Technological change both modifies the world of work and impacts the lives of billions of workers throughout the world. Growing evidence from developed countries suggests that the advance of automation technologies, both traditional and those related to artificial intelligence (AI), will cause many jobs to become automated and disappear. Other jobs will be modified in their task composition, while new jobs and occupations will emerge. It remains to be seen how such trends will affect workers, particularly in developing countries in which informal work is the norm. Drawing on microdata from the Key Indicators of Informality based on Individuals and their Household (KIIbIH) and estimates from the literature, this chapter assesses the potential impacts of traditional and AI automation on informal employment across a large sample of developing countries. The chapter starts by exploring the impacts of technological change on the world of work from a global perspective. It then examines the risk of automation for jobs held by informal and formal workers in developing countries and identifies the socio-economic groups that face higher risks. The chapter concludes by proposing policy options to mitigate the adverse effects of automation on workers and their families while strengthening some of its benefits.
Securing the Livelihoods of Informal Economy Workers in Times of Global Changes
Abstract
Automation and informal workers employment prospects: Evidence and policies
Automation is shaping the future of work by disrupting and changing job tasks, labour productivity, employment patterns, job availability and the skill sets required for workers in developing countries.
Robots are displacing industrial workers in routine manual task occupations (e.g. assembly line operators) while new technologies, such as generative AI, are increasingly transforming jobs in non-routine cognitive tasks (e.g. interpreters). AI-empowered robots could also impact workers in non-routine manual occupations.
The opportunities and challenges for workers to adapt to technological change, use both industrial robots and AI, and update their skill sets for new forms of production differ significantly in formal and informal economies.
Informal workers, compared with formal workers, typically hold jobs that are much more exposed to high risks of replaceability by traditional automation, as they are extensively engaged in jobs with manual routine and non-routine tasks. Informal, young and older workers living in rural areas, with low or middle earnings, are the most vulnerable.
Formal workers face higher risks of AI automation as generative AI tends to automate information processing and codifiable cognitive tasks that are more prevalent within the content of formal jobs.
Due to sector-specific technologies and production processes, the occupational structure of economies shapes the task content of jobs. This is the main channel for automation risks that affect formal and informal workers according to their respective positions in diverse industries. In fact, automation risks vary due to the different occupations demanded by firms.
Tackling diverse technological and economic shocks requires differentiated and targeted policies
Policies should support skill development plans that enable workers to complement the tasks taken over by new technologies, coupled with measures that facilitate mobility across sectors and occupations. Foundational skills, such as literacy and numeracy as well as digital skills, problem-solving and interpersonal skills are critical for lifelong learning. As such, they should be at the heart of skills development strategies
In anticipation of significant shocks linked to traditional automation and, to a lesser extent, AI automation, social protection systems must integrate broad, scalable social protection floors. Such programmes need to be easily and quickly activated to protect all workers at risk, especially informal workers. Modalities for delivering social protection also need to acknowledge the limit of targeting mechanisms in contexts of high informality, in which means testing and other methods can produce substantial inclusion and exclusion errors.
Several labour‑intensive activities remain at relatively low risk of traditional and AI automation. Governments should aim to expand employment in those labour‑intensive sectors while raising productivity through better skills, management and digital tools that augment1 workers – rather than replacing them. Policies promoting innovation, entrepreneurship and job creation in such sectors can offset or mitigate the impacts of job losses in automated industries.
Automation technologies such as industrial robots, machine learning and artificial intelligence (AI) are transforming the global economy at an unprecedented pace. Automation generally refers to the use of tools or technologies to perform tasks with minimal human intervention; usually, it aims to increase productivity, consistency and speed while reducing costs. As such, it is shaping the future of work by altering job tasks, labour productivity, employment patterns, job availability and the skill sets required for the workforce of tomorrow. As automation technologies continue to advance, several traditionally human-performed tasks may disappear while new tasks will certainly emerge. This could lead to job losses in certain sectors while simultaneously creating new job opportunities in others.
Understanding the impacts of automation on workers and jobs is essential for countries to navigate the changing labour landscape. Unless countries proactively respond and adapt to the unprecedented acceleration in technological advancement and implementation, opportunities may be lost. In turn, many workers may find themselves unprepared for the consequences. This is particularly true for emerging market and developing economies (EMDEs), where informal employment dominates.
Research increasingly shows that automation technologies affect employment in many complex ways. Technological change brings risks and opportunities for workers, many of which vary across different types of automation technologies, skill levels and time horizons. Often left out of the discussion, however, is how automation impacts the vast majority of workers who operate in the informal economy. To date, assessing the implications of automation technologies for informal workers in developing countries has proven challenging, particularly because harmonised employment and occupation data for informal workers across countries are lacking and research on informal workers remains scarce. This gap is important for at least two reasons. First, the penetration of automation technologies is expected to grow rapidly over the next decades, especially in developing countries in which economic structures and labour markets often depend on low-skill and routine work. Second, the opportunities and challenges for workers to adapt to technological change, use industrial robots and AI, and update their skill sets for new forms of production differ significantly in the formal and informal economies.
This chapter aims to fill this knowledge gap by exploring the impacts of automation technologies on informal economy workers in developing countries. Using the OECD Development Centre’s Key Indicators of Informality based on Individuals and their Household (KIIbIH) database, the chapter probes the following key questions: How may technological change affect informal work in developing countries? What are the employment prospects for formal and informal workers? How exposed and at risk are their jobs? Which population groups and economic sectors are most affected? How can policies mitigate the adverse effects of automation and maximise some of its benefits?
To investigate these questions, the chapter is structured as follows. The first section examines the impacts of automation technologies from a global perspective. The second section assesses the impacts of automation technologies on jobs, based on the risks of occupations to be technically replaceable, with a special focus on informal economy workers. It also identifies the groups of workers more exposed to traditional and AI automation and investigates the welfare implications for households, performing sensitivity analysis to test the robustness of the findings. Finally, the chapter concludes by summarising the main findings and discussing policy implications.
In recent decades, the world has been transformed by a rapid expansion of automation technologies. Since 2000, robot adoption has been rapid. In parallel, the share of mobile phone and internet users globally has increased tremendously. More recently, the spread of AI has been impressive, especially in the most advanced countries. These technologies are affecting both labour markets and the nature of work by making certain occupations obsolete, creating new ones, and changing the task composition of jobs and their skills requirements. Automation technologies also lead to new forms of work and employment, as evidenced by the emergence of platform work.
In practice, automation involves replacing human labour with technology for specific tasks, ranging from mechanised equipment in factories to software that performs office processes (Acemoglu, 2021[1]). In recent years, the employment impacts of these technologies have been widely analysed, especially in the context of advanced countries. Research findings suggest that the links between automation and employment are complex and largely depend on how such impacts are measured, the type of automation technologies being assessed, and the context in which these technologies operate (Filippi, Bannò and Trento, 2023[2]; Acemoglu and Restrepo, 2020[3]).
Different ways to measure the impacts of automation technologies on employment
To measure the impacts of automation technologies on employment, different approaches are usually adopted, with four complementary strands of the literature that build on one another. First are studies that develop measures of technical exposure at the occupation or task level, thereby identifying which tasks or occupations are potentially affected by automation. Second are studies that use these exposure measures to classify occupations into exposure categories and link them to labour force, household or administrative data. These findings can be used to estimate which workers, sectors, demographic groups and countries are most exposed. Third are studies that examine the actual adoption and use of automation technologies in workplaces, documenting how they are implemented across firms as well as which industries and occupations are impacted. Finally, an emerging body of research seeks to estimate the economy-wide effects of automation by analysing how exposure translates into realised outcomes via five interlinked channels: technology adoption, organisational change, productivity effects, labour reallocation and general equilibrium adjustments.
Occupations and tasks that are most exposed are used as the basis to derive a probability that a job is likely to be automated
When assessing the probability of automation, two approaches can be applied. The occupation-based approach identifies entire occupations that can be automated. The task-based approach investigates which specific tasks within an occupation can be automated. The risk of automation is then obtained by measuring a job’s probability of being automated based on the specific content of its tasks or occupation. Compared with the occupation-based approach, task-based measures offer finer granularity and recognise that workers in the same occupation may perform very different mixes of tasks. This helps provide a more detailed picture of which jobs – and which aspects of those jobs – are most exposed to technological change.
Analysis of automation risk began with a landmark study by Frey and Osborne (2017[4]) that considers a sample of occupations. Using expert evaluations on a sample of jobs in the United States (US), the authors classified occupations based on their likelihood of being automated “by means of computer-controlled equipment”. They then extended the expert judgments to the overall distribution of US occupations. Their headline finding was striking: 47% of US jobs were at high risk of automation (defined as a probability above 70%) over an unspecified timeframe, possibly one to two decades. This estimate triggered widespread concern and inspired similar studies across other countries, often reporting comparably high figures.
Subsequent research refined this approach by emphasising that job automation risk depends not only on occupations but on the specific tasks workers perform. For example, a retail salesperson carries out many tasks (e.g. operating a cash register or restocking inventory) that might be automated, while other tasks (e.g. engaging with customers) may not be. The task-level approach makes it possible to distinguish occupations that can be fully automated from those more likely to experience partial automation of certain activities. Such research has significantly advanced the understanding of how technological change affects employment and wages (Quintini, 2024[5]).
This task-level approach, first introduced by Autor, Levy and Murnane (2003[6]), assesses the task-content of occupations across two dimensions: routine vs. non-routine and cognitive vs non-cognitive. In recent decades, literature on how automation affects employment has focused on the task-content of occupations, characterising tasks carried out in occupations according to whether or not they are automatable. Arntz, Gregory and Zierahn (2016[7]) combined Frey and Osborne’s expert assessments with data from the OECD’s Survey of Adult Skills (PIAAC), which provides detailed information on tasks carried out at work. Their findings – that only 9% of jobs in the United States were deemed at high risk of automation – were far less alarming than Frey and Osborne (2017[4]). Nedelkoska and Quintini (2018[8]) extended this task-based approach to 32 OECD countries, estimating that 14% of jobs faced high risk, while 32% were likely to undergo significant changes in task composition. These results underscore the importance of accounting for task variation within occupations, which can dramatically alter risk estimates compared with occupation-level analyses.
Other studies explore alternative ways to measure automation potential. Mihaylov and Tijdens (2019[9]) examined 3 264 occupation-specific tasks to assess “routine task intensity” across 427 occupations listed in the International Standard Classification of Occupations 2008 (ISCO-08). They concluded that 16% of occupations were highly automatable, as most of their tasks were routine in nature. This task-based perspective highlights that automation risk is not uniform across jobs and depends heavily on the nature of work performed.
Beyond the impacts of past technology adoption on the probability of automation, another main concern is the potential impacts of current and future technological developments. Rapid advances in AI, robotics and digital technologies have renewed concerns that an increasing range of tasks may now be automated. These developments have motivated efforts to identify and apply indicators that capture occupational tasks’ exposure to automation as a means to measure the extent of possible labour-market disruption.
High exposure can lead to automation; but it can also foster human-machine complementarity and productivity gains. Webb (2019[10]) proposed a method to estimate exposure by comparing terms in AI patent titles with occupational task descriptions, using the degree of overlap to create an exposure score. Felten, Raj and Seamans (2021[11]) adopted a different approach, mapping AI capabilities to job-related skills using expert input and O*NET data to assess how critical those skills are to each occupation. An OECD study by Lassébie and Quintini (2022[12]) surveyed experts on the automatability of around 100 skills across technologies such as robotics and AI. Their findings suggest that while high-skilled jobs are often more exposed to AI, they are less likely to be fully automated. Similarly, low-skilled jobs rarely disappear entirely, although specific tasks within them may change substantially.
Recent developments in generative AI and large language models (LLMs) are expanding the technological frontier of automation. While previous technologies automated primarily routine tasks and did not lead to significant losses in labour demand, generative AI is substantively different (OECD, 2019[13]). It is a machine‑based system that can, for a given set of human-defined objectives, generate media objects, codes, make predictions, recommendations or decisions influencing real or virtual environments (OECD, 2019[14]). The current technological frontier of generative AI is multimodal, tool-using and increasingly using agentic foundation models. Agentic models are advanced AI systems trained or fine-tuned to act autonomously, undertake multi-step plans, use external tools and execute complex workflows. Unlike traditional models that only respond to prompts, these agentic models reason, adapt to feedback and interact dynamically with software environments. Current systems can understand and generate text, code, images, audio and video. They can also operate over longer contexts and now show rapid performance gains on advanced reasoning, programming and media-generation benchmarks (OECD, 2023[15]; Maslej et al., 2025[16]).
