The OECD/Boston Consulting Group/INSEAD study conducted in 2022-23 examines enterprises using artificial intelligence (AI) across the Group of Seven (G7) countries, focusing on the manufacturing and information and communication technology (ICT) services sectors. Novel survey questions address topics relevant to public policy, such as workforce skills and qualifications; types of collaboration with universities and research organisations; barriers to using AI; enterprises’ use of public programmes of financial support; and spending on research and development for AI. Manufacturers, especially small ones, face more adoption obstacles, while larger ICT enterprises invest heavily in employee training and hiring for AI. This chapter also reports on the widespread use of various types of public sector support, as well as enterprises' perspectives on the usefulness of these different support types and the priority they assign to related policy initiatives.
The Adoption of Artificial Intelligence in Firms
Abstract
As described in Chapter 1, the OECD/Boston Consulting Group/INSEAD survey conducted in 2022‑23 focuses on enterprises using artificial intelligence (AI) and compares relevant characteristics of AI use across the Group of Seven (G7) countries, two economic sectors (manufacturing and information and communications technology [ICT] services), and two enterprise size classes (from 50 to 249 employees and from 250 employees upwards).
The sample sizes are not statistically representative of the population of enterprises in each country. However, the 30-enterprise cell size adheres to widely used statistical norms. Despite the lack of representativeness to national enterprise populations, the findings suggest correlations that other institutions may wish to explore using larger samples. In addition, with a total of 840 enterprises and very few missing data points, rigorous within-sample analysis is feasible.
Relative to previous national and supranational surveys, many of the questions in the survey are novel and of direct policy relevance, such as which types of public support enterprises find useful. All the surveyed enterprises use AI in at least one application, so part of the analysis focuses on the number of AI applications adopted and how this relates to enterprise and industry characteristics. The sample comprises relatively advanced AI users. Consequently, from a policy standpoint, the findings may become increasingly relevant as the number of enterprises seeking to become (advanced) users of AI grows.
A range of possible obstacles to adopting AI were considered – from difficulties in estimating the return on investment (ROI) in AI to lack of external finance. Manufacturers experience all the obstacles more frequently than enterprises in ICT. Small manufacturers are the most likely to experience barriers to adopting AI.
Spending on research and development (R&D) for AI, as a share of all R&D spending, is positively related to how critical enterprises deem AI to be. Perhaps unsurprisingly, enterprises with a higher share of R&D spending going to AI are more likely to establish collaborations on AI with researchers in public research organisations. They also use public training services more often. Performing R&D with the help of AI is also one of the most widespread applications of AI.
Nearly three-quarters of enterprises in both sectors rely on employee training to adopt AI. More than 60% hire new staff to help develop AI technologies. Large enterprises in ICT are the most likely to train employees and hire staff to develop AI technologies. Around 20% of enterprises fail to find suitably qualified candidates when attempting to hire. However, many enterprises also appear to have difficulties fully understanding the skills they need. Most enterprises support developing qualification frameworks for graduates in AI.
Between 51% and 61% of enterprises make use of external data, whether from private data providers (such as organisations dedicated to producing and selling data), from a partner enterprise, or from the public sector.
A significant share of enterprises uses some form of public sector service to aid AI adoption. Most enterprises consider public sector services and initiatives “helpful” or “very helpful”. Enterprises that use more AI applications are more likely to use public support. Perhaps unsurprisingly, large enterprises are less likely to use public support. Enterprises that use AI intensively or face many obstacles to using AI find public services and initiatives more helpful than those that use AI less intensively or experience fewer obstacles. The generally positive view of potential public sector initiatives varies little in terms of industry and firm size.
Both in ICT and manufacturing, the most frequently used services are those that provide access to diverse forms of information or advice. Even in this sample of enterprises, many of which use AI in advanced ways, additional information on various domains of AI is often sought. Initiatives to develop human capital are also among the most widely used and highly valued. Roughly 58% of enterprises make use of training services provided by the public sector. In addition, 42% use programmes that promote access to finance, such as tax credits on R&D spending, grants or credit guarantees.
To help adopt and develop AI, many enterprises in the sample collaborate with universities, public research organisations, and other partners. More than half have worked with university faculty, PhD or postdoctoral students over the past 12 months.
The tables in Annex E report the aggregated responses to each survey question.
All the surveyed enterprises use AI in at least one application. As shown in Figure 3.1, 53% of the sampled enterprises consider AI critically important to their operation. Some 39% hold AI to be one among several important elements in the enterprise, and only 8% view AI as of minor importance.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Because all the sampled enterprises use AI, part of the analysis focuses on what economists refer to as the intensive margin of AI use, i.e. the number of AI applications adopted and how this relates to enterprise and industry characteristics. The intensive margin of AI use has been little studied in the previous literature.
Furthermore, the survey categorises AI applications by the business function they are used for, while in much of the wider literature, AI applications are typically characterised by technology. Examples of business functions are Product Design, Human Resources (HR) and R&D. Examples of AI technologies are speech recognition, image recognition and natural language generation. Each AI technology can be used in many functions. For instance, natural language processing (NLP) can be used for staff recruitment and human resource management, training and cognitive support for workers, customer-facing services and many more. Thus, enterprises may exploit economies of scope associated with AI technologies, using them in several business functions once they are introduced. Indeed, in the current survey, the number of applications enterprises use is higher than found in most other surveys, which may reflect this phenomenon. However, while complementing other studies, direct comparisons are not possible since the set of questions is not identical.
AI can be adopted as a point solution (i.e. solving one specific problem), an application solution or a systems solution. Furthermore, AI used as point solutions can be introduced independently of other functions in an enterprise, and this is typically a first step as enterprises adopt AI. For instance, an enterprise may start using an NLP application to manage customer data, extend its use over time into recommendation systems, and extend further to customer relation management systems, including sales forecasting.
