoglu, D. and P. Restrepo (2022), “A task-based approach to inequality”, IFS Deaton Review of Inequality, https://ifs.org.uk/inequality/a-task-based-approach-to-inequality.
[35] Acemoglu, D. and P. Restrepo (2021), “Tasks, Automation, and the Rise in Us Wage Inequality”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.3866352.
[34] Acemoglu, D. and P. Restrepo (2019), Automation and new tasks: How technology displaces and reinstates labor, https://doi.org/10.1257/jep.33.2.3.
[5] Aghion, P. et al. (2022), “Modern manufacturing capital, labor demand, and product market dynamics: Evidence from France”, Working Papers hal-03943312, HAL, https://ideas.repec.org/p/hal/wpaper/hal-03943312.html.
[63] AI Medical Services Corporation (2024), About AIM, https://en.ai-ms.com/.
[60] Aillis (2024), About Aillis, https://www.aillis.jp/.
[58] AISmiley (2024), Utilizing AI in nursing care and welfare! Case studies and future prospects introduced from the perspective of resolving labor shortages, https://aismiley.co.jp/ai_news/examples-of-utilization-of-ai-in-the-field-of-long-term-care-and-welfare/.
[1] Akst, D. (2013), What can we learn from past anxiety over automation? by Daniel Akst — Summer 2013: Where Have All The Jobs Gone? | The Wilson Quarterly, Wilson Quarterly, https://www.wilsonquarterly.com/quarterly/summer-2014-where-have-all-the-jobs-gone/theres-much-learn-from-past-anxiety-over-automation.
[37] Albanesi, S. et al. (2023), “New Technologies and Jobs in Europe”, Banco de España. Documentos de Trabajo 2322, https://www.bde.es/f/webbe/SES/Secciones/Publicaciones/PublicacionesSeriadas/DocumentosTrabajo/23/Files/dt2322e.pdf.
[57] AMED (2024), Research project for medical-engineering collaboration and implementation of artificial intelligence, https://www.amed.go.jp/en/program/list/14/05/014.html (accessed on 15 November 2024).
[12] Anyoha, R. (2017), “The History of Artificial Intelligence - Can Machines Think ?”, Harvard University.
[44] Australian Department of Employment and Workplace Relations (2023), Australian Digital Capability Framework: Version 1.0.
[9] Autor, D. (2015), Why are there still so many jobs? the history and future of workplace automation, https://doi.org/10.1257/jep.29.3.3.
[18] Autor, D., F. Levy and R. Murnane (2003), “The skill content of recent technological change: An empirical exploration”, Quarterly Journal of Economics, Vol. 118/4, https://doi.org/10.1162/003355303322552801.
[103] Baiocco, S. et al. (2022), “The Algorithmic Management of work and its implications in different contexts”, JRC Working Papers Series on Labour, Education and Technology.
[4] Banco de España (2023), El mercado de trabjo español: situación actual, tendencias estructurales y políticas de empleo, https://www.bde.es/f/webbe/SES/Secciones/Publicaciones/PublicacionesAnuales/InformesAnuales/23/Fich/InfAnual_2023_Cap3.pdf.
[10] Bessen, J. (2015), Toil and technology: Innovative technology is displacing workers to new jobs rather than replacing them entirely.
[36] Böckerman, P. (ed.) (2020), “Automation, workers’ skills and job satisfaction”, PLOS ONE, Vol. 15/11, p. e0242929, https://doi.org/10.1371/journal.pone.0242929.
[13] Bonney, K. et al. (2024), “The impact of AI on the workforce: Tasks versus jobs?”, Economics Letters, Vol. 244, p. 111971, https://doi.org/10.1016/j.econlet.2024.111971.
[94] Borgonovi, F. et al. (2023), “Emerging trends in AI skill demand across 14 OECD countries”, OECD Artificial Intelligence Papers, No. 2, OECD Publishing, Paris, https://doi.org/10.1787/7c691b9a-en.
[116] Brioscú, A. et al. (2024), “A new dawn for public employment services: Service delivery in the age of artificial intelligence”, OECD Artificial Intelligence Papers, https://doi.org/10.1787/5dc3eb8e-en.
[117] Broecke, S. (2023), Artificial intelligence and labour market matching, OECD Publishing, Paris, https://doi.org/10.1787/2b440821-en.
[59] Brown, D. et al. (2024), “A behaviourally informed chatbot increases vaccination rates in Argentina more than a one way reminder”, Nature Human Behaviour, https://doi.org/10.1038/s41562-024-01985-7.
