This chapter reports the overall findings of the OECD/Boston Consulting Group/INSEAD study, conducted in 2022-23, that explored the adoption of artificial intelligence in firms and how governments can support this. Beyond an opening review of prior research, the core of the study is a novel policy-oriented survey of enterprises implemented in the G7 countries and Brazil. This is complemented by studies of public sector institutions that help the diffusion of technology as well as structured interviews with enterprises. This chapter also sets out the main policy implications.
The Adoption of Artificial Intelligence in Firms
New Evidence for Policymaking
Abstract
Increasing the rate of growth of economic productivity is one of the greatest policy challenges facing OECD countries. OECD countries have experienced a decades-long period of stagnating productivity. Raising productivity is critical to raising living standards and enabling countries to cope with the consequences of rapid population ageing. Across the economy, artificial intelligence (AI) could be an important source of the needed productivity increases.
This study is the fruit of a partnership between the OECD, the Boston Consulting Group (BCG) and INSEAD Business School. It combines several types of data and information to explore AI adoption in firms and how governments can support this. Beyond an opening review of prior research, the core of the study is a novel policy-oriented survey of 840 enterprises across the Group of Seven (G7) countries plus 167 in Brazil. The survey in G7 countries was conducted between November 2022 and January 2023. The survey in Brazil was implemented between February and July 2023. Complementary insights came from structured interviews with leaders of 19 major public institutions from G7 countries and Singapore that work to accelerate the spread of digital and other technologies, including AI. These include organisations like Germany’s Fraunhofer IAO/IPA, the United Kingdom’s Digital Catapult and the United States’ Manufacturing Extension Partnership programme.
A further element of the study is a synthesis of findings from in-depth interviews with managers in enterprises adopting AI. These experts hold positions such as chief information officer, chief technology officer, head of R&D, and chief technology officer, among others. The interviews serve to test the survey findings and further elaborate what the private sector most seeks from government.
The 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises has several distinctive features. The first is a significant focus on policy. This entailed developing, testing and including novel survey questions on topics such as enterprises’ views on the value of public policies relevant to AI uptake and their priorities for future policy. Other novel questions seek information on familiar topics, such as enterprises’ use of cloud computing, but with new emphases – in this case, probing why cloud computing, an important adjunct to AI, might be underutilised.
A second distinctive feature of the survey is that it provides standardised data on AI in firms across G7 countries, plus Brazil. Other standardised supranational surveys focus only on countries in the European Union. While the survey covers a numerically small sample of enterprises and countries, G7 countries play an outsized role in AI globally. By one estimate, for example, France, Germany, and the United Kingdom alone account for around one-half of all AI talent in Europe (LinkedIn Economic Graph, 2019[1]). Similarly, among OECD countries, the United States is the dominant source of venture capital investment in AI-related early-stage firms (Tricot, 2021[2]).
An additional feature of the survey is its exploratory character. A longer-term goal is that some survey questions, or variants thereof, which all underwent cognitive testing, might be incorporated in subsequent surveys performed by national statistics offices (NSOs) or supranational bodies. Indeed, NSOs and other organisations with experience in the design and administration of large-scale surveys of AI in firms helped design the OECD/BCG/INSEAD questionnaire.
The book has six chapters. This opening chapter reports the key findings from all elements of the study. Chapter 2 provides an overview of previous survey-based research on AI in firms, focusing on the extent of adoption in the business sector and the barriers to uptake. Chapter 3 reports the findings of the 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises. Chapter 4 reports the results of structured interviews with public and public-private organisations that work to assist firms’ use of digital technologies, including AI. Chapter 5 summarises the key policy-relevant insights from interviews with C-suite managers and technical experts in firms that have adopted AI. Finally, Chapter 6 presents the results of a survey of enterprises in Brazil using essentially the same OECD/BCG/INSEAD questionnaire.
The remainder of this chapter has several subsections. These are dedicated, sequentially, to the goals of the study; choice of enterprise size and sector; novel survey content; implementation and statistical features of the survey; a summary of previous research literature; a description and analysis of the survey findings; summaries of the main insights gleaned from structured interviews with diffusion institutions and senior staff in enterprises; and the key findings from the survey in Brazil. The chapter concludes with an overview of the principal policy insights from the entire study.
An important point to note is that the survey was developed and administered before the November 2022 release of ChatGPT. ChatGPT and successor large language and/or multimodal models (LLMs and LMMs) promise many uses in business, for instance, in programming, creating intuitive conversational interfaces, handling customer enquiries, generating maintenance manuals, integrating and transferring data and information from across a business when the former exist in otherwise incompatible formats, multi-lingual communication with suppliers, and so on.
The speed of adoption of LLMs and LMMs, their business impact, and the challenges involved in deploying are all pressing questions. This study's view on the state of AI adoption as of early to mid-2023 provides a critical perspective on what was already happening prior to the latest generative models. This is a strong starting point for researchers trying to ascertain the adoption trajectory of the latest family of AI technologies, especially insofar as their technical characteristics may change the factors that have aided or hindered AI adoption in the recent past.