Most of the survey questions have binary answers (e.g. “is not an obstacle/is an obstacle”). The analysis, therefore, applies probit regressions to individual measures or correlates the sum of positive answers to enterprise or industry characteristics. Probit regression is a statistical technique used to estimate the probability of a binary outcome occurring for a population, in this case, a population of AI-using enterprises.
Number of uses of AI
The analysis starts by investigating the number of AI applications each enterprise uses and whether systematic differences exist across countries, industries, and enterprise size. The number of AI applications is the number of “Yes” answers to a screening question presenting the 11 possible applications shown in Table 3.1.
Source: OECD authors. 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
The average number of uses across all enterprises is 5.7, and the distribution around the average is slightly skewed towards the right (Figure 3.2). Figure 3.2 can be read as follows: The vertical axis shows the percentage of enterprises adopting the corresponding number of AI applications indicated on the horizontal axis. Thus, about 10% of enterprises use three AI applications, and another 10% use eight AI applications. Or – looking at the horizontal axis – if, for example, the reader is interested in seeing the share of enterprises that uses AI in ten applications, about 7.5 % of the sample is in that category. Given that the sample includes only enterprises that use AI, no enterprises register a zero on the horizontal axis.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Of the sample of 840 enterprises, only 4 responded “yes” to using AI in all 11 applications in Table 3.1. Of these, three are in subsectors of ICT: two with their main activity in Data Processing and Hosting Activities, and one in Writing, Testing and Supporting Software. A fourth enterprise operates in Oil and Gas. Three of these four enterprises have more than 250 employees.
Figure 3.3 breaks down the average number of AI uses in relation to enterprise size. The two size categories, in fact, show a very similar pattern. The average number of uses between the size groups is about the same, and so is the skewness. This result is perhaps surprising, as it is well-established that enterprise size is a strong predictor of AI adoption (Bughin et al., 2017[2]; Kinkel, Baumgartner and Cherubini, 2022[3]; Zolas et al., 2020[4]).Considering the average across G7 countries, the OECD Database on ICT Access and Usage by Businesses reports that the share of enterprises using AI is about twice as high in enterprises with more than 250 employees than in enterprises with 50‑249 employees (OECD, 2023[5]).
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
However, as previously mentioned, earlier studies mainly focus on AI technologies that can be used in several business functions. A possible explanation for the current and somewhat surprising survey result is that small enterprises are less specialised in their business functions than larger enterprises. Having incurred considerable investment costs to adopt one or more AI technologies, multitasking teams in small and medium-sized enterprises may use them in several functions.
Number of AI uses (horizontal axis) and percentage of the sample population of enterprises (vertical axis)
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
The patterns of AI adoption are quite different across industries (Figure 3.4). The highest average number of AI uses is in the subsector of Manufacturing of Machinery. The lowest is in Data Processing and Hosting and Online Platforms. While the former is slightly skewed to the left (that is, with a slight tendency to use fewer applications), the latter is slightly skewed to the right. Managing and Operating Clients’ Computer Systems is the industry with the highest skewness, which is skewed to the right. The intensity of AI use by industry largely corresponds with the findings of comprehensive surveys from Sweden (Statistics Sweden, 2021[6]) and the United States (Zolas et al., 2020[4]). However, in Sweden, the highest AI intensity is found in knowledge-intensive business services.
The average number of AI uses also varies across the G7 countries (Figure 3.5). The highest average number of uses (6.4) is in enterprises in France, while the lowest (4.9) is in the United States. The ranking of countries at the intensive margin of AI use is in line with the ranking of the share of enterprises using AI in the OECD Database on ICT Access and Usage by Businesses (OECD, 2023[5]). The results suggest that in countries where the adoption rate of AI is high, enterprises that use AI employ it in more applications. In other words, the extensive and intensive margins are closely related at a country level.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Table 3.2 cross-tabulates the average number of AI uses by country and industry using a heat map. The darker shading represents a high average number of AI uses, while lighter shading depicts industry and country combinations with a low average number of AI uses. French enterprises exhibit both the highest and lowest average number of AI uses, respectively, in Manufacturing of Machinery and in Writing, Modifying, Testing, and Supporting Software. The United States stands out with the highest average use of AI in Managing and Operating Clients' Computer Systems.
The stated importance of AI
Figure 3.6 depicts the number of uses of AI by the importance that enterprises give to AI. Enterprises that report that AI is of critical importance (Panel A) also report the highest average number of uses of AI. Enterprises that report that AI is of minor importance (Panel C) have the lowest average number of AI uses. In this latter category, no enterprises use more than seven AI applications.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
A closer look at the uses of AI: A probit analysis
This section studies the conditional probability of enterprises using AI for each of the 11 applications considered. The application “Detecting defects and anomalies” did not pass standard thresholds for goodness of fit and is therefore excluded.
As discussed earlier, the data show no systematic variation in AI use and enterprise size at the aggregate level. However, looking more closely at this finding, but by application, it turns out that for 3 of the 11 applications, there is a significant relationship between AI use and enterprise size. Table 3.3 reports the probit coefficient on enterprise size and the associated predicted probability of AI use for medium-sized and large enterprises, respectively. Enterprise size is positively related to the use of AI in R&D and in a residual open category of “Other core use”. By contrast, enterprise size is negatively related to the use of AI in Product Design (hence the negative sign on the probit coefficient).