[55] Cabinet Secretariat (2024), Digital Administrative and Financial Reform.
[2] Calvino, F. and L. Fontanelli (2023), “A portrait of AI adopters across countries: Firm characteristics, assets’ complementarities and productivity”, OECD Science, Technology and Industry Working Papers, No. 2023/02, OECD Publishing, Paris, https://doi.org/10.1787/0fb79bb9-en.
[121] cBrain (2024), “F2 AI Assistant for government”, https://cbrain.com/f2-ai-assistant.
[95] Cedefop (2024), “AI skills in the workplace? Cedefop asked the workers”, https://www.cedefop.europa.eu/en/press-releases/ai-skills-workplace-cedefop-asked-workers.
[43] Central Japan Startup Ecosystem (2024), About the Central Japan Startup Ecosystem, https://central-startup.jp/en/.
[74] Chan, G. and L. McDonough (2024), “Project Insights Report: Accelerating the Appropriate Adoption of Artificial Intelligence in Healthcare”, Future Skills Centre, https://fsc-ccf.ca/projects/ai-health-care/.
[108] CLJE LAB (2024), “Worker Power and Voice in the AI Response”, Center for Labor and a Just Economy at Harvard Law School, https://clje.law.harvard.edu/app/uploads/2024/01/Worker-Power-and-the-Voice-in-the-AI-Response-Report.pdf.
[123] Cyabra (2024), “Public Sector”, https://cyabra.com/solutions/public-sector/.
[101] Dell’Acqua, F. (2021), “Falling Asleep at the Wheel: Human / AI Collaboration in a Field Experiment on HR Recruiters”, Working paper June.
[100] Dell’Acqua, F. et al. (2023), “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.4573321.
[32] Dingel, J. and B. Neiman (2020), “How many jobs can be done at home?”, Journal of Public Economics, Vol. 189, https://doi.org/10.1016/j.jpubeco.2020.104235.
[83] EDA (2022), “U.S. Department of Commerce Invests Approximately $65 Million to Accelerate Integration of Artificial Intelligence Technologies in Industry in Georgia Through American Rescue Plan Regional Challenge”, US Economic Development Administration, https://www.eda.gov/news/press-release/2022/09/02/us-department-commerce-invests-approximately-65-million-accelerate.
[90] Elige Educar (2020), Análisis y proyección de la dotación docente en contextos rurales, https://eligeeducar.cl/investigaciones-realizadas/analisis-y-proyeccion-de-la-dotacion-docente-en-contextos-rurales/.
[47] Eloundou, T. et al. (2023), GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models, https://arxiv.org/abs/2303.10130.
[15] Epoch (2024), Training computation (petaFLOP) dataset, Parameter, Compute and Data Trends in Machine Learning, https://ourworldindata.org/grapher/training-computation-vs-dataset-size-in-notable-ai-systems-by-researcher-affiliation (accessed on 18 October 2024).
[112] European Parliament (2024), “AI Act. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence”, https://eur-lex.europa.eu/eli/reg/2024/1689/oj.
[124] Eurostat (2018), Methodological manual on territorial typologies, https://data.europa.eu/doi/10.2785/930137.
[79] EY (2022), EY Venture Capital Barometer Italia.
[84] fDi Intelligence (2023), “Piedmont’s hopes for start-ups to future-proof its economy”, https://www.fdiintelligence.com/content/feature/piedmonts-hopes-for-startups-to-futureproof-its-economy-82993 (accessed on 30 September 2024).
[113] Federal Ministry for Economic Affairs and Climate Action (2024), “Mittelstand-Digital”, https://www.mittelstand-digital.de/MD/Redaktion/DE/Publikationen/mittelstand-digital-flyer.html.
[21] Felten, E., M. Raj and R. Seamans (2023), “How will Language Modelers like ChatGPT Affect Occupations and Industries?”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.4375268.
[48] Felten, E., M. Raj and R. Seamans (2021), “Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses”, Strategic Management Journal, Vol. 42/12, https://doi.org/10.1002/smj.3286.
[69] Filippucci, F. et al. (2024), “Should AI stay or should AI go: The promises and perils of AI for productivity and growth”, VoxEU Column: Productivity and Innovation, https://cepr.org/voxeu/columns/should-ai-stay-or-should-ai-go-promises-and-perils-ai-productivity-and-growth.
[16] Filippucci, F. et al. (2024), “The impact of Artificial Intelligence on productivity, distribution and growth: Key mechanisms, initial evidence and policy challenges”, OECD Artificial Intelligence Papers, No. 15, OECD Publishing, Paris, https://doi.org/10.1787/8d900037-en.