In practice, these models are used to draft, summarise, search and translate information. They power chatbots and virtual assistants, and can write and debug code. In turn, they are able to support customer operations, marketing and sales, and software engineering. They can efficiently carry out some elements of both research and development (R&D) and administrative work. The economic activities most exposed to related impacts are therefore knowledge- and information-intensive sectors, notably banking/finance, high-tech and information and communication technologies (ICTs), life sciences, retail and other business-service activities. OECD evidence also shows that, because generative AI performs many cognitive, non-routine tasks, exposure is especially high in urban, high-skilled regional economies (OECD, 2024[17]). Previously, it was widely believed that humans had a comparative advantage over machines in these sorts of complex tasks. Generative AI may reverse this paradigm and render these tasks more prone to automation (Agrawal, Gans and Goldfarb, 2019[18]). Nordhaus (2021[19]) even suggests that AI may potentially “increase [its] productivity and breadth to the extent that human labour and intelligence will become superfluous.”
Recent contributions from international organisations have further refined understanding of AI-related labour market exposure. The International Labour Organization (ILO) introduced a granular, task-level index to measure exposure to generative AI across countries (Gmyrek et al., 2025[20]). It finds that roughly one-quarter of global employment is in occupations with some degree of exposure, but only a small fraction – around 3% – falls into the highest exposure category in which most tasks could, theoretically, be automated. The ILO emphasises transformation, rather than displacement, as the dominant trend.
Similarly, the International Monetary Fund (IMF) examined exposure patterns in advanced and emerging economies using worker-level data (Pizzinelli et al., 2023[21]; Cazzaniga et al., 2024[22]). Distinguishing between AI acting as a complement versus a substitute for labour can significantly affect displacement risk. Advanced economies show higher exposure overall, largely due to their concentration of professional and managerial roles. At the same time, these roles often exhibit strong complementarity with AI, reducing the likelihood of job loss. The IMF also finds that women and highly educated workers tend to face greater exposure, but notes that exposure does not equate to unemployment risk. Rather, it signals potential for task transformation and productivity gains.
Net impacts of automation on employment are measured as the combination of displacement, productivity and reinstatement effects
Acemoglu and Restrepo (2018[23]) suggest that the impacts of technology on employment can be organised in three effects: displacement, productivity and reinstatement. Automation may displace workers from performing certain tasks (displacement effect). Productivity gains resulting from new technologies may increase economic output and incomes, leading to more consumption and potentially for more demand for labour in non-automated tasks (productivity effect). Automation may also spur the development of new industries and new tasks (or more complex versions of existing tasks) in which labour has a comparative advantage (reinstatement effect).
As tasks are automated, workers can shift to more human-centred tasks. Typically, robots can do basic assembly while humans move into monitoring, co-ordination, creative design or complex problem-solving roles. This task extension can improve job quality, as workers spend more time on the higher-skilled and interpersonal or creative aspects of work that machines cannot do. This work can be more fulfilling and command higher wages. This transformation underlines that the opportunity of automation lies in human-machine collaboration – i.e. letting machines do what they do best (repetition, computation at scale, etc.) while humans focus on areas in which their strengths are complementary (creativity, empathy, complex judgement, etc.).
Technological change has profound and complex effects on employment
Historically, technological change has had profound and complex effects on employment. While it has consistently raised productivity and enabled long-term economic growth, it has also disrupted labour markets, displaced certain types of jobs, and reshaped the demand for skills. The nature of this impact has evolved over time and has varied significantly across countries, sectors and type of technology.
Automation may increase overall employment in the long run but causes disruption in the short term
Automation can dramatically increase productivity, thereby boosting economic output and improving employment outcomes. By taking over repetitive, time-consuming tasks, automation allows workers (and firms) to do more with less. Today, as many advanced countries face ageing populations and labour shortages in certain sectors, automation offers a way to maintain output and improve services even as labour force growth slows down (BIAC, 2025[24]). Estimates suggest that widespread adoption of existing automation technologies could increase global labour productivity growth by around 0.8 to 1.4 percentage points per year (Manyika et al., 2017[25]).
Estimates may, however, overstate the likely economic impacts as they often rely on simplified assumptions regarding technology adoption and productivity gains. According to Acemoglu (2025[26]), combining the occupational exposure measures developed by Eloundou et al. (2024[27]) with realistic assessments of task-level cost savings and the economic feasibility of AI implementation, as in Svanberg et al. (2024[28]), yields considerably more modest projections. Under these assumptions, the diffusion of AI technologies is estimated to increase total factor productivity (TFP) by only about 0.71% over a ten-year period. When accounting for tasks that remain difficult to automate or augment, the expected gain falls to approximately 0.55% of TFP, translating into additional GDP growth of roughly 0.92% over the same time horizon (Acemoglu, 2025[26]). In a more optimistic scenario, the OECD estimates that AI could add 0.4 to 1.3 percentage points annually to labour productivity growth in relatively exposed and adoption-ready economies such as the United States and United Kingdom. In other G7 countries, estimated gains are up to 50% lower because of their different sectoral structures and slower assumed adoption (Filippucci et al., 2025[29]).
Acemoglu (2025[26]) also emphasises that the long-term productivity effects of AI hinge on the nature of the new tasks and activities created alongside technological adoption. Not all emerging tasks are necessarily associated with substantial economic value creation; some may contribute little to productive capacity while others could generate socially undesirable outcomes or “public bads”. Consequently, measures of economic growth that capture only the monetary value of these activities may overestimate the extent to which AI contributes to genuine welfare improvements and sustainable productivity growth.
Arguably, automation has often served as an engine of economic growth (Acemoglu, 2021[1]). Major technological transformations – such as the mechanisation of agriculture, the rise of industrial manufacturing and, more recently, the spread of digital technologies – have led to structural changes in employment (Apicella, 2025[30]). While automation has eliminated some jobs, others have emerged in new sectors, often requiring different skills.
Fears of technological unemployment (i.e. that displaced workers cannot transition to new jobs) have been one of the most prominent anxieties related to technological progress (Mokyr, Vickers and Ziebarth, 2015[31]). Evidence from past technological changes suggests that automation has not produced long-term unemployment overall, although it has caused short-term disruptions in the labour market as workers adapted (Autor, 2015[32]). However, the economic literature has documented that labour-market adjustments in advanced economies have been shaped not only by automation but also by increased import competition. For example, using regional variation in exposure to Chinese imports, Acemoglu et al. (2016[33]) show that import penetration from the People’s Republic of China (hereafter “China”) accounted for a substantial share of US manufacturing employment losses during the 2000s and generated persistent negative effects on earnings and labour-force participation. Their findings complement the evidence of Acemoglu and Restrepo (2020[3]), showing that industrial robots exert an independent negative effect on employment and wages, even after controlling for trade exposure. These findings suggest that globalisation and automation have jointly contributed to labour-market restructuring.
Overall, declining employment in some economic sectors concurred with employment expansion in others and, in the long run, wages increased along with productivity. While past automation eventually led to new industries and job creation, short-run disruptions can be considerable for displaced workers. Regions that are heavily dependent on the economic sectors likely to become automated may see local employment decline.
Automation technologies are advancing rapidly. The development of AI and robotics has expanded the scope of potentially automatable tasks, raising concerns about broader labour market disruption. As machines and algorithms increasingly replicate tasks that are predictable and rules-based, occupations that rely heavily on such tasks face higher risks of substitution. Industrial robots can now carry out complex, repetitive operations with precision. Similarly, software algorithms can perform some cognitive and decision-making tasks that were previously thought to be the exclusive domain of humans (Filippi, Bannò and Trento, 2023[2]).
Caution is needed as some recent findings suggest that generative AI may primarily function as an augmentation technology rather than a pure replacement technology. AI appears particularly valuable in occupations in which codifiable knowledge can be transferred from experts to less-skilled workers. This contrasts with more pessimistic narratives of widespread automation and supports the view that the economic impact of AI depends heavily on organisational adoption and task design (Brynjolfsson, Li and Raymond, 2025[34]).
Surveys on the expected effects of technology on the world of work have analysed perceptions and even produced forecasts. These exercises incorporate not only technical potential, but also business sentiments and planned investments. They use qualitative economic indicators to capture the optimism or pessimism of business leaders and executives regarding the current and future state of the economy, underpinning their decisions on hiring, investment and expansion.
One recent survey of employers suggests that between 2025 and 2030, 170 million jobs (equivalent to 14% of current employment) will be created worldwide in new task areas such as AI development, engineering and care services. In parallel, 92 million jobs (around 8% of current employment) will be destroyed (WEF, 2025[35]).
Another survey finds that 81% of business leaders believe that AI and other technologies will force organisations to radically rethink skills and human resources (WEC, 2024[36]).
In 2022, the OECD conducted parallel surveys targeting workers and firms in seven countries (Austria, Canada, France, Germany, Ireland, the United Kingdom and the United States) on the impacts of AI on the labour market (Lane, Williams and Broecke, 2023[37]). The results suggest that while both workers and their employers are generally very positive about the impacts of AI on performance and working conditions, workers have concerns about the impacts of AI on job stability.
Automation can contribute to job polarisation and the hollowing out of the middle class
While technology can create jobs, it can also lead to job displacement and exacerbate inequality for certain groups of workers. Automation and digitalisation may replace certain types of work, leading to unemployment and income loss (Carbonero, Ernst and Weber, 2018[38]). Automation can be applied to both routine manual tasks (such as assembly line work or data entry) and, increasingly, to non-routine tasks that require adaptive or cognitive skills (Arntz, Gregory and Zierahn, 2016[7]; Frey and Osborne, 2017[4]). An overarching challenge is that the benefits of technology are often unevenly distributed, with those lacking digital literacy or access to technology being left behind (OECD, 2025[39]). The digital divide remains a significant challenge, particularly in developing countries where high costs, lack of infrastructure and low digital literacy are key factors. This is crucial as limited access to technology and the internet can prevent vulnerable workers from fully benefiting from technological advancements (World Bank, 2016[40]; Gmyrek, Viollaz and Winkler, 2026[41]; Gmyrek, Winkler and Garganta, 2024[42]).
Empirical research in developed countries suggests that automation accounted for a significant share of the decline in demand for routine manual and routine cognitive jobs in recent decades. In the United States, between 50% and 70% of the change in the wage structure since 1980 – particularly the erosion of jobs in occupations based on skilled in routine tasks – can be attributed to the adoption of automation technologies (Acemoglu and Restrepo, 2022[43]). Similarly, in Europe, increased use of industrial robots from 2006 to 2018 has been linked to lower wages and employment in regions with many automatable jobs (Doorley et al., 2023[44]).
These findings reflect a pattern of job polarisation observed in many advanced economies. Automation has hollowed out mid-skill, routine occupations, while employment has grown at the high-skill and low-skill ends of the spectrum (Autor, 2015[32]). High-skill professionals and low-paid service workers have increased in share, but middle-skill factory, office and administrative roles have declined. Ultimately, these trends have contributed to wage inequality and a “squeeze” on the middle class (OECD, 2019[45]).
The rise in digital platforms and so-called “click workers” are other by-products of technological advances
Digital platforms (also referred to as online platforms) are a by-product of technological innovation that has been growing rapidly in recent years (OECD, 2023[46]). These online entities cover a broad range of activities that use digital technologies to connect the demand and supply of particular services and products (OECD, 2023[46]; ILO, 2022[47]). While the rise of digital platforms has produced numerous opportunities and benefits for workers around the world, it has also raised important concerns about job creation, working conditions and social protection.
One important aspect is the extent to which digital platforms create new jobs or mainly reorganise existing forms of work. Often, digital platforms use technology to mediate work and help outsource services, rather than to create new jobs. Many jobs that existed before the emergence of these platforms have changed in nature (Schwellnus et al., 2019[48]; Graham, Hjorth and Lehdonvirta, 2017[49]). This is especially true for location-based platforms, which have reorganised work that already existed in particular sectors (ILO, 2022[47]), as well as some online web-based platforms serving local markets.
These changes raise major questions about the legal classification of digital platform workers, their working conditions and the protection of their rights (Box 4.2). Platform work can widen access to income-generating opportunities, particularly for young people, women and workers outside major urban centres. The World Bank estimates that online gig work may account for up to 12% of the global labour force and that low- and middle-income countries generate around 40% of traffic to gig-work platforms. However, this rise of digital platforms is introducing new dimensions to informal employment. Many platform workers face uncertain earnings, limited career progression, low remuneration, unpaid time spent searching for tasks, exposure to rejection or non-payment, and limited channels to contest platform decisions. Since they are often classified as self-employed, they may also be excluded from employment-based social protections, including health insurance, retirement benefits and unemployment insurance. These gaps can increase vulnerability and poverty (Berg, 2015[50]; Rosenblat and Stark, 2016[51]; Datta et al., 2023[52]). Digital platforms can also create power imbalances between workers and employers, leading to potential exploitation. They can also complicate taxation and regulation, as tax authorities often struggle with enforcement in the absence of well-established mechanisms to track platform-generated income.