Note: The probit regressions are run with country and industry fixed effects, controlling for enterprise age. There were 840 observations for the applications “Product design” and “R&D” and 512 for “Other core use”. The two asterisks (**) and single asterisk (*) signify statistical significance at the 5% and 10% levels, respectively.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
The variation in AI adoption by application and country is depicted in Figure 3.7. To understand how to read this figure, take the example of Product Design (upper left chart). The solid dots represent a point estimate of the predicted probability of using AI in Product Design for each country. For example, the average probability that an enterprise will use AI in Product Design in France is 0.7. The band indicated by the thin line shows how precise the estimate is, i.e. the probability that French enterprises adopt AI for Product Design lies between 0.62 and 0.78 with 95% certainty.
Figure 3.7 reveals substantial variation both in applications across the G7 countries and across the applications. French enterprises have the highest probability of using AI in all applications except for Product Design and Customer-facing Services. Japanese enterprises are the most likely to use AI applications in Product Design, while Canadian enterprises are the most frequent users of Customer-facing Services applications. Italy is at the opposite end of the scale, where enterprises are the least likely to use AI applications in Process Control and Optimisation, Supply Chain Management, and Logistics and Customer-facing Services. Enterprises in the United States are the least likely to use AI applications in Product Design and HR, while enterprises in the United Kingdom are the least likely to use AI applications in Fabrication and Assembly and in R&D.
95% confidence intervals
Note: The figure depicts the predicted probability of using AI in the applications indicated in the heading of each graphic in the figure. The underlying probit regressions control for enterprise size and have industry fixed effects, using robust standard errors.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
The survey data also reveal considerable variation in the incidence of AI use both across industries and across applications (variation in the use of AI within applications across industries is presented in Annex F). Across industries, the highest-probability application of AI is R&D. This application also has the smallest variation within and across industries. In other words, the application of AI to R&D is the most likely and consistently used application of AI across enterprises in any given industry and across industries overall.
The industry with the highest probability of AI use is Operating and Supporting the Operation of Web Search Portals. AI is least likely to be used in HR (although AI is widely used in HR in enterprises that specialise in Managing and Operating Clients’ Computer Systems and/or Data-processing Facilities). The low frequency of this particular use is perhaps unsurprising, as many enterprises have concerns about inadvertent misapplication of AI in recruitment, a possibility widely acknowledged in public discussion of AI.
Enterprise age and the use of AI
Previous studies have found that enterprise age matters for technology adoption, including AI (Calvino, Criscuolo and Menon, 2016[7]; Haller and Siedschlag, 2011[8]). Older and well-established enterprises may have more resources and experience to absorb the fixed costs of adopting a new technology. However, older enterprises may also face substantial switching costs from old to new technology. Having moved down the average cost curve using their present technology, older enterprises may have less incentive to switch technology than young enterprises and startups that have not incurred the associated sunk costs (Cho et al., 2023[9]); see also Kinkel, Baumgartner and Cherubini (2022[3]) who find that size, R&D and services orientation are important prerequisites for adopting AI in manufacturing enterprises. Which effect dominates is an empirical question.
In this connection, Annex G graphs the relationship between AI uptake, age of enterprise and business function, distinguishing between manufacturing and ICT industries. Older enterprises are more likely to use AI in Fabrication and Assembly as well as Logistics, possibly because these are among the most basic functions in manufacturing. Younger enterprises are the most likely to use the remaining AI applications. The difference between old and new enterprises is particularly large in Customer-facing Services and HR. One possibility is that this reflects high switching costs for older enterprises that may have built up large customer services and HR functions.
The survey data show that enterprises in manufacturing and ICT have different patterns of AI adoption. For enterprises of all ages, the probability of adopting AI is higher in manufacturing in traditional manufacturing business functions, including Logistics; Supply Chain Management; Process Control and Optimisation; Fabrication and Assembly, and Product Design. The probability of adoption is higher in services industries in generic functions such as HR; Training and Cognitive Support for Workers; and Customer-facing Services.
Enterprises need access to large quantities of high-quality data to reap the potential benefits of AI. A first point to note is that, overall, the surveyed enterprises were relatively data mature. The literature provides no single measure of data maturity in enterprises, as data can be used in many ways. The survey included a question that would serve as a proxy, namely whether enterprises use a data management solution, such as a data lake.1
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Most enterprises in the sample – 78% – use such a solution, and only 1% were unfamiliar with the technology (Figure 3.8, Panel A). Smaller enterprises are a little less likely to use a data management solution (Figure 3.8, Panel B).
A survey question was also used to gather insights into enterprises’ data collection practices. Specifically, enterprises were asked if, during the past 12 months, they had collected or otherwise acquired data from any of the following sources:
internally from processes and staff
customers and users
private data providers, such as organisations dedicated to producing and selling data
partner enterprises
research institutes
the public sector.
Roughly 78% of enterprises in the manufacturing and ICT sectors reported collecting data internally from their own processes and staff (Figure 3.9). With almost the same frequency, enterprises also draw data from customers and users (75%). This source encompasses information gathered from customer interactions, feedback and monitoring of usage patterns, which potentially helps enterprises to enhance customer-centric decision making. The widespread use of these data sources highlights the importance of proprietary data for AI usage and development.
In addition to internal and other proprietary data, access to high-quality external data allows enterprises to supplement their internal data to gain broader analytic insight. The survey shows that between 51% and 61% of enterprises make use of external data, whether from private data providers (such as organisations dedicated to producing and selling data), from a partner enterprise, or from the public sector. Manufacturers are somewhat more likely to use data from research institutes than enterprises in ICT (57% and 48%, respectively).
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
The likelihood that any specific data sources are used changes little by size of enterprise. The difference between large and smaller enterprises amounts to less than 10 percentage points for almost all data sources. The exception is for large enterprises in ICT. These are 12 percentage points more likely to use data from public sources than smaller enterprises.