[76] Future Skills Centre (2024), “Facing the challenge of digital transformation in the insurance sector: women at work”, https://fsc-ccf.ca/projects/facing-the-challenge-of-digital-transformation-in-the-insurance-sector-women-at-work/.
[75] Future Skills Centre (2024), “From data to decision: AI training and professional certification”, https://fsc-ccf.ca/projects/from-data-to-decision-ai-training-and-professional-certification/.
[82] Georgia AIM (2024), “Projects”, https://georgiaaim.org/projects/.
[26] Georgieff, A. (2024), “Artificial Intelligence and Wage Inequality”, OECD Artificial Intelligence Papers, OECD, https://doi.org/10.1787/bf98a45c-en.
[40] Georgieff, A. and R. Hyee (2021), Artificial intelligence and employment: New cross-country evidence, OECD Social, Employment and Migration Working Papers, https://doi.org/10.1787/c2c1d276-en.
[8] Georgieff, A. and A. Milanez (2021), “What happened to jobs at high risk of automation?”, OECD Social, Employment and Migration Working Papers 255, https://doi.org/10.1787/10bc97f4-en.
[19] Gmyrek, P., J. Berg and D. Bescond (2023), “Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.4584219.
[77] Gobierno de La Rioja (2024), “Belinda León destaca la importancia “de la implementación de tecnologías avanzadas para reforzar la competitividad de las empresas riojanas””, https://www.larioja.org/innovacion/es/noticias/noticia-innovacion/belinda-leon-destaca-importancia-implementacion-tecnologias (accessed on 30 September 2024).
[93] goMARTI (2024), “About”, https://www.gomarti.com/about.
[87] Gow, G. (2022), The Labor Shortage Is Killing American Manufacturing. Here’s How AI Can Bring It Back To Life..
[104] GPAI (2024), “Fairwork Amazon Report 2024: Transformation of the Warehouse Sector through AI”, https://gpai.ai/projects/future-of-work/fairwork-amazon-report-2024.pdf.
[102] Hechler, E., M. Oberhofer and T. Schaeck (2020), AI and Change Management, Apress, Berkeley, CA, https://doi.org/10.1007/978-1-4842-6206-1_10.
[23] Hering, A. (2023), Indeed’s AI at Work Report: How GenAI Will Impact Jobs and the Skills Needed to Perform Them, INDEED.
[81] IBISWorld (2024), “Georgia - State Economic Profile”, https://www.ibisworld.com/united-states/economic-profiles/georgia/.
[61] J-Startup (2024), J-Startup summary, https://www.j-startup.go.jp/en/about/.
[42] J-Startup Kansai (2024), Creating a Role Model from the Kansai Region, https://next-innovation.go.jp/j-startup-kansai/english/.
[68] Kalliamvakou, E. (2022), “Research: quantifying GitHub CoPilot’s impact on developer productivity and happiness”, The GitHub Blog, Sep.
[66] Khan, B. et al. (2023), “Drawbacks of Artificial Intelligence and Their Potential Solutions in the Healthcare Sector”, Biomedical Materials & Devices, Vol. 1/2, pp. 731-738, https://doi.org/10.1007/s44174-023-00063-2.
[110] Kinder, M. et al. (2024), “Generative AI, the American worker, and the future of work”, Blookings, https://www.brookings.edu/articles/generative-ai-the-american-worker-and-the-future-of-work/.
[122] Knudsen, C. and F. Søndergaard-Gudmandsen (2023), “Use Cases for AI in Government: Accelerating Permitting Processes”, cBrain, https://cbrain.com/ai-for-government.
[3] Koch, M., I. Manuylov and M. Smolka (2021), “Robots and Firms”, The Economic Journal, Vol. 131/638, pp. 2553-2584, https://doi.org/10.1093/ej/ueab009.
[119] Lanbide – Basque Employment Service (2024), AI-Machine Learning Serving People and Employment – Profiling Jobseekers.
[99] Lane, M., M. Williams and S. Broecke (2023), “The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers”, OECD Social, Employment and Migration Working Papers, No. 288, OECD Publishing, Paris, https://doi.org/10.1787/ea0a0fe1-en.
[27] Lassébie, J. and G. Quintini (2022), “What skills and abilities can automation technologies replicate and what does it mean for workers? New evidence”, OECD Social, Employment and Migration Working Papers 282, https://doi.org/10.1787/646aad77-en.