The expansion of artificial intelligence and digital platforms in advanced economies also has implications for workers in EMDEs. Many services that appear automated rely on geographically dispersed workers, often referred to as “click workers”, who label and classify data, transcribe audio, moderate content, verify outputs and complete other microtasks needed to train or operate digital systems. This “hidden” human contribution to automation, described by Casilli (2025[53]), places workers in EMDEs within global AI production chains even where domestic AI adoption remains limited.
Although this chapter does not examine these cross-border labour arrangements in detail, they reinforce the case for policy co-operation on platform governance. Relevant priorities include transparency in algorithmic management, fair remuneration, effective dispute-resolution mechanisms, data protection and portable social protection, regardless of where workers are located (Berg et al., 2018[54]).
In June 2026, the 114th Session of the International Labour Conference adopted the Decent Work in the Platform Economy Convention (Convention No. 193), marking the first binding international labour standard specifically designed to regulate digital platform work (ILO, 2026[55]). The Convention emerged in response to the rapid expansion of the platform economy – including ride-hailing, food delivery, crowd work and other app-mediated services. It also addressed growing concerns that existing labour regulations were insufficient to protect workers operating outside of traditional employment relationships.
Convention No. 193 establishes a broad framework of rights and protections applicable to platform workers, regardless of their formal employment status. Among its key provisions are requirements for fair remuneration, occupational safety and health, social protection, and protection against violence and harassment. It also covers transparency of algorithmic management systems, data protection and privacy rights, human review of automated decisions, safeguards against unjustified account suspension or deactivation, and access to dispute resolution mechanisms. Importantly, Convention No. 193 adopts a fact-based approach to employment classification, emphasising that workers' rights should not depend solely on the contractual labels used by platforms.
Diverse factors influence the likelihood that highly automatable jobs become automated
As automation technologies continue to advance, an increasing number of work activities could, a priori, be handled by automated systems. The extent to which this occurs will depend on more than just the technology itself. What is technically feasible is not always economically profitable, and the speed and scope of automation adoption also depends on factors such as cost-effectiveness, legal permissibility, and stakeholder responses – including those of trade unions, political actors and consumers (Schlogl and Sumner, 2020[56]). The automation of tasks is also shaped by economic, political, social and cultural considerations. For instance, a task may be 95% automatable in theory, but if the required machinery is too expensive compared with local real wages, firms might not adopt it for a long time. These broader factors are often underexplored, yet they play crucial roles in determining not only if automation takes place, but how it unfolds in relation to local institutions and labour markets.
This is especially true in developing countries, in which the expected impact of generative AI is strongly moderated by historically slow rate of adoption of digital technology. Exposure measures based solely on occupational tasks tend to overstate the near-term effects of generative AI because they overlook a critical prerequisite: the levels of workers' actual access to and use of digital technologies in the workplace. Using indicators of computer use at work, Gmyrek, Winkler and Garganta (2024[42]) adjust occupational exposure estimates and find that the existing digital divide substantially constrains AI diffusion. More precisely, almost half of the jobs that could potentially benefit from AI-driven augmentation are unlikely to realise these gains because workers lack adequate access to digital technologies, with the constraint being particularly severe in lower-income countries. Consequently, while a significant share of occupations in developing countries have technical characteristics that could be replaced by AI applications, the actual pace of adoption is expected to be considerably slower than in advanced economies. This anticipated lag implies more gradual productivity and labour-market effects.
Workers need a minimum level of foundational technical skills to fully reap the benefits of automation technologies (Autor, 2024[57]), which are generally lacking in developing countries (OECD, 2024[58]). Statistical evidence from experimental studies with generative AI tend to focus on very specialised groups of workers who are engaged in sets of very complex tasks. Although empirical evidence suggests that AI may reduce skills inequality within these groups, it should not be directly linked to the whole set of workers, especially in developing economies (Brynjolfsson, Li and Raymond, 2025[34]; Dell’Acqua et al., 2026[59]; Noy and Zhang, 2023[60]).
Finally, increased productivity may bring employment and wage gains in sectors facing a rapidly growing consumer demand (Autor, 2024[57]). For sectors facing more stable consumer demand, these demand-side effects are likely to differ across countries. In developing economies with a large informal sector, and in which technology adoption and private sector investment are typically concentrated among a small share of formal firms (Cirera, Comin and Cruz, 2022[61]), formal workers displaced from jobs may face more challenges finding high quality jobs than their counterparts in high-income countries.
Automation risks refer to a wide range of technological change, including traditional and AI-related automation (see Annex 4.B for more details on automation measurement).
Growing evidence, mostly from advanced countries, suggests that automation technologies affect employment in complex ways. The impacts of automation technologies vary across countries and depend on the type of technology involved. Within countries, the impacts on workers are uneven.
What remains unclear is how automation will affect the vast majority of workers that operate in the informal economy in developing countries. While technological change brings risks and opportunities for informal workers and labour formalisation, there has been only limited assessment of the implications of automation technologies for informal workers in developing countries (Gmyrek, Winkler and Garganta, 2024[42]).
Methodology
The risk of job automation can be assessed using a wide range of methodologies. The chapter analyses two types of technical exposure (or susceptibility): traditional automation and AI automation (Acemoglu and Restrepo, 2019[62]; OECD, 2023[46]).
Traditional automation refers to the extent that specific tasks required to perform a job are technically fully substitutable by the waves of automation that occurred up to the early 2020s. It generally includes industrial robot innovations and early AI digital technologies based on data computations, predictions and technical assistance in decision making.
AI automation focuses on the frontier of generative AI technologies and explicitly considers the expanding capabilities of these systems beyond text and LLMs, including voice, image and video generation. These capabilities increase the potential exposure of some occupations, especially in media- and web-related work.
To measure the traditional automation risk, this chapter follows the methodology of Lassébie and Quintini (2022[12]), which was replicated in the OECD Employment Outlook (2023[46]). The O*NET database, produced by the National Center for O*NET Development, is used to extract occupation-specific information on the skills and abilities required for each job. This information is then combined with expert assessments of the automatability of around 100 skills and abilities and ranked across different scores (O*NET Resource Center, 2026[63]). The O*NET-based information makes it possible to compute, for each occupation, the share of skills and abilities that are highly automatable and the share that remain bottlenecks to automation, focusing exclusively on skills and abilities deemed essentials to the occupation. The resulting occupation-level indicators are then used to assess which jobs are most exposed to automation and which groups of workers are most at risk, with the latter being those having a job for which at least 25% of tasks can be technically automated (based on the technical judgments of experts in technological innovation).
To measure the AI automation risk, the chapter follows the statistical approach of Gmyrek et al. (2025[20]). The risk of AI displacement is measured at the task level within each occupation code, combining task-level information for nearly 30 000 tasks (using detailed ISCO-08 job descriptions). In parallel, human expert input and AI-based predictions estimate the extent to which generative AI can perform the tasks that make up each occupation. Each occupation is then classified into four exposure gradients based on its overall average task exposure and the degree of exposure variability across the task bundle. The last gradient includes occupations at high risk of AI technical displacement (i.e. occupations with both high average task exposure, similar to Lassébie and Quintini (2022[12]), and low exposure variability across tasks, meaning that exposure is high across most tasks performed in the occupation).
Both methods use a granular task-based approach, reflecting the idea that task automation does not automatically translate into full job displacement. Importantly, some important tasks may remain non-automatable or require substantial human involvement.
Data used for the estimates
The last update of the KIIbIH is used to analyse the actual potential risk of displacement (OECD, 2026[64]). These surveys are harmonised consistently across countries, making it possible to construct standardised indicators of demographic, labour market and socio-economic conditions of individuals and their families (see Chapter 2 for more details on the KIIbIH).
Information on the risk of automation is matched with data on occupations of workers in the KIIbIH database, using the ISCO-08 international standard either at three-digit (for the risk of traditional automation) or four-digit level (for the risk of AI automation) (ILO, 2012[65]). The variables on occupations have been standardised across countries using either official correspondence tables to reclassify national occupations into ISCO-08 equivalents, or an AI-based classification procedure that matches national and international occupational codes based on detailed job descriptions (see Annex 4.B for more details).
Regarding the definition of informality, the KIIbIH follows ILO guidelines and statistical concepts defined by the 20th International Conference of Labour Statisticians (ILO, 2018[66]) (see Annex 2.B in Chapter 2 for more details). Analysis in this chapter focuses primarily on the employment classification of informality (i.e. the formal or informal status of the job or work activity itself), rather than sector classification, which focuses on the type of economic unit in which production takes place.
In developing countries, informal labour markets dominate and often depend on manual and low-skilled work
In developing countries, in which informality dominates and economic structures and labour markets often depend on manual and low-skilled work (especially in the agricultural sector), the disruptive potential of traditional automation on employment may be particularly pronounced.
In fact, tasks performed in a particular job are the principal driver of exposure to the risks of traditional or AI automation (Acemoglu and Restrepo, 2020[67]; Acemoglu and Restrepo, 2019[62]). These tasks can be broadly categorised along two main axes: manual versus cognitive tasks, and routine versus non-routine tasks. Research findings suggest that manual routine and non-routine tasks are much more exposed to traditional automation. In contrast, cognitive tasks, which include routine tasks such as prediction, classification, retrieval and summarisation (Agrawal, Gans and Goldfarb, 2019[18]; Eoundou et al., 2024[27]), and non-routine tasks, such as content generation and editing, code generation, knowledge management and process orchestration may be more exposed to AI automation (Noy and Zhang, 2023[60]; Almorsi, Ahmed and Gomaa, 2025[68]; OpenAI, 2025[69]; Cao, Ma and Zhai, 2024[70]).
In developing countries, among workers in general and informal workers in particular, manual tasks – whether routine or non-routine – are the most common. On average, across the sample of KIIbIH countries for which data are available, 77% of informal workers perform routine or non-routine manual tasks in their job, compared with 42% for formal workers (Figure 4.1). In contrast, the share of workers performing cognitive tasks is as low as 23% among informal workers and as high as 58% for formal workers.
Large country disparities exist in the relative importance of manual versus cognitive tasks performed by informal workers. The proportion of informal workers holding jobs that involve non-routine or routine manual tasks is the highest in the sample of African countries (90% on average), followed by countries from Latin America and the Caribbean and Asia (75% for both), and Europe (44%).
Distribution of informal and formal workers aged above 15 years by occupational task group, latest year available
Note: Routine cognitive refers to jobs requiring codifiable and repetitive information-processing tasks. Non-routine cognitive analytical refers to jobs requiring creative, problem-solving tasks. Non-routine cognitive personal refers to jobs requiring interpersonal tasks requiring judgement. Routine manual refers to jobs with repetitive physical tasks in structured, machine-paced environments. Non-routine manual refers to jobs requiring physical tasks in changing environments requiring dexterity, adaptation or spatial judgement. KIIbIH and regional averages are calculated as simple unweighted averages of countries for which data are available.
Source: Estimates based on OECD (2026[64]) and Lewandowski et al. (2020[71]).
The level of skills required for the job is also an important predictor of the risk of automation. Using the occupational structure of the labour market and based on the ISCO-08 classification that associates a skill level to each occupation, it is possible to estimate the skill content of each job (ILO, 2023[72]; ILO, 2012[65]).
Across KIIbIH countries, informal workers are disproportionately holding low- and medium-skilled jobs: 17% of informal workers perform a low-skilled job and 71% hold a medium-skilled job, compared with, respectively, 9% and 49% for formal workers (Figure 4.2, Panels A and B). The shares of informal workers in low- and medium-skilled jobs are remarkably high across all regions, notably across African countries for which data are available (7% of them are low- and 90% are medium-skilled).
Distribution of informal and formal workers aged above 15 years by occupational skill level, latest year available
Note: Skill levels are based on ISCO-08 major groups: “Low skill (level 1)” refers to workers employed in elementary occupations (Group 9 of ISCO-08). “Medium skill (level 2)” refers to: clerical workers; sales and services workers; skilled agricultural, craft and trades workers, and plant and machine operators and assemblers (Groups 4, 5, 6, 7, 8). “High skill (level 3)” refers to technicians and associate professionals (Group 3) as well as managers (Group 1) from sub-major group 14. “High skill (level 4)” refers to professionals (Group 2) as well as managers (Group 1) not otherwise classified in sub-major group 14. Armed forces (Group 0) are classified using 2-digit ISCO-08 codes: sub-major group 03 is assigned to “Low skill (level 1)”, sub-major group 02 to “Medium skill (level 2)”, and sub-major group 01 to “High skill (level 4)”. KIIbIH and regional averages are calculated as simple unweighted averages of countries for which data are available.