The pattern of use of different data sources differs quite substantially across countries in some instances (Figure 3.10). While 86% of the surveyed enterprises in Japan use internal data, this is only the case for 68% of enterprises in Canada. At the same time, in the United States, most enterprises use data from customers and users as well as private data providers (86% and 70%, respectively). In Italy, this holds for only 60% and 50% of enterprises, respectively. By far the highest incidence of use of data from the public sector is in the United Kingdom (63%).
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Enterprises were asked about the practices they employ to adopt and develop AI. In both sectors, more than 70% report that they carry out R&D on AI technologies for their own use (Figure 3.11). Nearly three-quarters of enterprises in both sectors rely on employee training. In addition, more than 60% hire new staff to further develop AI technologies. Between 53% and 64% of enterprises rely on customised systems built by third parties or purchase off-the-shelf software or hardware. About every second enterprise has institutionalised the development of AI by creating a senior management role or a team with responsibilities for AI. Finally, many enterprises also expedite AI uptake through partnerships with national or international enterprises with capabilities in AI (51% in manufacturing, 41% in ICT).
Large enterprises in ICT are the most likely to train employees (78%) and to hire staff (73%) to develop AI technologies. In contrast, 56% of smaller enterprises in both ICT and manufacturing hire staff for this purpose, compared with 64% of large manufacturers. Consequently, large enterprises in ICT are the least likely to purchase off-the-shelf software or hardware.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Within the sample, almost 86% of enterprises in the United States carry out some level of R&D to develop AI technologies for their own use. This is considerably higher than in all other countries, with Germany having the second highest share, at 75%. Notably, France has the highest share of enterprises that train their employees, hire new staff, or partner with other enterprises to develop AI.
A deeper look at R&D
Prior work has shown that investment in R&D relates to the use and development of AI in several ways. Enterprises with more researchers are better placed in terms of skills to adopt, adapt and innovate with AI. A correlation also exists between the use of AI and spending on R&D, given that AI is an increasingly prevalent research tool (Nolan, 2021[10]).
In the survey sample, spending on R&D for AI, as a share of all R&D spending, is positively related to how critical enterprises deem AI to be. Some 38% of enterprises that allocate between 0‑10% of their R&D spending to AI consider this technology critically important to their core business processes. By comparison, among enterprises that spend between 11% and 30% of their R&D outlays on AI, 68% consider AI critical to their business. This is the case for 87% of enterprises spending more than 30% (Table 3.4).
Variation by country, enterprise size, industry and criticality of AI to the enterprise
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Enterprises that spend more intensively on R&D for AI also use public training services more often. The survey showed that 65% of enterprises allocating more than 30% of their R&D spending to AI have also used public training services in the past 12 months, compared to 51% of companies with R&D spending on AI of up to 10% of their total outlay on R&D.
Perhaps unsurprisingly, enterprises that spend more of their R&D on AI are more likely to establish collaborations on AI with researchers in public research organisations. Between 60% and 65% of enterprises with R&D spending on AI higher than 11% have such collaborations, compared to 44% of enterprises that spend less than 10%. The importance of R&D in connection with AI is noteworthy for policy makers, who possess various tools for incentivising and giving direction to this form of investment. Educational and research institutions also possess a range of tools to facilitate such investments and collaborations.
Collaboration with universities and public research organisations
Many enterprises in the sample collaborate with universities, public research organisations and other partners to aid the use and development of AI. More than half have worked with university faculty, PhD or postdoctoral students over the past 12 months (Table 3.5). Partnerships with researchers in public research organisations are somewhat more prevalent among manufacturers (55%) than enterprises in ICT (48%). Roughly one-third of all enterprises work with undergraduate students.
In a context where AI skills are almost scarce everywhere, one reason why enterprises collaborate with universities is to secure access to talented graduates. As shown in Table 3.5, a high share (76%) of enterprises collaborating with universities recruited graduates in AI in the previous 12 months.
A further indication of the importance of human capital is the high share of enterprises that consider government investment in university education and vocational training related to AI as “very helpful” or “helpful”. Even among enterprises that do not consider AI as central to their core business process, 73% hold that such initiatives are “very helpful” or “helpful” (Table 3.6).
Collaborations with researchers in public research organisations are particularly widespread among smaller manufacturers (64%). By contrast, such collaborations were reported by about half of enterprises in other sectors (Figure 3.12).
Enterprises in the United States are less likely than enterprises in other countries to collaborate with university faculty, undergraduate students and researchers in public research organisations (Figure 3.13). However, 60% collaborate with other partners, which is the third-highest incidence in the G7 group.
The highest share of enterprises working with other partners to develop AI is observed in France (67%). In addition, France has the highest share of enterprises collaborating with university faculty (72%) and undergraduate students (43%). In the sample, enterprise collaboration with public research organisations is most frequent in Japan (58%).
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.Obstacles to using and adopting AI
Workforce skills
In the survey sample, most enterprises had been active in hiring AI skills during the previous 12 months. Indeed, around 60% of enterprises hired employees with AI skills during this period (Figure 3.14). However, an almost universal finding from studies internationally is that a shortage of workforce skills presents a main bottleneck for firms seeking to implement AI. The current survey echoes those findings. Around 20% of enterprises with 50‑250 employees report being unable to find appropriately qualified candidates for available vacancies. Even many large enterprises – approximately 17% – experience the same problem.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Do enterprises understand which skills they need?
A less-frequently addressed question is whether firms fully understand their skills needs and whether formal academic qualifications provide sufficient information to employers making recruitment decisions. There are several reasons why fully understanding skills needs in AI might be a problem for some enterprises. “AI” is, in fact, an umbrella term encompassing many subdisciplines. Firms in sectors without a tradition of data analytics or small firms that do not have the in-house expertise to make the necessary technical distinctions might lack a strong basis on which to search for skills. In addition, AI technologies are changing quickly, complicating the assessment of job seekers’ suitability.