[72] Lopez Cobo, M. (ed.) (2022), AI Watch Index 2021, Publications Office of the European Union, https://doi.org/10.2760/921564.
[25] Manning, S. (2024), AI’s impact on income inequality in the US Interpreting recent evidence and looking to the future.
[11] McCulloch, W. and W. Pitts (1943), “A logical calculus of the ideas immanent in nervous activity”, The Bulletin of Mathematical Biophysics, Vol. 5/4, https://doi.org/10.1007/BF02478259.
[105] McKensey & Company (2021), “An on-demand revolution in customer-experience operations?”, https://www.mckinsey.com/capabilities/operations/our-insights/an-on-demand-revolution-in-customer-experience-operations.
[49] Mehdi, T. and R. Morissette (2024), “Experimental Estimates of Potential Artificial Intelligence Occupational Exposure in Canada”, Analytical Studies Branch Research Paper Series - Statistics Canada, https://doi.org/10.25318/11f0019m2024005-eng.
[56] Ministry of Health, L. (2023), 17th Healthcare AI Development Acceleration Consortium Roadmap.
[45] Ministry of Science and ICT of Korea (2021), Strategy to realize trustworthy artificial intelligence.
[125] MIT Technology Review (2013), 10 Breakthrough Technologies, https://www.technologyreview.com/10-breakthrough-technologies/2013/.
[24] Moll, B., L. Rachel and P. Restrepo (2022), “Uneven Growth: Automation’s Impact on Income and Wealth Inequality”, Econometrica, Vol. 90/6, https://doi.org/10.3982/ecta19417.
[73] Muro, M. and S. Liu (2021), “The geography of AI: Which cities will drive the artifical intelligence revolution?”, Metropolitan Policy Program at Brookings, https://www.brookings.edu/wp-content/uploads/2021/08/AI-report_Full.pdf.
[17] Nedelkoska, L. and G. Quintini (2018), “Automation, skills use and training”, OECD Social, Employment and Migration Working Papers, https://doi.org/10.1787/2e2f4eea-en.
[65] NHS (2023), “The Topol Review: Preparing the healthcare workforce to deliver the digital future”, NHS: Health Education England, https://topol.hee.nhs.uk/.
[64] NHS England (2024), “NHS Long Term Workforce Plan”, https://www.england.nhs.uk/long-read/nhs-long-term-workforce-plan-2/.
[67] Nicholson Price II, W. (2019), Risks and remedies for artificial intelligence in health care, Brookings, https://www.brookings.edu/articles/risks-and-remedies-for-artificial-intelligence-in-health-care/.
[51] O*NET (2024), National Center for O*NET Development. Browse by Job Family. O*NET OnLine, https://www.onetonline.org/find/family (accessed on 10 October 2024).
[50] O*NET (2024), National Center for O*NET Development. Browse by Job Zone. O*NET OnLine, https://www.onetonline.org/find/zone (accessed on 21 October 2024).
[70] OECD (2024), “Miracle or Myth? Assessing the macroeconomic productivity gains from Artificial Intelligence”, Working Party No. 1 on Macro-Economic and Structural Policy Analysis, https://one.oecd.org/official-document/ECO/CPE/WP1(2024)14/en.
[86] OECD (2024), OECD Database on Regions, cities and local, http://oe.cd/geostats.
[30] OECD (2024), OECD Employment Outlook 2024: The Net-Zero Transition and the Labour Market, OECD Publishing, Paris, https://doi.org/10.1787/ac8b3538-en.
[41] OECD (2024), OECD Regions and Cities databases, http://oe.cd/geostats.
[115] OECD (2024), “SME Digitalisation to manage shocks and transitions: 2024 OECD D4SME survey”, OECD SME and Entrepreneurship Papers 62, https://doi.org/10.1787/eb4ec9ac-en.
[107] OECD (2024), “TUAC Statement to the OECD Ministerial Council Meeting (MCM) 2024”, https://one.oecd.org/document/C/MIN(2024)20/FINAL/en/pdf.
[80] OECD (2024), “Twin Transition Tracker: Assessing Regional Resilience”, https://www.oecd.org/en/data/dashboards/twin-transition-tracker-assessing-regional-resilience.html.
[28] OECD (2023), Job Creation and Local Economic Development 2023: Bridging the Great Green Divide, OECD Publishing, Paris, https://doi.org/10.1787/21db61c1-en.
[88] OECD (2023), OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market, https://doi.org/10.1787/19991266.