Source: (OECD, 2026[64]), “Key Indicators of Informality based on Individuals and their Households (KIIbIH)”, https://data-explorer.oecd.org.
Understanding the distribution of workers across industries is also important to examine automation risks. Different industries require indeed distinct bundles of tasks and skills, which influences their susceptibility to automation (Acemoglu and Restrepo, 2020[67]; Acemoglu and Restrepo, 2019[62]). On average across the KIIbIH, informal workers disproportionately (38%) work in the primary sector2 compared with formal workers (9%) (Figure 4.3, Panels A and B). The prevalence of informal workers in the primary sector is particularly strong in several countries in Africa and Asia, but less so in countries from Latin America and the Caribbean. A large share of informal workers (44%) is also engaged in the tertiary sector, in which formal workers are also heavily engaged (75%). In contrast, relatively few informal workers (17%) work in the secondary sector, a share equal to that for formal workers.
Distribution of informal and formal workers aged above 15 years by broad sector, latest year available
Note: Broad sectors are defined based on ISIC Rev. 4 sectors. Primary activities refer to the direct extraction or use of natural resources, such as agriculture, forestry, fishing, mining and quarrying. It includes ISIC Rev. 4 sectors A and B. Secondary activities refer to the transformation of raw materials and the production of goods, including manufacturing, utilities and construction. It includes ISIC Rev. 4 sectors C to F. Tertiary activities refer to services, including trade, transport, accommodation, finance, public administration, education, health and other service activities. It includes ISIC Rev. 4 sectors G to U.
Source: (OECD, 2026[64]), “Key Indicators of Informality based on Individuals and their Households (KIIbIH)”, https://data-explorer.oecd.org.
Informal workers face considerable risks of traditional automation; AI automation is more of a risk for formal workers in more advanced countries
Data from the KIIbIH show that, overall, one-third of workers hold a job with considerable traditional automation risks (Figure 4.4). This is slightly more than the OECD average of about one-quarter (OECD, 2023[46]). Informal workers (35%) are almost twice as likely to be at risk of automation as formal workers (19%). The risk of traditional automation varies across regions. Informal workers in African and Asian countries are the most exposed (41%), followed by those from Latin American and Caribbean (30%). Among European countries, the exposure of informal workers is the lowest (17%), but is the highest across all regions for formal workers.
By and large, the greater exposure to traditional automation risks for jobs held by informal workers from developing countries mirrors the differences in occupations between economies at different stages of development. Since less advanced economies employ more people in occupations involving routine, manual tasks – and fewer people in cognitive tasks – they are more vulnerable than advanced countries to job displacement by industrial robots.
Share of informal and formal workers aged above 15 years in jobs at high risk of traditional automation, latest year available
Note: Jobs are considered to be at high risk of traditional automation if they have a significant share of important skills and abilities (more than 25%) that are highly automatable, following Lassébie and Quintini (2022[12]).
Source: Estimates based on OECD (2026[64]) and Lassébie and Quintini (2022[12]).
By contrast, very few workers (2%) in the sample of KIIbIH countries hold a job exposed to the risks of AI automation (Figure 4.5). Such risks are even smaller for jobs of informal workers (1%) than of formal workers (5%). The risks are more concentrated among workers in Latin America and the Caribbean and Europe, reflecting the differences in the distribution of high skill and cognitive tasks across workers with different formal status and place of residence.
Share of informal and formal workers aged above 15 years in jobs at high risk of AI automation, latest year available
Note: Jobs are considered to be at high risk of AI automation if they fall into the highest of four exposure gradients, following Gmyrek et al. (2025[20]). This gradient includes occupations with both high average task exposure and low exposure variability across tasks, meaning that exposure is high across most tasks performed in the occupation. Axis scale is adapted for clarity and ranges from 0% to 14%.
Source: Estimates based on OECD (2026[64]) and Gmyrek et al. (2025[20]).
Of particular concern is the magnitude of the possible negative disruption effects of traditional automation on informal workers, who usually have fewer means to adjust to labour market shocks and protect their livelihoods. Indeed, as compared with workers in formal jobs, informal workers have higher odds of being undereducated and/or unskilled, are more likely to live in poor or near-poor households, and have a low probability of transitioning into a formal job (OECD, 2024[73]; Aleksynska, La and Manfredi, 2023[74]; OECD, 2023[75]; Aleksynska and Kolev, 2021[76]; OECD/ILO, 2019[77]).
The risk of traditional automation is spread unequally across informal workers and disproportionately affects those in the lower tier of the informal economy
Tackling the negative job displacement effects that traditional automation may have on workers in general – and on workers in the informal economy in particular – requires in-depth understanding of which worker characteristics are linked to high automation risk. Figure 4.6 presents the share of workers aged 15 or more holding a job at high risk of traditional automation, by informality status and selected socio-demographic characteristics. Despite being based on univariate correlates, most of these results tend to be relatively robust to multivariate analysis and confirm the greater vulnerability to job displacement of informal workers in the middle and lower tier of the informal economy.
Share of informal and formal workers aged above 15 years in jobs at high risk of traditional automation by selected socio-demographic characteristics
Note: Jobs are considered to be at high risk of traditional automation if they have a significant share of important skills and abilities (more than 25%) that are highly automatable, following Lassébie and Quintini (2022[12]). Axis scale is adapted for clarity and ranges from 0% to 75%.
Source: Estimates based on OECD (2026[64]) and Lassébie and Quintini (2022[12]).
Across KIIbIH countries, the traditional automation risk is similar for jobs of informal male workers (36%) and female workers (37%) (Figure 4.6, Panel A), although patterns differ across regions, with Asian and European female workers more exposed than males. Such gender differences stem largely from the fact that informal male workers are more frequently employed in agriculture and in jobs requiring manual and routine tasks that score higher on the risks of technical replaceability. As shown in Figure 4.6, among informal workers, the risk of automation is also disproportionately higher for jobs held by younger and older workers (Panel B), workers with no or little education (Panel C), rural workers (Panel D), and workers in the primary sector (Panel E). Also affected are employees and unpaid family workers (Panel F), low- and middle-paid workers (Panel G) and workers living in poor households (Panel H).
Multivariate analysis confirms that, compared with formal workers, informal workers are at a significant disadvantage and shows that, “all else equal”, gender, age, education, earnings category and place of residence are important correlates of holding a job at high risk of automation (Annex Table 4.A.1). Similarly, informal workers with low levels of skills and middle- and upper-paid jobs are more exposed to traditional automation risk than formal counterparts with similar characteristics (Annex Table 4.A.2).
Probability of workers aged above 15 years of holding a job at high risk of traditional automation by sector of employment
Note: The chart shows estimates of the predicted probabilities of being employed in a job at high risk of traditional automation in each sector. Estimates are based on logit regressions for which the dependent variable is a dummy identifying if the worker is in a job at high risk of traditional automation. 95% confidence intervals (CI) are shown with point estimates. Controls include age and age squared, gender, educational attainment, marital status, location, formality status of workers, sector of employment (using the one-digit ISIC Rev. 4), earnings quality and status in employment. To account for differences in sample and population size across countries, probability survey weights are normalised so that each country equally contributes to the estimates. Regressions are run for the sample of countries with available data: Bolivia, Cameroon, Chile, China, the Dominican Republic, El Salvador, Honduras, Mexico, Nicaragua, Nigeria, Peru, Tanzania and Uruguay.
Source: Estimates based on OECD (2026[64]) and Lassébie and Quintini (2022[12]).
Regression results highlight the critical role of industry as a significant mediator of automation exposure in KIIbIH countries (Figure 4.7). This is consistent with the fact that sector-specific technologies and production processes shape the task content of jobs demanded by firms. Predicted exposure is particularly high in primary and secondary sectors, such as activities of households with very low level of productivity (55%), mining (49%), manufacturing (48%) and agriculture (44%). It remains elevated in service activities such as transport (40%), administrative and support service and other services (27%), water and waste management (42%) and utilities (37%). Exposure is substantially lower in health (11%), education (10%), ICT services (7%) and financial services (4%).
While AI automation may impact few informal workers, those with jobs at high risk are disproportionately female, young, urban and highly educated workers
Across KIIbIH countries, AI automation risks are disproportionately higher for jobs held by formal workers (5%) than by informal ones (1%) (Figure 4.5). This aligns with the fact that AI (especially generative applications) targets information‑processing and codifiable cognitive tasks that are more prevalent in the context of formal job structures. If those formal workers were to lose their job through the dissemination of AI technologies, they could end up in the informal economy and contribute to further enlarging informal employment.
KIIbIH data also show that the few informal workers holding a job at risk of AI automation have specific characteristics. As shown in Figure 4.8, the share of informal workers at risk of AI automation is usually higher for female (Panel A), younger (Panel B) and urban workers (Panel D). It rises markedly with the level of education (Panel C) and for workers in the secondary and tertiary sector (Panel E). It is also higher for employees (Panel F), better-paid workers (Panel G), and workers who live in non-poor households (Panel H).
Share of informal and formal workers aged above 15 years in jobs at high risk of AI automation by selected socio-demographic characteristics
Note: Jobs are considered to be at high risk of AI automation if they fall into the highest of four exposure gradients, following Gmyrek et al. (2025[20]). This gradient includes occupations with both high average task exposure and low exposure variability across tasks, meaning that exposure is high across most tasks performed in the occupation. Axis scale is adapted for clarity and ranges from 0% to 10%.
Source: OECD (2026[64]) and Gmyrek et al. (2025[20]).
Based on multivariate analysis, the formal employment status of workers remains positively and significantly correlated with AI automation risks when controlling for other factors. However, a large part of this apparent formality effect tends to be mediated by differences between formal and informal workers in terms of skills, status in employment and sector-specific tasks. The sector of employment, in particular, seems to be an important factor shaping the likelihood of jobs being at high risk of AI automation (Figure 4.9). Even holding worker characteristics constant, the predicted probabilities of being at risk of AI automation vary widely across sectors, with significantly higher risks for workers who operate in the financial sector (10%), public administration (6%), and professional scientific and technical roles (4%).
Probability of workers aged above 15 years of holding a job at high risk of AI automation by sector of employment
Note: The chart shows estimates of the predicted probabilities of being employed in a job at high risk of AI automation in each sector. Estimates are based on logit regressions for which the dependent variable is a dummy identifying if the worker is in a job at high risk of AI automation. 95% confidence intervals (CI) are shown with point estimates. Controls include age and age squared, gender, educational attainment, marital status, location, formality status of workers, sector of employment (using the one-digit ISIC Rev. 4), earnings quality and status in employment. To account for differences in sample and population size across countries, probability survey weights are normalised so that each country equally contributes to the estimates. Regressions are run for the sample of countries with available data: Chile, China, the Dominican Republic, El Salvador, Honduras, Liberia, Maldives, Peru, and Tanzania. Axis scale is adapted for clarity and ranges from 0% to14%.
Source: OECD (2026[64]) and Gmyrek et al. (2025[20]).
Automation risks are amplified at the household level, especially for individuals who live in informal households
Automation risks can affect livelihoods through job displacement and earnings disruption, at both the individual and household level. Factors such as a household’s incidence of informality and whether some or all workers of a household hold jobs exposed to automation risks will ultimately determine the overall level of vulnerability of workers and their families to income shocks. This makes it essential to assess, separately, the overall level of exposure to automation risks for formal, mixed and informal households.
As regards traditional automation, KIIbIH data point to substantial heterogeneity across households while showing that such risks are high and disproportionately concentrated among informal households (Figure 4.10). Among households in which all workers are informal, 47% of members live in a household in which at least one worker is employed in a job exposed to traditional automation. More precisely, 23% live in a household where some working members are exposed and others are not, while 24% live in a household where all are exposed. By contrast, in mixed households (where formal and informal workers co-habit), 44% of members live with at least one exposed worker (39% with a mix of exposed and non-exposed workers, and 5% with all workers exposed). In formal households, exposure is less common overall: 18% of members live in a household in which at least one working member is exposed (9% with a mix of exposed and non-exposed workers, and 9% with all workers exposed). Across regions, Africa (30%) and Asia (28%) exhibit the largest share of people who live in informal households in which all working members are exposed to traditional automation risks.
Distribution of the population by level of household informality and number of household members exposed to traditional automation risk
Note: Jobs are considered to be at high risk of traditional automation if they have a significant share of important skills and abilities (more than 25%) that are highly automatable, following Lassébie and Quintini (2022[12]).The population of reference are people living in households with at least one worker (household with no workers are excluded). “Informal” refers to a household where all workers are informal workers; “mixed” refers to a household where at least one worker is an informal worker and one worker is a formal worker; and “formal” refers to a household where all workers are formal workers. KIIbIH and regional averages are calculated as simple unweighted averages of countries for which data are available: the Bahamas, Benin, Bolivia, Brazil, Cameroon, Chile, China, Colombia, the Dominican Republic, El Salvador, Gambia, Ghana, Honduras, Jamaica, Lao PDR, Madagascar, Maldives, Mali, Mexico, Nicaragua, Peru, Senegal, Sierra Leone, Togo, Uganda, Uruguay and Zambia.