To explore this topic, the survey asked if enterprises had experienced difficulties during the preceding 12 months in understanding what skill sets to look for in potential AI recruits. Almost 19% of respondents had experienced this problem (Figure 3.15). This number varies somewhat by enterprise size, affecting around 20% of smaller enterprises but less than 17% of enterprises with more than 250 employees. This finding has several possible implications for policy, especially concerning the possible development of new qualifications frameworks.
A large majority (86%) of enterprises that highly value support for partnerships with educational and vocational institutions also consider the development of new qualification frameworks to be either “very useful” or “moderately useful” (
Table 3.7). In other words, many enterprises interested in or searching for increased AI skills also feel they need a better practical understanding of how to identify and use the necessary skills.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Other obstacles to using AI
Respondents were asked to indicate which, if any, of 8 conditions had limited the enterprise in implementing AI applications in the preceding 12 months. The options presented were:
1. difficulties in estimating the returns on investment in AI applications
2. concerns related to data privacy, data protection or data security
3. scarcity of cloud computing solutions that guarantee data security and regulatory compliance
4. lack of clarity about the legal consequences in case of damage caused by using AI
5. lack of vendors of AI systems offering solutions tailored to the enterprise’s needs
6. lack of external finance for investment to support AI adoption
7. reluctance of staff to adopt AI
8. difficulties in retraining or upskilling staff.
Manufacturers experience all the above obstacles more frequently than enterprises in ICT. This might have several causes. For example, manufacturing has always been product rather than data-led and has less of a tradition of working with big data (although differences exist within manufacturing, especially as regards continuous flow manufacturing; for instance, of petrochemicals and manufacturing of discrete products, such as cars). There is only one exception to this higher incidence of obstacles experienced by manufacturers, namely concerns related to data privacy, data protection or data security (Figure 3.16). Some 55% of manufacturers and 57% of enterprises in ICT report that these concerns have limited their use of AI in the past 12 months. The comparatively high frequency of concerns among manufacturers about data privacy might be somewhat surprising in that manufacturers generally gather less confidential data than do many enterprises in ICT. However, worries over data security rather than data privacy may be the primary concern here for manufacturers.
Percentage of all enterprises stating the issue to be a concern
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
The most frequently experienced obstacle is the difficulty in estimating a priori the ROI in AI applications. Some 62% of manufacturers and 56% of enterprises in ICT cite this as problematic. This result echoes findings from many previous surveys, as well as the experience of agencies across the G7 countries charged with accelerating the spread of digital technologies in firms (see Chapter 4). Part of the reason for uncertainty around the ROI is that many AI projects involve a degree of experimentation, with no guarantee of success. In addition, the key processes of data cleaning and model development involve an element of art. Compounding these uncertainties, investment decisions might also have to include complex strategic considerations, such as the need for the firm to remain viable in future supply chains.
More than 40% of enterprises in both manufacturing and ICT cite challenges in finding AI system vendors that offer solutions tailored to their needs. This observation, noted in other work, has motivated some public sector agencies – for instance, in Singapore – to signpost vendors with suitable track records, with the aim of lowering search costs, especially for small firms.
Around 40% of enterprises also report that they encounter a lack of clarity around the legal consequences of damages caused by AI, as well as a scarcity of cloud computing solutions that guarantee data security and regulatory compliance (see the following section on cloud computing). Approximately 40% of enterprises affirm that a lack of external finance for investment to support AI adoption limited the use of AI in the past 12 months. However, again, this result is sensitive to enterprise size: larger enterprises are considerably less likely to report such financial barriers (33% in manufacturing and 30% in ICT). This finding might be explained by the fact that smaller enterprises generally possess less capital to invest, as well as greater constraints linked to cash flow, which are especially important limitations when investment returns are relatively uncertain. In such circumstances, externally provided subsidies are much sought after.
Small manufacturers are more likely to experience barriers to adopting AI than any of the remaining groups of enterprises (Figure 3.17). For instance, while about 50% of smaller manufacturers find it hard to identify vendors of AI solutions tailored to their needs, this is true for only 37% of large enterprises in ICT.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Enterprises in the United States are the least likely to report problems for five out of the eight cited obstacles to AI adoption (Figure 3.18). The most pronounced difference between the United States and other countries is in external finance to support the uptake of AI. Only 24% of enterprises in the United States consider this a problem, compared with 38% in Germany. At 45%, Canada has the highest share of enterprises that cite external finance as a challenge.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Roughly every second enterprise reports difficulties in retraining or upskilling staff, a finding which might be amenable to change through education and training policies. A further challenge is the apparent reluctance of some staff to retrain or upskill, as cited by 45% of manufacturers and 34% of enterprises in ICT.
Obstacles to using cloud computing
Previous studies on AI adoption have found a hierarchy of technology adoption where early users of websites and computer systems are also early users of cloud services, followed then by AI use (Zolas et al., 2020[4]).
A significant share of enterprises reports challenges in using cloud computing. This matters because of the complementarity between cloud computing and many AI applications. For instance, Industry 4.0 requires increased data sharing across production sites and company boundaries. Leading edge manufacturers may wish to know the real-time status of production equipment in companies that produce key components for their products. Increasingly, machine data and data analytics, and even monitoring and control systems, will operate in the cloud. However, it is already known from previous studies that cloud use varies significantly between small and large firms, and across countries. For example, in 2021, in Finland, around 99% of firms with 250 or more employees purchased cloud services. By comparison, in Japan, in 2019 (the latest year for which data are available), only 49% of firms of the same size used cloud services. The OECD and EU averages were 74% (2022) and 72% (2021), respectively (OECD, 2023[11]).
