Measuring the impacts.

To assess the impact of AI on productivity and the workforce, economists generally use two metrics. One focuses on the associated increases in productivity, that is, the amount of goods and services produced for given inputs such as labour and capital. The other considers workforce exposure, that is, the degree to which their tasks can be performed by AI systems; the higher the exposure, the greater the potential for complementation or substitution.

To date, research that employs systematic applied methods on data sets with good coverage and adequate scale is mostly based on micro-level studies on early adopters in developed countries. It is far from conclusive, yet suggests that firms using AI can make substantial productivity gains, particularly those employing skilled workers and those in service industries. A summary of recent firm-level studies indicates that AI can increase both labour productivity and total factor productivity, although the range of the estimates is wide, reflecting the differing capacities of firms to benefit from AI (figure II.2).

 For example, in some firms in Germany, sales achieved per worker increased substantially with higher levels of AI use. In some firms in Italy, total factor productivity increased by 2.2 per cent with the adoption of AI. A study of large firms from a range of countries showed that the accumulated stock of AI knowledge increased total factor productivity by 6.7 per cent. 

The impact may also depend on firm characteristics, such as size, although the evidence is mixed. Some studies showed higher productivity gains in larger firms that could benefit from scale effects and greater financial resources. Other studies showed advantages for smaller firms that could integrate new technologies more rapidly within existing production systems. Most of the literature concentrates on developed countries, for which there is more detailed firm-level data. However, similar benefits could also arise in developing countries, as indicated by an analysis of listed firms in China. The early evidence thus suggests that the use of AI can enhance productivity, yet does not clarify the exact drivers.  Similarly, at a business consultancy, consultants supported by ChatGPT were 12 per cent more efficient and had a 40 per cent increase in work quality. Other studies demonstrate notable productivity enhancements in professional writing and computer coding. These micro-level studies used experimental or quasi-experimental designs to infer causal links between the use of GenAI tools and gains in labour productivity. They showed significant differences between workers at different skill levels, and it is therefore not clear from the studies whether the use of AI can reduce or increase inequality across workers. For example, one study found that the largest productivity improvements in a customer service centre were from the least-skilled and least-experienced workers, who used an AI assistant to learn the good practices of the highest-skilled workers. On the other hand, another study, on science material researchers, showed much higher productivity gains for leading researchers.

This may be because the most experienced scientists were able to take advantage of their knowledge to prioritize the most promising AI suggestions, while the 30 per cent of least-productive researchers spent time on testing less promising options. Most of the evidence to date comes from early adopters, and whether similar productivity gains apply to latecomers, particularly from developing countries far from the technological frontier, remains to be ascertained. Overall, the impact of AI, particularly the use of GenAI, tends to be greater for particular service-related tasks. Yet the benefits can also extend indirectly to other firms. Therefore, it is important to foster inter-industry synergies and complementarities between knowledge-based services and manufacturing and the primary sector in order to transmit productivity gains through the economy and drive an AI-powered industrial transformation. More comprehensive studies that consider complex tasks that are more difficult for AI to learn can help better understand the impact of AI across the economy. Nonetheless, the early evidence on GenAI complements the findings from firm level studies that show that the use of AI can increase productivity (box II.1).

Previous waves of technology primarily impacted blue-collar occupations, but those most exposed by AI are in knowledge-intensive sectors. A recent OECD survey on job markets in Europe and North America listed the top industries prone to AI automation as those in finance, advertising, consulting and information technology. Similarly, a study in India based on online job postings between 2016 and 2019 found that AI-related skills requirements were concentrated in information technology, finance and professional services. A recent global survey found that GenAI was being adopted least in manufacturing and more commonly in marketing and sales, product and services development and information technology functions. It should be noted, however, that even in non-knowledge intensive sectors, there are jobs highly exposed to AI. One study estimated that AI would affect 40 per cent of global employment, showing that one third of jobs in developed countries had high potential for AI automation and around 27 per cent were exposed to AI augmentation. Workforces in advanced economies are at greater risk since more of their jobs involve cognitive tasks. However, these economies are also better positioned than emerging and low-income economies to capitalize on the benefits of AI. For individual countries, the impacts depend on their occupational structures. For example, the United Kingdom has a significant share of employment in professional and managerial occupations that are highly exposed to AI augmentation, as well as in clerical support and technician occupations that could be exposed to AI-related automation. Developed countries are in general more likely than developing countries to face more immediate labour market adjustments and an increase in wage inequality.

In contrast, in India, for example, most workers are agricultural workers and craftspeople, who are less exposed. Developing countries might, therefore, have time to gain insights from the experiences in developed countries. A similar picture is seen when considering the impact of GenAI. Workers with higher levels of education are more exposed but also more likely to benefit. Overall, GenAI offers greater potential for labour augmentation than automation, particularly in low- and middle-income countries (figure II.3). Technicians and associate professionals can gain from augmentation while clerical support workers are highly exposed to automation. 

Exposure to GenAI within job categories is relatively balanced from a gender perspective, but the over-representation of women among clerical support workers makes them more exposed to automation, particularly in the United States and Europe. A study in Latin America showed that GenAI was more likely to lead to augmentation than automation and to favour urban, educated and higher-income workers in formal occupations, with the benefits fairly evenly distributed across gender and age. 

The study highlighted that nearly half of the occupations that could benefit from augmentation faced digital barriers. In addition, there is a significant gender-related imbalance in automation, largely because women are more likely to perform the most exposed jobs; the proportion of women-held jobs that are exposed to automation can be up to twice that of men. This, combined with the gender divide in digital skills and access to ICTs, can limit the benefits of AI adoption for women, thus widening existing inequalities. It should be emphasized that the impact of AI on the labour market depends on the rate of technology adoption, as well as on other non-technological factors, such as the relative prices of capital and labour, economic structures and the social acceptance of new technology. These factors amplify or reduce expected AI-related impacts between sectors and countries.

Despite concerns about widespread job losses, the pace of automation has been slower than initially predicted. In one survey conducted in 2020, employers expected that 42 per cent of their business tasks would be automated by 2027 but, subsequently, employers have reduced their estimates. As in previous waves of technological innovation, the use of AI has also created new jobs. 

One study of seven high-income countries found that while the use of AI had automated some tasks in finance and manufacturing, it had also introduced new tasks, and most employers reported higher productivity but no overall impact on employment. Box II.2 provides further discussion on the impact of AI in knowledge-intensive sectors.

Current evidence suggests that the future scenario is likely to be a complex interplay of automation, augmentation and the emergence of new roles. Automation is likely to reduce the labour share in value added in favour of capital, which will result in slower growth in wages than productivity and increasing wealth concentration. However, this tendency can be counterbalanced by the benefits of augmentation and of generating new tasks for workers. 

It is important to understand and plan for all eventualities. Increasing inequalities have already been stirring social discontent and weakening trust in public institutions, while increasing political polarization and undermining democratic governance. Policymakers and businesses need to understand these dynamics to ensure that the benefits of AI are distributed equitably and to facilitate smooth transitions