Source: OECD (2026[64]) and Lassébie and Quintini (2022[12]).
For AI automation, KIIbIH data show an overall low level of household exposure, especially in less advanced regions, and that most of the exposed workers live in formal and mixed households (Figure 4.11). Among formal households, 7% of people live in households with some (5%) or all (2%) working members exposed, compared to 7% in mixed households and 1% in informal households.
Distribution of the population by level of household informality and number of household members exposed to AI automation risk
Note: Jobs are considered to be at high risk of AI automation if they fall into the highest of four exposure gradients, following Gmyrek et al. (2025[20]). This gradient includes occupations with both high average task exposure and low exposure variability across tasks, meaning that exposure is high across most tasks performed in the occupation. The population of reference are people living in households with at least one worker (household with no workers are excluded). “Informal” refers to a household where all workers are informal workers; “mixed” refers to a household where at least one worker is an informal worker and one worker is a formal worker; and “formal” refers to a household where all workers are formal workers. KIIbIH and regional averages are calculated as simple unweighted averages of countries for which data are available: the Bahamas, Benin, Bolivia, Brazil, Cameroon, Chile, China, Colombia, the Dominican Republic, El Salvador, Gambia, Ghana, Jamaica, Honduras, Lao PDR, Madagascar, Maldives, Mali, Mexico, Nicaragua, Peru, Senegal, Sierra Leone, Togo, Uganda, Uruguay and Zambia.
Source: OECD (2026[64]) and Gmyrek et al. (2025[20]).
Among informal workers exposed to high risk of traditional automation, substantial shares live in poor- (20% of the total) or lower-welfare households (20%). This exposes their material welfare to income shocks in case of a technology-related dismissal or worsening labour market conditions (Figure 4.12, Panel A). Most (48%) live in middle-welfare households, while relatively few (12%) live in upper-welfare households. This fact especially holds for Asian countries and those in Latin America and the Caribbean. In contrast, formal workers facing high risks of traditional automation are more likely to live in middle- (53%) and upper-welfare (26%) households (Figure 4.12, Panel B). If workers were displaced by technology, the resulting loss of income could weaken their families’ material welfare and push some middle-welfare households towards lower-income or poorer conditions. This would mirror the hollowing out of the middle class already observed in many OECD economies (OECD, 2023[46]; OECD, 2019[45]). This risk appears substantial in Latin America and the Caribbean, where most formal workers exposed to traditional automation are part of the middle-welfare category (60% of the total).
Distribution of informal and formal workers aged above 15 years at high risk of traditional (Panels A and B) and AI (Panels C and D) automation by welfare group, latest year available
Note: The population of reference are people living in households with at least one worker (household with no workers are excluded). Jobs are considered to be at high risk of traditional automation if they have a significant share of important skills and abilities (more than 25%) that are highly automatable, following Lassébie and Quintini (2022[12]). Jobs are considered to be at high risk of AI automation if they fall into the highest of four exposure gradients, following Gmyrek et al. (2025[20]). This gradient includes occupations with both high average task exposure and low exposure variability across tasks, meaning that exposure is high across most tasks performed in the occupation. Welfare classes are defined relative to the household median welfare (income or consumption, depending on the country and survey) observed in each country, using the categories and underlying concepts developed by OECD (2019[45]). Poor: below 50% of median welfare; lower welfare (or “vulnerable”): between 50% and 75% of median welfare; middle welfare (or “middle class”): 75% and 200% of median welfare; upper welfare (or “affluent”): above 200% of median welfare. KIIbIH and regional averages are calculated as simple unweighted averages of countries for which data are available.
Source: Estimates based on OECD (2026[64]), Lassébie and Quintini (2022[12]) and Gmyrek et al. (2025[20]).
The risk of material welfare loss due to AI automation is concentrated among more affluent workers than for traditional automation. Formal workers highly exposed to AI automation mostly live in middle- (42%) or upper-welfare (52%) households. The figures are similar for at-risk informal workers (57% and 32% respectively) (Figure 4.12, Panels C and D). In reality, given the low shares of workers highly exposed to AI automation risks, the expected overall effects on income levels and distribution is likely to be small, at least in the short term. If technological innovation accelerates and new AI applications (such as agentic models) become able to automate more complex bundles of tasks than is currently possible, many workers who now appear relatively protected could face skills obsolescence (OECD, 2025[39]; Maslej et al., 2025[16]; OECD, 2023[46]). This could weaken their position in formal labour markets and reduce their chances of accessing better-quality jobs.
Throughout the world, technological change has always entailed risks and opportunities for workers. Traditional and AI automation technologies are no exceptions. What makes a significant difference in impacts is the environment in which these technological transitions are taking place. In contrast to developed OECD countries, governments in developing countries often face the dual challenge of high informality and limited fiscal space, which heightens the level of vulnerability of the workforce and restricts the scope for policy response. At the same time, the pace of technology adoption in developing countries may be slower and more conducive to gradual tasks restructuring rather than immediate, large job losses. This is particularly true for generative AI, for which infrastructure and organisational bottlenecks can slow deployment. Ultimately, policymakers have a responsibility to manage the spread of automation technologies. To a large extent, policy action will determine whether the opportunities brought by technological change could outweigh its challenges.
The evidence presented in this chapter shows that traditional automation, if investment and capital deepening were to happen abruptly, could become a mass risk for informal workers, who disproportionately hold low-skill jobs. For AI automation, the magnitude of the risk is more limited and largely contained among few formal and highly educated workers with more adaptive capacities. In this context, a strategy to manage the technological transition could be articulated around the following five pillars: i) invest in education and reskilling; ii) strengthen social protection systems; iii) encourage decent job creation in emerging sectors; iv) regulate automation; and v) address equity issues and profit sharing.
Investing in education and reskilling
As automation technologies create uncertainty and change skills demands, governments and firms must develop skill strategies that prioritise lifelong learning, vocational training and digital literacy to equip workers for new roles. Technical and vocational education and training (TVET) systems are particularly well-placed to respond to shifting task demands, given their direct connections to employers and their capacity to update curricula in response to emerging skill needs (OECD, 2023[78]). Skills and occupations are central channels shaping exposure to technological replacement, with sectoral composition being one of the characteristics shaping the distribution of risks. For traditional automation risk, exposure is broader and more concentrated in routine-intensive sectors. Predicted automation probabilities across broad sectors show very high risk in manufacturing relative to many services. For AI automation, secondary and tertiary educational attainment significantly increases the probability of being in a high-risk AI automation occupation – even after controlling for the industry composition (the main source of variation in risk).
These results support the need for skill development plans that are task-focused and enable workers to complement the tasks performed by new technologies, coupled with measures that facilitate mobility across sectors and occupations. Specific attention is also needed on foundational skills such as literacy and numeracy, as well as digital skills, problem-solving and interpersonal skills. All of the preceding are critical for lifelong learning. The high risk of traditional automation faced by workers in the informal economy also requires specific measures to bridge the skills gap for informal workers and facilitate their access to skills development and lifelong learning programmes. Informal workers are less likely to participate in adult education and training, which impedes skills formation and makes the jobs held by informal workers more prone to automation. Such measures could include: recognition of prior learning and the extension of formal TVET programmes to informal economy workers. Community-based and other non-formal training initiatives also represent a critical source of skills for informal workers, as does offering training incentives as part of broader formalisation policies. Apprenticeship programmes, which help build task-relevant skills by embedding structured learning in the workplace, help strengthen employer engagement and improve system responsiveness to shifting skill demands (OECD, 2018[79]).
Lifelong learning is becoming more and more relevant, amid the changes in skill needs induced by AI technologies. While versatile and transferable skills foster adaptability, the returns to these competencies vary. Digital roles may allow more employment prospects for workers, especially informal ones, but the scale of these gains differs by occupation, educational attainment and country contexts. Skills are critical for professional resilience, inclusion and opportunity, but they must be developed, recognised, valued and used in ways that reflect evolving labour markets and societal needs – across worker’s entire life cycle. Lifelong learning must be at the centre of transition strategies to tackle informality and ensure that no one is left behind. Expanding comprehensive, inclusive and forward-looking skills offers is vital for a fast-changing world of work (ILO, 2026[80]).
Strengthening social protection systems
The analysis of automation risks undertaken in this chapter highlights the importance of strengthening social protection programmes in parallel with the technological transition. The findings show that traditional automation risks disproportionately impact informal economy workers and a meaningful share of households in which all workers are exposed, implying the potential for large welfare loss if displacement occurs. By contrast, individual and household exposure to AI automation risk is much lower and affects disproportionately formal workers from middle/upper classes. This suggests that immediate welfare impacts stemming from AI automation may be smaller, though job transformation remains important.
These concerns call for the strengthening of social protection programmes with a view to establish broad, scalable social protection floors in anticipation of significant shocks linked to traditional automation and, to a lesser extent, AI automation. Such programmes need to be easily and quickly activated to protect all at-risk workers, especially informal workers. Social protection delivery modalities also need to acknowledge the limit of targeting mechanisms in high-informality contexts, where proxy-means testing and other methods can produce substantial inclusion/exclusion errors. As delivery-system quality is often as important as the targeting rule itself (Hanna and Olken, 2018[81]), three action streams appear critical: i) investing in integrated delivery systems (registries, digital ID where feasible, grievance redress, payment infrastructure, etc.); ii) progressively moving towards universal or quasi-universal floors as capacity and fiscal space allow (Sabates-Wheeler, Hurrell and Devereux, 2015[82]); and iii) using adaptive targeting (categorical elements plus community validation) to reduce inclusion and exclusion errors and improve legitimacy.
Encouraging decent job creation in emerging sectors
Policies promoting innovation, entrepreneurship and job creation in sectors with low exposure to automation risks can offset or mitigate the impacts of job losses in automated industries. As shown in this chapter, several labour‑intensive activities remain at relatively low risk of both traditional and AI automation. This includes jobs in health and social work, accommodation and food, public administration, and professional services. In this context, expanding employment in those labour‑intensive sectors – while boosting productivity through better skills, management and digital tools that augment rather than replace workers – is particularly important and should be integral to any job-rich growth strategy (Nayyar, Vorisek and Yu, 2023[83]; Rodrik and Sandhu, 2024[84]; Nayyar, Hallward-Driemeier and Davies, 2021[85]; African Development Bank, 2025[86]; García Zaballos, Iglesias Rodríguez and Puig Gabarró, 2020[87]). In parallel, promoting innovation and effective eco-systems for the clusters of high-productivity firms in AI-intense sectors remains essential to reap the benefits of innovation and support new and highly productive jobs.
Regulating automation
Automation technologies, including generative AI, can bring many benefits to workers and societies. They also present new risks that require thoughtful responses from governments, businesses and individuals. There is a strong rationale for regulating the pace and scope of automation in order to influence technology adoption speed and pervasiveness, protect vulnerable workers, and minimise potential misuse. Often, the challenge is to balance innovation incentives and the need to address potential risks. Diverse approaches have proven effective. Importantly, in a time of rapid technological change, any regulatory efforts may need to frequently evolve and incorporate mechanisms for periodic review and flexibility.
One approach is to enhance existing governance frameworks and develop new regulatory mechanisms to address emerging risks. For instance, to promote safety and avert risks to people, the EU AI Act classifies AI systems into risk categories with different degrees of requirements and obligations embedded.
Another approach is to focus on voluntary commitments from AI firms themselves and to promote co-regulation rather than purely state regulation. Industry actors are expected to develop internal governance systems, conduct risk assessments, implement safety frameworks, disclose incidents, and participate in voluntary standards and codes of conduct (OECD, 2024[88]). The implication is that the effective frontier of AI capabilities is partially determined by firms' own governance choices. Without credible self-regulation, governments may impose stronger ex-ante restrictions. Industry self-regulation serves both as a safety function, reducing harms and incidents, but also as a legitimate function, preserving public acceptance of AI deployment (OECD, 2025[89]). From a labour-market perspective, legitimacy may be as important as capability. Occupations are substituted only when society accepts delegation of tasks to AI systems. In this regard, many professional occupations remain protected not by technical bottlenecks alone, but rather by social expectations regarding responsibility, liability, and trust.