While many surveys have collected data on enterprise use of cloud computing, few have explored the reasons for non- or problematic use. The current survey does not directly ask questions about previous use of cloud services. However, the respondents were asked to indicate what kind of obstacles they encountered when using cloud services. The obstacles considered in the survey were the following:
high cost of retooling systems
concerns about corporate compliance
concerns about customisation of applications
concerns about network stability
lack of availability of adequate cloud computing services
do not see the advantages of cloud computing
lack of support from top management
lack of IT skills.
The cost of retooling systems is the most frequently cited obstacle to using cloud computing, both in manufacturing (60%) and ICT (56%) (Figure 3.19). Approximately every second enterprise in both sectors has concerns about customisation of applications, corporate compliance, or network stability. Roughly one-third report that a lack of IT skills – for instance, in cloud engineering – limits their use of cloud computing. Finally, and somewhat surprisingly, a substantial share of enterprises in manufacturing (34%) state that they do not see advantages in cloud computing.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Manufacturers – especially smaller manufacturers – are more likely to report obstacles than enterprises in ICT (Figure 3.20). Nevertheless, the ranking of the various obstacles is similar across both sectors. The importance of enterprise size is particularly pronounced when it comes to cost constraints on retooling systems. While nearly 70% of enterprises with 50 to 250 employees indicate that the cost of retooling limits their use of cloud computing, this is true for only 44% of enterprises with more than 250 employees.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
A salient feature of the survey – not examined in previous studies – is an assessment of the extent to which enterprises use public sector services to support the adoption of AI (Figure 3.21).
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Foremost among the survey findings is that a significant share of enterprises uses such services. The most frequently used services in ICT and manufacturing are those that provide access to information or advice (75% in manufacturing, 69% in ICT). Roughly 58% of enterprises make use of training services provided by the public sector, and around 42% use programmes that promote access to finance, such as tax credits on R&D spending, grants, or credit guarantees.
Public sector services are used most by manufacturers with 50 to 250 employees. For instance, 85% of such enterprises use information or advisory services, compared with roughly 68% for other groups of enterprises.
Among the surveyed countries, enterprises in Japan are the most frequent users of public sector services to assist in using AI (Figure 3.22). This is particularly so for services that provide information or advice (82%). The differences between Japan and other countries are less pronounced for other service types.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Enterprises in the United States are much less likely to use public sector services than enterprises in other countries. For instance, only 19% of enterprises in the United States use services that promote access to finance, as compared with 50% of enterprises in Japan.
Supporting growth in workforce skills in AI
Firms can increase the skills of their workforce in a variety of ways. Enterprises were asked about the usefulness of three support mechanisms, all of which are amenable to change by policy makers. Specifically, enterprises were queried on how helpful the following types of support could be to increase staff skills in AI:
partnerships with educational and vocational institutions
tax allowances or tax credits for training in AI
support to develop qualification frameworks for graduates in the field of AI.
Most enterprises indicate that one or another form of public support would help to strengthen staff skills in AI. Some 84% of enterprises indicate that partnerships with educational and vocational institutions would be either “very useful” or “moderately useful” (Figure 3.23). A similar share states that they would value support in developing qualification frameworks for graduates in the field of AI. Finally, 67% of enterprises indicate that tax allowances or tax credits for training in AI would be “very useful” or “moderately useful” (recall that most of the surveyed enterprises invest in R&D as part of the process of adopting and using AI).
Percentage of enterprises expressing agreement
Note: The industry-size symbols present the sum of “very useful” and “moderately useful”.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
As Figure 3.24 shows, large and small enterprises vary little in their assessment of the utility of support on human capital. Regardless of size, most enterprises consider that all three of the proposed ways of supporting the strengthening of workforce skills in AI are useful. The United States stands out as the country with the lowest share of enterprises that consider the surveyed means of support either moderately or very useful (Figure 3.24).
Percentage of enterprises that consider each type of support either moderately or very useful
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Across the survey sample just over 50% of enterprises use AI itself to facilitate training or to give cognitive support for workers. Such applications frequently combine AI with other technologies, such as augmented (AR) and virtual reality (VR). For example, using AR, workers might view useful information – such as how best to repair breakdowns in complex machine environments – on wearable visors. Among other applications, VR can enable safe and inexpensive “learning by doing”, which is especially useful for beginners in tasks that entail safety risks or using expensive machinery. Using AI to provide training and cognitive support is one of the more recent applications of AI, which underscores the advanced-adopter character of some enterprises in the sample.
Public sector information services to assist in the adoption of AI
Enterprises were queried on how helpful the following types of mostly information services provided by the public sector could be to their use and development of AI:
information on and examples of business use cases in the firm’s industry
information on expected rates of ROI in AI
information on available and reliable technology vendors
information on available and reliable sources of private-sector advice and expertise
certification or accreditation schemes for AI solution providers
information on current or forthcoming regulations around data or AI.
A large majority of enterprises judge that information services provided by the public sector would be “helpful” or even “very helpful” to their use of AI. For any of the services considered, no less than 76% of enterprises indicate that they would be at least “helpful” (Figure 3.25). Fully 83% of enterprises judged that having more information on current or forthcoming regulations around data or AI or on expected ROI in AI would be either “helpful” or “very helpful”.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
It is notable that even in this sample of enterprises that often use AI in advanced ways, additional information on various domains of AI is sought. This suggests that such information may be even more important for firms that do not already use AI. Concerning differences across sectors and size of enterprise, Figure 3.26 shows that smaller manufacturers most often indicate that information services would be “helpful” or “very helpful”. Differences due to enterprise size are much less pronounced among enterprises in ICT.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Except for the United States, the share of enterprises indicating that information services are “helpful” or “very helpful” varies little across countries. In the sample, enterprises in the United States are least likely to consider public services as “helpful” or “very helpful”.