[109] OECD (2023), OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market, OECD Publishing, Paris, https://doi.org/10.1787/08785bba-en.
[106] OECD (2023), Platform cooperatives and employment: An alternative for platform work, OECD Publishing, https://doi.org/10.1787/3eab339f-en.
[54] OECD (2023), Summer Academy on Cultural and Creative Industries and Local Development 6th Edition - Disrupting tradition: How digital technology is changing the cultural and creative processes.
[92] OECD (2023), Using AI to support people with disability in the labour market: Opportunities and challenges, OECD, https://doi.org/10.1787/008b32b7-en.
[14] OECD (2022), OECD Framework for the Classification of AI systems, OECD Digital Economy Papers, https://doi.org/10.1787/cb6d9eca-en.
[53] OECD (2022), The Culture Fix: Creative People, Places and Industries, Local Economic and Employment Development (LEED), OECD Publishing, Paris, https://doi.org/10.1787/991bb520-en.
[78] OECD (2021), Regional Innovation in Piedmont, Italy: From Innovation Environment to Innovation Ecosystem, OECD Regional Development Studies, OECD Publishing, Paris, https://doi.org/10.1787/7df50d82-en.
[31] OECD (2020), OECD Digital Economy Outlook 2020, OECD Publishing, Paris, https://doi.org/10.1787/bb167041-en.
[118] OECD (2020), Preparing the Basque Country, Spain for the Future of Work, OECD Reviews on Local Job Creation, OECD Publishing, Paris, https://doi.org/10.1787/86616269-en.
[6] OECD (2019), “Digitalisation and productivity: A story of complementarities”, OECD Publishing, Paris, https://doi.org/10.1787/5713bd7d-en.
[33] OECD (2018), Job Creation and Local Economic Development 2018: Preparing for the Future of Work, OECD Publishing, Paris, https://doi.org/10.1787/9789264305342-en.
[96] OECD (forthcoming), OECD Regions and Cities at a Glance 2024, OECD.
[22] Pizzinelli, C. (2023), “Labor Market Exposure to AI: Cross-country Differences and Distributional Implications”, IMF Working Papers, Vol. 2023/216, https://doi.org/10.5089/9798400254802.001.
[91] Post, R. (2024), How Can AI Help Solve Teacher Shortages?, https://www.aaspa.org/news/how-can-ai-help-solve-teacher-shortages (accessed on 30 September 2024).
[120] Public Employment Service Austria (2024), “AMS Berufsinfomat”, https://www.ams.at/arbeitsuchende/aus-und-weiterbildung/berufsinformationen/berufsinformation/berufsinfomat.
[85] Regione Piemonte (2024), Agevolazioni per le imprese, https://www.regione.piemonte.it/web/temi/sviluppo/agevolazioni-per-imprese (accessed on 30 September 2024).
[97] Saes, C. (2024), “Digital Twins For Mining Industries: A Transformative Technology”, Forbes, https://www.forbes.com/councils/forbesbusinesscouncil/2024/02/21/digital-twins-for-mining-industries-a-transformative-technology/.
[114] SCALE AI (2024), “Projects”, https://www.scaleai.ca/projects/.
[46] Song, K. (2022), “Korea is leading an exemplary AI transition. Here’s how.”, The AI Wonk, https://oecd.ai/en/wonk/korea-ai-transition.
[71] Statista (2024), “Artificial intelligence (AI) market size worldwide from 2020 to 2030”, https://www.statista.com/forecasts/1474143/global-ai-market-size.
[111] U.S. Department of Labor (2024), “Artificial Intelligence And Worker Well-being: Principles And Best Practices For Developers And Employers”, https://www.dol.gov/sites/dolgov/files/general/ai/AI-Principles-Best-Practices.pdf.
[62] Ubie (2024), About Ubie, https://ubie.life/about_ubie.
[52] UNESCO (2012), International Standard Classification of education ISCED 2011, https://isced.uis.unesco.org/.
[89] Virtasant (2024), Stress, Burnout, and Labor Shortages: How AI in the Workplace is Making Core Business Problems a Thing of the Past, https://www.virtasant.com/ai-today/ai-in-the-workplace.
[29] Vona, F., G. Marin and D. Consoli (2019), Measures, drivers and effects of green employment: Evidence from US local labor markets, 2006-2014, https://doi.org/10.1093/jeg/lby038.
[20] Webb, M. (2019), “The Impact of Artificial Intelligence on the Labor Market”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.3482150.