Addressing equity issues and profit sharing
If productivity gains are widely shared and workers are provided help to retrain for new tasks, automation can support inclusive growth. Conversely, if gains accrue mainly to owners of capital and a narrow group of innovators, with weak support for displaced workers, automation can widen and deepen inequality. In the absence of policy guidance, automation technologies may amplify inequality and reduce worker bargaining power. Policies that redirect technology, strengthen education and labour-market institutions, and extend social protection can tilt outcomes towards broad societal benefits (Acemoglu (2024[90]). While social dialogue can play an important role to address equity issues, union coverage and collective bargaining are often weaker in countries with pervasive informality (ILO, 2026[91]), reducing workers’ capacity to negotiate how technology is introduced (task redesign, redeployment, training commitments, and safeguards against abusive monitoring). Evidence from international social dialogue research emphasises that effective bargaining institutions can improve fairness in transitions (including digital transitions) (ILO, 2024[92]). The interventions may include strengthening or incentivising legal and practical freedoms of association, investing in dispute resolution and sectoral dialogue, and enabling representation models for informal and self-employed workers (through co-operatives, associations or hybrid social-partner structures). Such interventions help to ensure that adoption decisions incorporate job-quality safeguards and training pathways.
Equity is also an issue for many workers and their families. Traditional automation risks are higher for male, low-skilled informal workers living in rural areas. In contrast, women in formal jobs with a high level of skills and living in urban settings are more likely to hold jobs at high risk of AI automation. These findings confirm that certain risks may be concentrated in particular clusters of labour markets. For most occupations, the impacts are more likely to be felt through changes in tasks, skills and working conditions rather than widespread job losses. Strengthening social dialogue, addressing occupational segregation and ensuring women’s representation in AI-related roles will be critical to ensuring that technological change supports decent work and advances gender equality (ILO, 2026[93]).
Percentage points change in the likelihood of being employed in a job at high risk of traditional automation
Note: The table reports the marginal effects from logit regressions. The dependent variable is a dummy variable identifying workers employed in jobs at high risk of traditional automation. Robust standard errors (SE) of the estimates are shown in parentheses. Stars (*) indicate the level of statistical significance (* p<0.10; ** p<0.05; *** p<0.01). To account for differences in sample and population size across countries, probability survey weights have been normalised so that each country equally contributes to the estimates. Multiple regressions are run with distinct specifications. Specification (1) controls only for socio-demographic variables, formality status and earnings category; (2) controls also for the educational attainment (based on ISCED classification); (3) further controls for the status in employment; (4) adds sector controls, using one-digit ISIC Rev. 4 and agricultural workers as the reference group. All specifications control for country fixed effects. Regressions cover all countries with available data: Bolivia, Cameroon, Chile, China, the Dominican Republic, El Salvador, Honduras, Mexico, Nicaragua, Nigeria, Peru, Tanzania and Uruguay.
Source: Estimates based on OECD (2026[64]) and Lassébie and Quintini (2022[12]).
Percentage points change in the likelihood of being employed in a job at high risk of traditional automation
Note: The table reports the marginal effects from logit regressions. The dependent variable is a dummy variable identifying workers employed in jobs at high risk of traditional automation. Robust standard errors (SE) of the estimates are shown in parentheses. Stars (*) indicate the level of statistical significance (* p<0.10; ** p<0.05; *** p<0.01). To account for differences in sample and population size across countries, probability survey weights have been normalised so that each country equally contributes to the estimates. Regressions are run separately for formal and informal workers. To test whether the estimated marginal effects differ between the two groups, an additional pooled specification is estimated in which each regressor is interacted with informality status (results are not presented in the table). This allows the marginal effect of each variable to vary between formal and informal workers. The “Difference” columns report the statistical significance of the difference between these two marginal effects, based on a t-test of the corresponding interaction term, with robust standard errors shown in parentheses. All specifications include the same set of regressors: age and age squared (the table reports the compounded effect), gender, marital status, location, educational attainment, employment status, earnings quality and sector controls (using one-digit ISIC Rev. 4 and agricultural workers as the reference group). They also control for country fixed effects and cover all countries with available data: Bolivia, Cameroon, Chile, China, the Dominican Republic, El Salvador, Honduras, Mexico, Nicaragua, Nigeria, Peru, Tanzania and Uruguay.
Source: Estimates based on OECD (2026[64]) and Lassébie and Quintini (2022[12]).
Percentage points change in the probability of being employed in a job at high risk of AI automation
Note: The table reports the marginal effects from logit regressions. The dependent variable is a dummy variable identifying workers employed in jobs at high risk of AI automation. Robust standard errors (SE) of the estimates are shown in parentheses. Stars (*) indicate the level of statistical significance (* p<0.10; ** p<0.05; *** p<0.01). To account for differences in sample and population size across countries, probability survey weights have been normalised so that each country equally contributes to the estimates. Multiple regressions are run with distinct specifications. Specification (1) controls only for socio-demographic variables and formality status; (2) controls also for the educational attainment (based on ISCED classification); (3) further controls for the status in employment and earnings category; (4) adds sector controls, using one-digit ISIC Rev. 4 and agricultural workers as the reference group. All specifications control for country fixed effects. Regressions cover all countries with available data: Bolivia, Chile, China, the Dominican Republic, El Salvador, Honduras, Mexico, Nicaragua, Peru, Tanzania and Uruguay.
Source: Estimates based on OECD (2026[64]) and Gmyrek et al. (2025[20]).
Percentage points change in the probability of being employed in a job at high risk of AI automation
Note: The table reports the marginal effects from logit regressions. The dependent variable is a dummy variable identifying workers employed in jobs at high risk of AI automation. Robust standard errors (SE) of the estimates are shown in parentheses. Stars (*) indicate the level of statistical significance (* p<0.10; ** p<0.05; *** p<0.01). To account for differences in sample and population size across countries, probability survey weights have been normalised so that each country equally contributes to the estimates. Regressions are run separately for formal and informal workers. To test whether the estimated marginal effects differ between the two groups, an additional pooled specification is estimated in which each regressor is interacted with informality status (results are not presented in the table). This allows the marginal effect of each variable to vary between formal and informal workers. The “Difference” columns report the statistical significance of the difference between these two marginal effects, based on a t-test of the corresponding interaction term, with robust standard errors shown in parentheses. All specifications include the same set of regressors: age and age squared (the table reports the compounded effect), gender, marital status, location, educational attainment, employment status, earnings quality and sector controls (using one-digit ISIC Rev. 4 and agricultural workers as the reference group). They also control for country fixed effects and cover all countries with available data: Bolivia, Chile, China, the Dominican Republic, El Salvador, Honduras, Mexico, Nicaragua, Peru, Tanzania and Uruguay.
Source: Estimates based on OECD (2026[64]) and Gmyrek et al. (2025[20]).
This chapter measures technical exposure (or susceptibility) to automation, rather than predicting realised job losses. Technical exposure captures the extent to which existing or frontier technologies could, in principle, perform a significant share of the tasks required in a job. Actual labour market outcomes depend on additional factors such as technology adoption costs, regulatory frameworks, firm-level capabilities, and the creation of new tasks and occupations. The analysis therefore follows a task-based approach, in which occupations are understood as “bundles of tasks” that may be differentially affected by technology (OECD, 2023[46]; Acemoglu and Restrepo, 2019[62]).
Two complementary dimensions of technical exposure are considered – traditional and AI automation exposure. Both rely on a task-based approach, reflecting the concept that occupations are bundles of tasks, many of which can be automated by technologies. As the extent of the technical replaceability differs from task to task, occupations may be transformed in terms of numbers, contents and the degree of automation of the tasks.
Traditional automation exposure reflects the substitutability of tasks by earlier waves of digital technologies and robotics (up to the early 2020s).
AI automation exposure captures the expanding capabilities of recent generative AI systems, including multi-modal models capable of processing text, voice, images and video.
This dual approach makes it possible to analyse how different technological waves have effects on occupations through distinct channels. It also accounts for the fact that estimates of exposure are sensitive to methodological choices and technological scope. A key strength of this dual approach is that it captures different technological frontiers, allowing comparison between mature automation technologies and emerging AI-driven risks. It also reflects the reality that technological change is cumulative rather than discrete.
A limitation is that both measures remain “technical” rather than behavioural or economic: they do not capture firms’ adoption decisions, institutional constraints or demand-side adjustments. As a result, the estimates should be interpreted as upper-bound exposure indicators rather than forecasts of employment outcomes.
The measurement of traditional automation risk follows the approach developed by Lassébie and Quintini (2022[12]) and subsequently applied in the OECD Employment Outlook (OECD, 2023[46]).
The methodology combines two main inputs:
Occupational requirements (O*NET database): The O*NET database provides detailed information on the skills, abilities and knowledge required for each occupation in the US labour market. Each descriptor is rated according to its importance, using a scale from 1 (“not important”) to 5 (“extremely important”).
Expert assessments of automatability: Expert surveys are used to classify approximately 100 skills and abilities according to their degree of automatability, distinguishing between highly automatable items and bottlenecks to automation (items that remain difficult to automate).
Computing occupation-level exposure indicators and defining high-risk occupations
For each occupation, the analysis computes two indicators:
the share of important skills and abilities that are highly automatable, and
the share of important skills and abilities that are bottlenecks to automation.
Occupations are decomposed by skills and abilities necessary to perform them. Each skill and ability is assigned a score ranging from 1 to 5 depending on its importance to perform the related occupation, with 1 indicating “not important” and 5 indicating “extremely important”. The calculation of the share of skills and abilities that are automatable is then done by only retaining those skills and abilities deemed as essentials (importance score of 4 = “very important” or 5 = “extremely important”). Thus, the analysis focuses on core job tasks, avoiding overestimation driven by marginal activities. The decomposition into automatable and bottleneck skills also provides a nuanced view of exposure. A drawback is that the choice of the importance threshold is somewhat arbitrary, and alternative thresholds could yield different exposure estimates. Importance scores are derived from US-based data (O*NET), which may not fully reflect task structures in developing economies.
An occupation is classified as at high risk of traditional automation when at least 25% of its important skills and abilities are highly automatable. This threshold identifies occupations in which a significant share of core tasks is technically substitutable, while recognising that most occupations retain some non-automatable components. Indeed, existing evidence shows that even highly exposed occupations typically have only a minority of fully automatable tasks, implying that job transformation is more likely than complete job disappearance. Because O*NET data are based on US occupations, exposure scores are mapped to international occupational classifications (ISCO) and, using occupational codes, subsequently assigned to workers in the KIIbIH microdata. This mapping enables consistent cross-country analysis while acknowledging that task content within occupations may vary across contexts.
A major strength of this approach is that it goes deeper than occupation-level classifications. Working at the level of skills and abilities allows a granular assessment of which components of work are automatable. The approach explicitly states that occupations contain both automatable and non-automatable elements. A potential shortcoming is that the methodology relies on expert judgements, which may become outdated as technologies evolve or reflect subjective biases. In addition, the approach captures primarily pre-generative AI technologies, potentially underestimating exposure in some high-skilled or cognitive jobs affected by recent AI advances.
AI automation exposure follows Gmyrek et al. (2025[20]) and is based on:
Task-level scoring for nearly 30 000 tasks. The paper uses a comprehensive global task library, derived from occupational classifications, namely the technical documentation of ISCO‑08. This tool breaks each occupation down into hundreds of detailed tasks, similarly to Lassébie and Quintini (2022[12]). Each task is scored for its potential automation by generative AI.
A combination of human expert input and AI-assisted predictions. The approach directly assesses a subset of tasks (around 2 800 to 3 000) using worker surveys (perceived AI automation potential), expert panels and iterative validation. This labelled subset is then used to train and calibrate an AI system, which predicts automation scores for the full task universe.
Compared to earlier approaches (such as O*NET skills/abilities), this task dataset has much finer granularity, facilitating the capture of specific activities rather than broad skills. It is closer to real work processes, in which tasks are the actual units of production. The dataset is partly built from one national classification (from Poland, compatible with ISCO-08) and extended globally. This may bias the measurements in other labour markets.
Computing occupation-level exposure indicators and defining high-risk occupations
Each task in the dataset is assigned an automation score between 0 and 1. In turn, these task-level scores are aggregated to the occupational level, producing a mean exposure score across tasks as well as a measure of task variability (dispersion of exposure across tasks). Combining these two dimensions, occupations are then classified into four exposure gradients: i) low exposure (low mean exposure, high variability); ii) moderate exposure; iii) significant exposure; and iv) high exposure (high mean exposure and relatively low variability across tasks). This framework reflects the idea that automation of some tasks does not necessarily imply job displacement as other tasks may still require human input.
In this chapter, occupations are considered at high risk of AI automation when they fall into the highest exposure gradient, i.e. those with both:
high average exposure to AI across tasks, and
low exposure variability across the task bundle (i.e. exposure is high across most tasks performed in the occupation).