Other public sector initiatives to support the uptake of AI
The views of enterprises were also surveyed on the value of a wider set of public initiatives to foster the use of AI beyond information services. The initiatives considered were:
investing in university education and vocational training in fields related to AI
investing in retraining and lifelong learning for employees who work with AI
improving understanding of AI among government officials
gathering and publishing administrative public datasets
promoting a competitive AI vendor market
upgrading IT infrastructure, such as high-speed broadband
By taking into account the views of enterprises on the utility of identified areas of support, policy makers and organisations might more effectively promote the successful integration of AI technologies.
Most enterprises in the sample perceive all the listed public sector initiatives as “helpful” or even “very helpful” (Figure 3.27).
Percentage of all enterprises
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Among the most widely and highly valued initiatives are those to develop human capital. Some 86% of enterprises consider that initiatives that foster investments in retraining and lifelong learning for employees who work with AI would be “helpful” or “very helpful”. Similarly, 82% of enterprises consider public investments in university education and vocational training in fields related to AI to be “helpful” or “very helpful”. Such initiatives would not only provide students with specialised skills, but they could also contribute to the overall development of a workforce capable of driving innovation in AI. In addition, but slightly less prominent, the surveyed enterprises recognise the importance of enhancing government officials' understanding of AI, with 74% of enterprises rating this either “helpful” or “very helpful”.
Some 78% of enterprises believe that any measures to foster a competitive marketplace for AI vendors would be “helpful” or “very helpful”. By promoting a diverse range of vendors, enterprises might benefit from increased access to cutting-edge AI solutions and services. A reliable IT infrastructure is important for ensuring seamless connectivity and efficient data transfer, thereby facilitating the successful integration and deployment of AI technologies. Public initiatives aiming to upgrade IT infrastructure, such as high-speed broadband, are also supported by 78% of firms. Finally, 73% of enterprises perceive public sector initiatives that aim to gather and publish administrative datasets as “helpful” or “very helpful” for their adoption of AI. This finding emphasises the potential benefits of making administrative public datasets accessible to firms. Such datasets might serve as a resource for training and developing AI algorithms and models.
As depicted in Figure 3.28, the generally positive view towards this sample of possible public sector initiatives varies little in terms of industry and firm size. Even so, there are some subtle differences: smaller manufacturers tend to perceive the surveyed initiatives as “helpful” or “very helpful” most frequently.
Percentage of all enterprises indicating “helpful” or “very helpful”
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
With respect to country of location, somewhat larger differences in enterprises’ perceptions can be detected (Figure 3.29). First, enterprises in the United States are least likely to report that public initiatives are “helpful” or “very helpful”. Nevertheless, even within the United States, most enterprises hold a positive view of the types of initiatives considered in the survey. For instance, roughly 63% of enterprises in the United States state that initiatives to gather and publish administrative datasets are “helpful” or “very helpful”. Overall, France exhibits the highest share of enterprises with a positive outlook on five out of six of the examined public sector initiatives.
Percentage of all enterprises indicating “helpful” or “very helpful”
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Relating the adoption of AI to the use of public sector support
This section examines the relationship between the use of public support and reported difficulties in adopting AI and cloud services, controlling for AI adoption rates.
First, enterprises that use more AI applications are more likely to use all three categories of public support. Large enterprises are less likely to use public support (Table 3.8). For instance, the estimated probability that a medium-sized enterprise will use public information and advice is 0.748, while the probability that a large enterprise will use the same public service is 0.684 (the probabilities come from a post-estimation that is not included in the table). For training services, the predicted probability is 0.640 for medium-sized and 0.508 for large enterprises. For access to finance, the predicted probabilities are 0.469 and 0.361, respectively, for medium-sized and large enterprises. These findings with respect to enterprise size are unsurprising since larger enterprises generally have more resources than medium-sized enterprises with which to resolve the challenges of AI adoption themselves. In addition, small and medium-sized firms are typically the main intended targets of public support for AI adoption. A further message here is that the probability of using these services is quite high overall and, as shown in Table 3.8, increases with the number of AI applications used.
Table 3.8 also shows that enterprises that report more obstacles to using cloud computing and more obstacles to using AI are more likely to use public sources of information and advice but not training services or access to finance and subsidies. A possible interpretation of these results is that lack of information is the biggest obstacle to overcome when adopting AI applications, while support for training and finance is only relevant once an AI application is adopted.
The parameters on the country and industry dummies were not statistically significant in the regressions, suggesting that the patterns of use of public support are similar across the G7 countries as well as across industries.
Note: Probit regressions use robust standard errors, country, and industry fixed effects. Robust standard errors are in parentheses. The three asterisks (***) and two asterisks (**) signify statistical significance at the 1% and 5% levels, respectively. Pseudo R2 signifies the overall explanatory power of the regression, which may take values between zero and unity. The low levels of explanatory power here are typical of probit analysis.
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
Earlier, it was seen that the questionnaire asked respondents to rate the helpfulness of the following five specific public sector services and initiatives:
investing in education in fields related to AI
investing in retraining for employees who work with AI
improving understanding of AI among government officials
gathering and publishing administrative public databases
promoting a competitive AI vendor market
The respondents assigned scores ranging from “very helpful” (score 1) to “not helpful at all” (score 4). These responses were regressed on the number of AI uses and the number of obstacles to using AI reported by the enterprise, controlling for enterprise size, age, and industry (Annex H, Tables 1 and 2 report the regression results). The average score for public sector services and public sector initiatives is around 2, representing “helpful”.
Notably, the scores do not differ significantly with enterprise size or age. The only exception is the public initiative entitled “Investing in retraining for employees who work with AI”, which medium-sized enterprises value more than large enterprises. Furthermore, for this initiative, AI intensity does not affect the enterprises’ evaluation.