[98] Zirar, A., S. Ali and N. Islam (2023), “Worker and workplace Artificial Intelligence (AI) coexistence: Emerging themes and research agenda”, Technovation, Vol. 124, p. 102747, https://doi.org/10.1016/j.technovation.2023.102747.
The two measures of exposure to Generative AI are used in this report: exposure now and exposure now or in the near future. The former is defined as the share of tasks within an occupation that can be completed in half the time with the use of LLMs in their current form i.e., Chat-GPT 3.5 or similar. In particular, this measure includes those tasks classified under E1 (Annex Table 3.A.1). The latter is defined as the share of tasks within an occupation that can be completed in half the time with LLMs in their current form or it is easy to imagine additional software that could be developed on top of the LLM that would reduce the time it takes to complete the task by half. In particular, this measure includes those tasks classified under E1 and E2.
Note: Tasks performed with the use of Generative AI should be of equivalent quality, this is, a third party would not notice or care about LLM assistance.
Source: (Eloundou et al., 2023[47])
Source: OECD elaboration
The degree of urbanisation (DEGURBA) is a classification that indicates the character of an area by classifying the territory of a country on an urban-rural continuum. The classification uses local administrative units (LAUs or communes) and classifies these as cities, towns and suburbs, or rural areas based on a combination of geographical contiguity and population density. The basis for the classification is the data for 1 km² population grid cells. The categories are described as follows:
Cities: densely populated areas where at least 50% of the population lives in one or more urban centres.
Towns and suburbs: intermediate density areas where less than 50% of the population lives in an urban centre and at least 50% of the population lives in an urban cluster.
Rural areas: thinly populated areas where more than 50% of the population lives in rural grid cells.
Source: (Eurostat, 2018[124])
← 1. Generative AI refers to artificial intelligence that can create new content by learning patterns from existing data. Current uses of Generative AI include content creation (including text and images), code generation, and personalised recommendations, with technologies such as ChatGPT, DALL-E, and GitHub Copilot being some recognised examples of Generative AI platforms.
← 2. MIT Technology Review (2013[125]). “10 Breakthrough Technologies 2013”. https://www.technologyreview.com/10-breakthrough-technologies/2013/ accessed on 18/07/2024.
← 3. Authors name this middle group the “Big Unknown” and estimate that a full 8.6% of global employment (281 million workers) falls within this category.
← 4. Note that this study uses the Generative AI platform GPT-4 to judge Generative AI itself, which might lead to bias. Nevertheless, the direction and magnitude of any potential bias is unclear.
← 5. This includes mechanical technologies, and non-generative AI. This may include some early forms of LLMs but only to a limited extent as the most prominent Generative AI platforms were released after this date. For the purpose of this chapter, Generative AI encompasses models and platforms released in or after 2022.
← 6. Note that regional dispersion is impacted by the size and number of regions, as countries with a higher number of smaller regions tend to have more regional dispersion. This explains, at least in part, the small variance in, for example, Australia and Canada.
← 7. Once country-specific factors are considered, the correlation is insignificant. Qualitatively similar results are found by (Georgieff and Milanez, 2021[8]).
← 8. Labour productivity is measured as gross value added per employment at place of work by main economic activity (OECD, 2024[41]).
← 9. Additionally, some productivity gains may not have manifested as increased sales or income, but rather as consumer surplus, which would not be reflected in this particular measure of productivity.
← 10. Exposure to generative AI does not make a job more or less likely to be displaced, it simply means that generative AI is a useful tool for enhancing efficiency in that occupation.
← 12. Healthcare occupations include general and specialist doctors, dentists, medical and dental practitioners, assistants and technicians, chiropractors, physical therapists and assistants, medical and health service managers, podiatrists, nurses, midwives, acupuncturists, veterinarians, opticians, genetic counsellors, pharmacists and complementary medicine associates. Exact occupations depend on occupation classification.
← 13. Software-related occupations includes developers, programmers, computer technicians, ICT support, web and digital designers, computer operators, network and computer systems administrators, computer scientists and specialists, system and security analysts, database managers and designers, data entry keyers, statisticians, and data scientists. Exact occupations depend on occupation classification.
← 14. An ageing region is one where the elderly dependency ratio has increased in the last 10 years (to 2022 or 2023 depending on last year available). Similarly, a region is losing population if it has less population than it did 10 years ago. This analysis is done at the TL-3 level where possible, and if not, it is done at the TL2 level.
← 15. The survey was conducted between February and May 2024 in 11 EU countries. The sample size was around 500 adult workers in each country, with a total of 5 342 interviews.