This definition is conceptually consistent with the threshold-based approach used for traditional automation. It enables meaningful comparison across different technologies while leveraging richer task-level information. This framework is particularly strong in capturing the idea that task heterogeneity matters: even highly exposed occupations may retain non-automatable tasks. The inclusion of variability is an important improvement over simpler averages The approach may, however, be less intuitive than threshold-based methods. Additionally, the interpretation of gradients can be less straightforward for policy implications.
All empirical results are based on the OECD Key Indicators of Informality based on Individuals and their Households (KIIbIH) database, which harmonises nationally representative household survey data across a large number of developing and emerging economies (OECD, 2026[64]).
The KIIbIH database provides, among others, detailed information on three major areas:
Individual labour market characteristics, including employment status, industry, occupation, earnings, education, demographic characteristics and location.
The formality status of workers, harmonised across countries based on survey information (e.g. social protection coverage for employees, registration/accounting practices for own-account workers) and following ILO recommendations (ILO, 2018[66]; ILO, 2013[94]).
Household composition and welfare, which enables analysis of labour market risks in a household context rather than solely at the individual level.
Because O*NET data are based on US occupations, exposure scores are first mapped to the International Standard Classification of Occupations (ISCO-08) and then assigned to workers in the KIIbIH microdata using their occupational codes. For traditional automation, the chapter uses data from Lassébie and Quintini (2022[12]), which translates O*NET-based scores into 3-digit ISCO-08 occupational groups. For countries where KIIbIH occupational data are only available at the 2-digit level, estimates are computed by applying the 3-digit distribution of high-risk occupations within each broader 2-digit group. In practice, workers in a 2-digit group are assigned the share of its 3-digit sub-occupations classified as being at high risk of traditional automation. For AI automation, the chapter uses 4-digit ISCO-08 data from Gmyrek et al. (2025[20]), mapped to KIIbIH data. When country-level occupational data are only available at the 2- or 3‑digit level, the same approach is applied: workers are assigned the average AI automation risk of the corresponding 4-digit occupations within their broader occupational group. Thus, if a 3-digit occupation contains three 4-digit occupations and only one is classified as being at high risk of AI automation, one-third of workers in that 3-digit occupation are counted as being at high risk.
The mapping is done at the level of detail of the occupational variable of the KIIbIH database. As this varies from among countries, it has been standardised with the methodology presented in OECD (forthcoming[95]). This enables consistent cross-country analysis while acknowledging that task content within occupations may vary across contexts.
A key advantage of KIIbIH is that it allows linking labour outcomes to household-level vulnerability, including the distinction between household socio-economic typologies and their welfare status. This framework enables the analysis of how exposure to automation may translate into welfare risks depending on the degree of income diversification within households.
References
[26] Acemoglu, D. (2025), “The simple macroeconomics of AI”, Economic Policy, Vol. 40/121, pp. 13-58, https://doi.org/10.1093/epolic/eiae042.
[1] Acemoglu, D. (2021), “Remaking the post-COVID world”, Finance & Development, International Monetary Fund, Washington DC, https://www.imf.org/external/pubs/ft/fandd/2021/03/COVID-inequality-and-automation-acemoglu.htm.
[33] Acemoglu, D. et al. (2016), “Import Competition and the Great US Employment Sag of the 2000s”, Journal of Labor Economics, Vol. 34/S1, pp. S141-S198, https://doi.org/10.1086/682384.
[43] Acemoglu, D. and P. Restrepo (2022), “Tasks, Automation, and the Rise in U.S. Wage Inequality”, Econometrica, Vol. 90/5, pp. 1973-2016, https://doi.org/10.3982/ecta19815.
[3] Acemoglu, D. and P. Restrepo (2020), “Robots and Jobs: Evidence from US Labor Markets”, Journal of Political Economy, Vol. 128/6, pp. 2188-2244, https://doi.org/10.1086/705716.
[67] Acemoglu, D. and P. Restrepo (2020), “Unpacking Skill Bias: Automation and New Tasks”, AEA Papers and Proceeding, Vol. 110, pp. 356-361, https://doi.org/10.1257/pandp.20201063.
[62] Acemoglu, D. and P. Restrepo (2019), “Automation and New Tasks: How Technology Displaces and Reinstates Labor”, Journal of Economic Perspectives, Vol. 33/2, pp. 3-30, https://doi.org/10.1257/jep.33.2.3.
[23] Acemoglu, D. and P. Restrepo (2018), “The Race between Man and Machine: Implications of Technology for Growth, Factor Shares, and Employment”, American Economic Review, Vol. 108/6, pp. 1488-1542, https://doi.org/10.1257/aer.20160696.
[86] African Development Bank (2025), Africa’s AI Productivity Gain: Pathways to Labour Efficiency, Economic Growth and Inclusive Transformation, African Development Bank, Abidjan, https://www.afdb.org/en/documents/africas-ai-productivity-gain-pathways-labour-efficiency-economic-growth-and-inclusive-transformation.
[18] Agrawal, A., J. Gans and A. Goldfarb (2019), “Artificial Intelligence: The Ambiguous Labor Market Impact of Automating Prediction”, Journal of Economic Perspectives, Vol. 33/2, pp. 31-50, https://doi.org/10.1257/jep.33.2.31.
[76] Aleksynska, M. and A. Kolev (2021), “Education-occupation mismatch in the context of informality and development”, OECD Development Centre Working Papers, No. 346, OECD Publishing, Paris, https://doi.org/10.1787/3291e65c-en.
[74] Aleksynska, M., J. La and T. Manfredi (2023), “Transitions to and from formal employment and income dynamics: Evidence from developing economies”, OECD Development Centre Working Papers, No. 349, OECD Publishing, Paris, https://doi.org/10.1787/bc186b3b-en.
[68] Almorsi, A., M. Ahmed and W. Gomaa (2025), “Guided Code Generation with LLMs: A Multi-Agent Framework for Complex Code Tasks”, Proceedings of the 2024 IEEE International Japan-Africa Conference on Electronics communications and Computations (JAC ECC), Institute of Electrical and Electronics Engineers, New York, https://doi.org/10.48550/arXiv.2501.06625.
[30] Apicella, A. (2025), “Automation in Agriculture: Occupational Trends, Worker Outcomes, and Labor Market Implications”, International Journal of Academic Research in Business and Social Sciences, Vol. 15/8, pp. 398-418, https://kwpublications.com/papers_submitted/17811/automation-in-agriculture-occupational-trends-worker-outcomes-and-labor-market-implications.pdf.
[7] Arntz, M., T. Gregory and U. Zierahn (2016), “The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis”, OECD Social, Employment and Migration Working Papers, No. 189, OECD Publishing, Paris, https://doi.org/10.1787/5jlz9h56dvq7-en.
[57] Autor, D. (2024), “Applying AI to Rebuild Middle Class Jobs”, NBER Working Paper, No. w32140, National Bureau of Economic Research, Cambridge, MA, https://ssrn.com/abstract=4722981.
[32] Autor, D. (2015), “Why Are There Still So Many Jobs? The History and Future of Workplace Automation”, Journal of Economic Perspectives, Vol. 29/3, pp. 3-30, https://doi.org/10.1257/jep.29.3.3.
[6] Autor, D., F. Levy and R. Murnane (2003), “The Skill Content of Recent Technological Change: An Empirical Exploration”, The Quarterly Journal of Economics, Vol. 118/4, pp. 1279-1333, https://doi.org/10.1162/003355303322552801.
[50] Berg, J. (2015), “Labour Market Institutions: The Building Blocks of Just Societies”, in Berg, J. (ed.), Labour Markets, Institutions and Inequality: Building Just Societies in the 21st Century, Edward Elgar Publishing, Cheltenham and International Labour Organization, Geneva, https://researchrepository.ilo.org/esploro/outputs/book/Labour-markets-institutions-and-inequality/995328484002676?institution=41ILO_INST.
[54] Berg, J. et al. (2018), Digital labour platforms and the future of work: Towards decent work in the online world, International Labour Organization, Geneva, https://www.ilo.org/publications/digital-labour-platforms-and-future-work-towards-decent-work-online-world.
[24] BIAC (2025), Boosting Productivity and Business Growth: The Role of Artificial Intelligence (AI) Skills, Business at OECD, Paris, https://www.businessatoecd.org/hubfs/website/documents/pdf/Education/Boosting%20Productivity%20and%20Business%20Growth%20The%20Role%20of%20AI%20Skills.pdf.
[34] Brynjolfsson, E., D. Li and L. Raymond (2025), “Generative AI at Work”, The Quarterly Journal of Economics, Vol. 140/2, pp. 889-942, https://doi.org/10.1093/qje/qjae044.
[90] Bullock, J. (ed.) (2024), Harms of AI, Oxford Handbooks, Oxford Academic, https://doi.org/10.1093/oxfordhb/9780197579329.013.65.
[70] Cao, H., R. Ma and Y. Zhai (2024), “LLM-Collab: a framework for enhancing task planning via chain-of-thought and multi-agent collaboration”, Applied Computing and Intelligence, Vol. 4/2, pp. 328-348, https://doi.org/10.3934/aci.2024019.
[38] Carbonero, F., E. Ernst and E. Weber (2018), “Robots worldwide: The impact of automation on employment and trade”, Research Department Working Paper, No. 36, International Labour Organization (ILO), Geneva, https://www.ilo.org/publications/robots-worldwide-impact-automation-employment-and-trade.
[53] Casilli, A. (2025), Waiting for Robots. The Hired Hands of Automation, The University of Chicago Press, Chicago, IL, https://press.uchicago.edu/ucp/books/book/chicago/W/bo239039613.html.
[22] Cazzaniga, M. et al. (2024), “Gen-AI: Artificial Intelligence and the Future of Work”, Staff Discussion Notes 2024, 001, https://doi.org/10.5089/9798400262548.006.
[61] Cirera, X., D. Comin and M. Cruz (2022), Bridging the Technological Divide. Technology Adoption by Firms in Developing Countries, International Bank for Reconstruction and Development / The World Bank, https://openknowledge.worldbank.org/bitstreams/5a5e55f7-edf8-530e-8e11-aa2e15421a9d/download.
[52] Datta, N. et al. (2023), Working Without Borders: The Promise and Peril of Online Gig Work, World Bank, Wasghinton DC, https://doi.org/10.1596/40066.
[59] Dell’Acqua et al. (2026), “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality”, Organization Science, Vol. 37/2, pp. 403-423, https://doi.org/10.1287/orsc.2025.21838.
[44] Doorley, K. et al. (2023), Automation and Income Inequality in Europe, IZA Discussion Paper 16499. Institute of Labor Economics (IZA), https://docs.iza.org/dp16499.pdf.
[27] Eoundou, T. et al. (2024), “GPTs are GPTs: Labor market impact potential of LLMs”, Science, Vol. 384/6702, pp. 1306-1308, https://doi.org/10.1126/science.adj0998.
[11] 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, pp. 2195-2217, https://doi.org/10.1002/smj.3286.
[2] Filippi, E., M. Bannò and S. Trento (2023), “Automation technologies and their impact on employment: A review, synthesis and future research agenda”, Technological Forecasting and Social Change, Vol. 191, p. 122448, https://doi.org/10.1016/j.techfore.2023.122448.
[29] Filippucci et al. (2025), “Macroeconomic productivity gains from Artificial Intelligence in G7 economies”, OECD Artificial Intelligence Papers, No. 41, https://doi.org/10.1787/a5319ab5-en.
[4] Frey, C. and M. Osborne (2017), “The future of employment: How susceptible are jobs to computerisation?”, Technological Forecasting and Social Change, Vol. 114, pp. 254-280, https://doi.org/10.1016/j.techfore.2016.08.019.
[87] García Zaballos, A., E. Iglesias Rodríguez and P. Puig Gabarró (2020), “Annual Report of the Broadband Development Index (IDBA) 2020: Digital Gap in Latin America and the Caribbean”, Discussion Paper, No. IDB-DP-824, Inter-American Development Bank, Washington DC, https://doi.org/10.18235/0003337.
[20] Gmyrek, P. et al. (2025), “Generative AI and Jobs: A Refined Global Index of Occupational Exposure”, ILO Working Paper, No. 140, International Labour Organization, Geneva, https://doi.org/10.54394/hetp0387.
[41] Gmyrek, P., M. Viollaz and H. Winkler (2026), “Disruption without Dividend? How the Digital Divide and Task Differences Split GenAI’s Global Impact”, ILO Working Paper, No. 166, International Labour Organization, Geneva, https://doi.org/10.54394/00033147.
[42] Gmyrek, P., H. Winkler and S. Garganta (2024), “Buffer or Bottleneck? Employment Exposure to Generative AI and the Digital Divide in Latin America”, ILO Working Paper, No. 112, International Labour Organization, Geneva, https://www.ilo.org/sites/default/files/2024-07/WP121_web.pdf.