Of specific interest here is to what extent enterprises that use AI intensively or face many obstacles to using AI find public services and initiatives more helpful than those that use AI less intensively. The regressions suggest that they do. For instance, one additional AI application at the enterprise level is associated with a score of about 0.1 higher on the usefulness of “Public information on and examples of business use cases” in its industry. The service where experience with AI plays the least important role in terms of usefulness is “Information on available and reliable technology vendors”.
These use patterns, of course, cannot shed light on the actual efficacy of public support; the assessment of these would require other analytic approaches.
Support to facilitate the management of regulatory change
The survey also elicited enterprises’ views on AI-related regulation. Some uses of AI that involve autonomous systems might be detrimental to clients, potentially exposing businesses to legal jeopardy. Enterprises were asked if they favour regulation that helps to overcome such a problem by establishing clear accountability when AI is used. A clear message is that enterprises seek clarity regarding accountability around the safe use of AI (Table 3.9). While the desire for clear accountability is unsurprising, these data underscore the need for policy makers to examine possible ambiguities that regulations might give rise to and how best to communicate information on regulation to firms.
Percentage of enterprises favouring clear regulatory accountability and the perceived usefulness of public services to this end
Source: OECD (2022-23[1]). 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
The OECD/BCG/INSEAD survey focuses on enterprises that utilise AI and compares AI use across G7 countries in two economic sectors (manufacturing and ICT) and two enterprise size classes. The sample size of 840 enterprises, while not statistically representative of national populations, allows for rigorous within-sample analysis. This within-sample analysis focuses on the intensive margin of AI use, exploring the number of AI applications adopted and their relationship to enterprise and industry characteristics.
The survey uses novel questions of direct policy relevance, addressing issues like the usefulness of public policies and support for AI adoption. Because the sample comprises many advanced AI users, the findings will likely become increasingly relevant to policy over time as more enterprises aim to become advanced AI users. However, as stated in Chapter 1, the survey was conducted prior to recent developments in generative AI and its rapidly expanding public use. It remains to be seen how the advent of generative AI affects patterns of AI adoption in firms.
A novel insight from the survey is the widespread positive view of public sector services to facilitate adoption, especially among enterprises facing obstacles to adoption or using AI intensively. Similarly, while initiatives to help develop human capital are highly valued, a non-trivial share of enterprises express uncertainty with respect to the precise skills they need, which suggests the value of developing new qualification frameworks. In addition, to aid in the use and development of AI, collaboration with universities and public research organisations is widespread, which underscores the importance of measures that facilitate these interactions, the need for which is also evident from interviews with enterprises (see Chapter 5).
As with any cross-sectional study, various results merit further examination using other methods. For instance, it would be helpful to better understand causal relationships associated with public sector support. For example, is the tendency for enterprises that use AI more widely to also use public support services driven by their encountering more diverse adoption challenges? Or might it be because more alert enterprise leadership will adopt AI more actively and more actively seek external assistance? Answers to questions such as these could inform decisions around the best allocation of public resources to support AI uptake.
References
[2] Bughin, J. et al. (2017), Artificial Intelligence: The Next Digital Frontier?, McKinsey Global Institute, https://apo.org.au/node/210501.
[7] Calvino, F., C. Criscuolo and C. Menon (2016), “No Country for Young Firms?: Start-up Dynamics and National Policies”, OECD Science, Technology and Industry Policy Papers, No. 29, OECD Publishing, Paris, https://doi.org/10.1787/5jm22p40c8mw-en.
[9] Cho, J. et al. (2023), “What’s driving the diffusion of next-generation digital technologies?”, Technovation, Vol. 119/102477, https://doi.org/10.1016/j.technovation.2022.102477.
[8] Haller, S. and I. Siedschlag (2011), “Determinants of ICT adoption: Evidence from firm-level data”, Applied Economics, Vol. 43/26, pp. 3775–3788, http://www.tandfonline.com/doi/full/10.1080/00036841003724411.
[3] Kinkel, S., M. Baumgartner and E. Cherubini (2022), “Prerequisites for the adoption of AI technologies in manufacturing – Evidence from a worldwide sample of manufacturing companies”, Technovation, Vol. 110, p. 102375, https://doi.org/10.1016/j.technovation.2021.102375.
[10] Nolan, A. (2021), “Artificial intelligence, its diffusion and uses in manufacturing”, OECD Going Digital Toolkit Notes, No. 12, OECD Publishing, Paris, https://doi.org/10.1787/249e2003-en.
[11] OECD (2023), “based on the OECD ICT Access and Usage by Businesses Database, http://oe.cd/bus The OECD Going Digital Toolkit”, The OECD Going Digital Toolkit, https://goingdigital.oecd.org (accessed on 26 July 2024).
[5] OECD (2023), ICT Access and Usage by Businesses (database), https://stats.oecd.org/Index.aspx?DataSetCode=ICT_BUS (accessed on 25 July 2023).
[1] OECD (2022-23), OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
[6] Statistics Sweden (2021), Artificial Intelligence in Sweden, http://www.scb.se/contentassets/048c2c293c404f3e899e91b844b6b9c2/artificiell-intelligens-i-sverige-2019_slutrapport.pdf.
[4] Zolas, N. et al. (2020), Advanced Technologies Adoption and Use by U.S. Firms: Evidence from the Annual Business Survey, National Bureau of Economic Research, Cambridge, MA, https://doi.org/10.3386/w28290.
← 1. In future work, an alternative question might be tested, such as “Has your enterprise implemented a system of data governance?” Data management solutions can be local to individual divisions or activities within a business. Data governance, however, is essentially corporate-wide, and might therefore serve as a good proxy for data maturity overall. Thanks are expressed to Benoit Bergeret for this observation.