[49] Graham, M., I. Hjorth and V. Lehdonvirta (2017), “Digital labour and development: impacts of global digital labour platforms and the gig economy on worker livelihoods”, Transfer: European Review of Labour and Research, Vol. 23/2, pp. 135-162, https://doi.org/10.1177/1024258916687250.
[81] Hanna, R. and B. Olken (2018), “Universal Basic Incomes versus Targeted Transfers: Anti-Poverty Programs in Developing Countries”, Journal of Economic Perspectives, Vol. 32/4, pp. 201-226, https://economics.mit.edu/sites/default/files/publications/JEP%20UBI%20vs%20Targeted%20Final.pdf.
[55] ILO (2026), “Decent Work in the Platform Economy Convention, 2026 (No. 193)”, 114th Session of the International Labour Conference (1 June 2026), International Labour Organization, Geneva, https://www.ilo.org/resource/record-decisions/convention-concerning-decent-work-platform-economy-2026.
[93] ILO (2026), “Gen AI, occupational segregation and gender equality in the world of work”, Research Brief, International Labour Organization, Geneva, https://doi.org/10.54394/00033798.
[80] ILO (2026), Lifelong learning and skills for the future, International Labour Organization, Geneva, https://doi.org/10.54394/00033011.
[91] ILO (2026), “Statistics on Social Dialogue”, ILOSTAT, International Labour Organization, Geneva, https://ilostat.ilo.org/topics/industrial-relations/ (accessed on 1 April 2026).
[92] ILO (2024), Social Dialogue Report 2024: Peak-level social dialogue for economic development and social progress, International Labour Organization, Geneva, https://doi.org/10.54394/LIXW6848.
[72] ILO (2023), The International Standard Classification of Occupations (ISCO-08) companion guide, International Labour Organization, Department of Statistics, Geneva, https://www.ilo.org/publications/international-standard-classification-occupations-isco-08-companion-guide.
[47] ILO (2022), Decent work in the platform economy: Reference document for the Meeting of experts on decent work in the platform economy, No. MEDWPE/2022, International Labour Organization, Geneva, https://www.ilo.org/resource/decent-work-platform-economy.
[66] ILO (2018), “Report of the Conference: Report III”, 20th International Conference of Labour Statisticians (Geneva, 10-19 October 2018), No. ICLS/20/2018/3, International Labour Organization, Department of Statistics, Geneva, https://www.ilo.org/sites/default/files/wcmsp5/groups/public/@dgreports/@stat/documents/publication/wcms_651209.pdf.
[94] ILO (2013), Measuring informality: A statistical manual on the informal sector and informal employment, International Labour Organization, Geneva, https://www.ilo.org/publications/measuring-informality-statistical-manual-informal-sector-and-informal.
[65] ILO (2012), “Volume 1: Structure, Group Definitions and Correspondence Tables”, in International Standard Classification of Occupations: ISCO-08, International Labour Organization, Geneva, https://www.ilo.org/publications/international-standard-classification-occupations-2008-isco-08-structure.
[37] 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.
[12] 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, No. 282, OECD Publishing, Paris, https://doi.org/10.1787/646aad77-en.
[71] Lewandowski, P. et al. (2020), “Ageing of routine jobs in Europe”, Economic Systems, Vol. 44/4, https://doi.org/10.1016/j.ecosys.2020.100816.
[25] Manyika, J. et al. (2017), A Future that Works: Automation, Employment, and Productivity, McKinsey Global Institute, New York, https://www.mckinsey.com/featured-insights/digital-disruption/harnessing-automation-for-a-future-that-works.
[16] Maslej, N. et al. (2025), The AI Index 2025 Annual Report, AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, https://doi.org/10.48550/arXiv.2504.07139.
[9] Mihaylov, E. and K. Tijdens (2019), “Measuring the Routine and Non-Routine Task Content of 427 Four-Digit ISCO-08 Occupations”, Tinbergen Institute Discussion Paper, No. TI 2019-035/IV, Tinbergen Institute, Amsterdam and Rotterdam, https://papers.tinbergen.nl/19035.pdf.
[31] Mokyr, J., C. Vickers and N. Ziebarth (2015), “The History of Technological Anxiety and the Future of Economic Growth: Is This Time Different?”, Journal of Economic Perspectives, Vol. 29/3, pp. 31-50, https://doi.org/10.1257/jep.29.3.31.
[85] Nayyar, G., M. Hallward-Driemeier and E. Davies (2021), At Your Service? The Promise of Services-Led Development, World Bank, Washington DC, http://hdl.handle.net/10986/35599.
[83] Nayyar, G., D. Vorisek and S. Yu (2023), Unlocking the promise of services-led growth, VoxEU, Centre for Economic Policy Research, London, https://cepr.org/voxeu/columns/unlocking-promise-services-led-growth (accessed on 1 July 2026).
[8] Nedelkoska, L. and G. Quintini (2018), “Automation, skills use and training”, OECD Social, Employment and Migration Working Papers, No. 202, OECD Publishing, Paris, https://doi.org/10.1787/2e2f4eea-en.
[19] Nordhaus, W. (2021), “Are We Approaching an Economic Singularity? Information Technology and the Future of Economic Growth”, American Economic Journal: Macroeconomics, Vol. 13/1, pp. 299-332, https://doi.org/10.1257/mac.20170105.
[60] Noy, S. and W. Zhang (2023), “Experimental evidence on the productivity effects of generative artificial intelligence”, Science, Vol. 381/6654, pp. 187-192, https://doi.org/10.1126/science.adh2586.
[63] O*NET Resource Center (2026), O*NET 30.2 Database, https://www.onetcenter.org/database.html (accessed on 1 March 2026).
[64] OECD (2026), “Key Indicators of Informality based on Individuals and their Households (KIIbIH)”, OECD International Development Statistics (database), https://data-explorer.oecd.org (accessed on 1 July 2026).
[39] OECD (2025), OECD Skills Outlook 2025: Building the Skills of the 21st Century for All, OECD Publishing, Paris, https://doi.org/10.1787/26163cd3-en.
[89] OECD (2025), “Towards a common reporting framework for AI incidents”, OECD Artificial Intelligence Papers, No. 34, OECD Publishing, Paris, https://doi.org/10.1787/f326d4ac-en.
[73] OECD (2024), Breaking the Vicious Circles of Informal Employment and Low-Paying Work, OECD Publishing, Paris, https://doi.org/10.1787/f95c5a74-en.
[58] OECD (2024), Do Adults Have the Skills They Need to Thrive in a Changing World?: Survey of Adult Skills 2023, OECD Publishing, https://doi.org/10.1787/b263dc5d-en.
[17] OECD (2024), Job Creation and Local Economic Development 2024: The Geography of Generative AI, OECD Publishing, Paris, https://doi.org/10.1787/83325127-en.
[88] OECD (2024), “OECD AI Principles overview”, OECD.AI Policy Observatory, https://oecd.ai/en/ai-principles (accessed on 19 August 2026).
[15] OECD (2023), “AI language models: Technological, socio-economic and policy considerations”, OECD Digital Economy Papers, No. 352, OECD Publishing, Paris, https://doi.org/10.1787/13d38f92-en.
[78] OECD (2023), Building Future-Ready Vocational Education and Training Systems, OECD Reviews of Vocational Education and Training, OECD Publishing, Paris, https://doi.org/10.1787/28551a79-en.
[75] OECD (2023), Informality and Globalisation: In Search of a New Social Contract, OECD Publishing, Paris, https://doi.org/10.1787/c945c24f-en.
[46] OECD (2023), OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market, OECD Publishing, Paris, https://doi.org/10.1787/08785bba-en.
[14] OECD (2019), Artificial Intelligence in Society, OECD Publishing, Paris, https://doi.org/10.1787/eedfee77-en.
[13] OECD (2019), “Overview: The future of work is in our hands”, in OECD Employment Outlook 2019: The Future of Work, OECD Publishing, Paris, https://doi.org/10.1787/e4718721-en.
[45] OECD (2019), Under Pressure: The Squeezed Middle Class, OECD Publishing, Paris, https://doi.org/10.1787/689afed1-en.
[79] OECD (2018), Seven Questions about Apprenticeships: Answers from International Experience, OECD Reviews of Vocational Education and Training, OECD Publishing, Paris, https://doi.org/10.1787/9789264306486-en.
[95] OECD (forthcoming), “Using Artificial Intelligence to harmonise household survey data on occupations and sectors”, OECD Development Centre Working Papers, OECD Publishing, Paris.
[77] OECD/ILO (2019), Tackling Vulnerability in the Informal Economy, Development Centre Studies, OECD Publishing, Paris, https://doi.org/10.1787/939b7bcd-en.
[69] OpenAI (2025), The state of enterprise AI, OpenAI, https://openai.com/index/the-state-of-enterprise-ai-2025-report/.
[21] Pizzinelli, C. et al. (2023), “Labor Market Exposure to AI: Cross-country Differences and Distributional Implications”, MF Working Papers 2023, 216, https://doi.org/10.5089/9798400254802.001.
[5] Quintini, G. (2024), “Job automation: complementarities and capabilities”, in Carcillo, S. and S. Scarpetta (eds.), Handbook on Labour Markets in Transition: Promoting Resilience in a World in Flux, Edward Elgar Publishing, Cheltenham, https://doi.org/10.4337/9781839106958.00008.
[97] Raisch, S. and S. Krakowski (2021), “Artificial Intelligence and Management: The Automation-Augmentation Paradox”, Academy of Management Review, Vol. 46/1, pp. 192-210, https://doi.org/10.5465/amr.2018.0072.
[96] Raj, R., A. Srivastava and R. Behera (2026), “Automation — Augmentation of Artificial Intelligence and Human Intelligence”, Technovation, Vol. 152, p. 103462, https://doi.org/10.1016/j.technovation.2025.103462.
[84] Rodrik, D. and R. Sandhu (2024), “Servicing Development: Productive Upgrading of Labor-Absorbing Services in Developing Economies”, NBER Working Paper, No. 32738, National Bureau of Economic Research, Cambridge, MA, https://doi.org/10.1111/1758-5899.70003.
[51] Rosenblat, A. and L. Stark (2016), “Algorithmic Labor and Information Asymmetries: A Case Study of Uber’s Drivers”, International Journal of Communication, Vol. 10/27, pp. 3758–3784, https://ssrn.com/abstract=2686227.
[82] Sabates-Wheeler, R., A. Hurrell and S. Devereux (2015), “Targeting Social Transfer Programmes: Comparing Design and Implementation Errors Across Alternative Mechanisms”, Jounal of International Development, Vol. 27/8, pp. 1521–1545, https://doi.org/10.1002/jid.3186.
[56] Schlogl, L. and A. Sumner (2020), “Technological Transformation”, in Disrupted Development and the Future of Inequality in the Age of Automation, Springer International Publishing, Cham, https://doi.org/10.1007/978-3-030-30131-6_4.
[48] Schwellnus, C. et al. (2019), “Gig economy platforms: Boon or Bane?”, OECD Economics Department Working Papers, No. 1550, OECD Publishing, Paris, https://doi.org/10.1787/fdb0570b-en.
[28] Svanberg, M. et al. (2024), “Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision?”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.4700751.
[10] Webb, M. (2019), “The Impact of Artificial Intelligence on the Labor Market”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.3482150.
[36] WEC (2024), The Work We Want, World Employment Confederation, Brussels, https://wecglobal.org/press-release-post/global-study-reveals-how-ai-is-transforming-work-as-we-know-it/.
[35] WEF (2025), Future of Jobs Report 2025, World Economic Forum, Geneva, https://www.weforum.org/reports/the-future-of-jobs-report-2025/.
[40] World Bank (2016), World Development Report 2016: Digital Dividends, World Bank, Washington DC, https://www.worldbank.org/en/publication/wdr2016.
← 1. Augmentation refers to the use of technology, including AI, to enhance the capabilities of workers rather than replace them (automation). Whereas automation implies that machines take over a human task, augmentation means that humans collaborate closely with machines to perform a task (Raisch and Krakowski, 2021[97]; Raj, Srivastava and Behera, 2026[96]).
← 2. Primary activities refer to the direct extraction or use of natural resources, such as agriculture, forestry, fishing, mining and quarrying. It includes ISIC Rev. 4 sectors A and B. Secondary activities refer to the transformation of raw materials and the production of goods, including manufacturing, utilities and construction. It includes ISIC Rev. 4 sectors C to F. Tertiary activities refer to services, including trade, transport, accommodation, finance, public administration, education, health and other service activities. It includes ISIC Rev. 4 sectors G to U.