In recent decades, nearly all OECD countries have experienced a decline in the rate of growth of economic productivity, that is, the balance between the volume of economic inputs and outputs. Productivity growth is the main driver of rising incomes and living standards. Across OECD countries as a group, labour productivity growth in 2013‑19 was only around half of that in 1996‑2001 (Figure 1.1). New technologies, including AI, hold the promise of raising labour productivity (later sections of this chapter describe several examples of how).
Note: To maintain consistency of the panel over time, Korea and Estonia are excluded from the OECD average (not available for all years in the first sub-period considered). The period 2008‑13 is excluded as it is largely influenced by the Great Financial Crisis and the European debt crisis.
Source: OECD Productivity Database, https://www.oecd.org/en/publications/serials/oecd-productivity-statistics_g1g72f69.html.
Increasing productivity is both a short and long-term imperative. In the short term, the economic aftermath of the coronavirus (COVID‑19) pandemic and the ongoing disruption of the war in Ukraine have increased the importance of productivity growth. Over the longer term, the productivity challenge could become more urgent still, owing to the economic and social consequences of demographic change. Old-age dependency ratios – the number of people older than 65 years per 100 persons of working age – are projected to at least double in most Group of Twenty (G20) countries by 2060. This means that those who are working will need to become more productive still, to offset the fact that they will be fewer in number (other things, especially immigration, unchanged). The problem will be made more acute because more young people may leave the workforce to care for ageing parents. In addition, more of society’s resources will be dedicated to retirement-related transfers and healthcare. OECD countries will adapt, for instance, by raising the age of retirement. However, without improvements in labour productivity, which AI could help bring, these developments could cause severe economic stress.
A further reason why it matters to better understand AI adoption has to do with the labour market. In recent years, many studies have sought to estimate the effects of AI on labour demand. Many studies have suggested significant disruption as more cognitive tasks workers perform are substituted by AI. An early example of this work is (Frey and Osborne, 2017[3]), which examined the probability of computation of over 700 occupations, concluding that 47% of total US employment is attributable to occupations potentially automatable over one or two decades. Other studies have used other methods and arrived at different estimates. For instance, Arntz, Gregory and Zierahn (2016[4]) account for the distribution of automatable tasks within occupations and conclude that across 21 OECD countries, around 9% of jobs are automatable. (Lassébie and Quintini, 2022[5]). focus on the automatability of skills and abilities, which they then link to occupations. They find that, on average across OECD countries, occupations at highest risk of automation account for about 28% of jobs.
More recently, in research for the International Monetary Fund, (Cazzaniga et al., 2024[6]) add to studies using a purely task-based approach by also examining whether AI complements or replaces job roles. The authors conclude that almost 40% of global employment is exposed to AI, rising to around 60% of jobs in advanced economies, owing to the greater presence of cognitive-task-oriented jobs. However, a main point with respect to survey work, of which this report is a part, is that the results of such studies hinge on assumptions about the rate of adoption of AI and other forms of automation. The studies cited above are concerned with what is automatable in principle. However, as the literature review in Chapter 2 and the data in this study from Brazil show, adoption at the aggregate level has been relatively limited to date.
Data and information from this study could also help inform the ongoing development of practical guidelines for the implementation of the OECD Recommendation on Artificial Intelligence (the OECD AI Principles) (OECD, 2019[7]). This is particularly so for those recommendations that have to do with policies fostering AI-related innovation, entrepreneurship, and productivity growth in firms.
Avoiding duplication of existing survey evidence
The starting point for deciding the content and scope of the current survey was to consider what prior national and supranational surveys already revealed about the diffusion of AI in firms. Over the past decade or so, several national and supranational institutions and other public and private organisations have conducted surveys to investigate the extent of AI uptake and some of its key characteristics. These surveys have different sample coverages and designs and yield different insights. However, as detailed in Chapter 2, several messages consistently emerge from this work, including that:
The extent of diffusion of AI in the business sector is generally low but still varies considerably across countries.
Small enterprises use AI much less frequently than larger enterprises.
Certain sectors, such as financial services and information technology (IT) services, consistently register the highest shares of AI use in firms.
Accordingly, designing a survey to replicate such facts would be redundant. In addition, these findings are based on survey sample sizes beyond the financial resources available for the current exercise. The same goal of avoiding duplications informed the choice of questions about several common barriers to AI adoption, often broached in public and private-sector surveys as well as case studies. Chief among these standard questions is the availability of skills and the cost of adopting AI.
One upshot of the aim of avoiding duplication is that the survey includes AI users only (rather than seeking to assess the aggregate extent of AI use in the corporate sector). Furthermore, the survey only includes enterprises that classify themselves as active rather than passive users of AI, again a difference from many national surveys. It is hoped that understanding the policy needs of active users will provide useful insights into measures to assist other enterprises as they become or aim to become active AI users. Annex A compares selected features of recent national and supranational surveys of AI in firms with the 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
The survey sought and achieved completed responses from 840 AI-using enterprises across the G7 countries. Completed responses were sought from 120 enterprises in each country. In each country, the survey was addressed to enterprises in two size classes: medium-sized enterprises (between 50 and 249 employees) and larger enterprises (250 or more employees). The survey focused on enterprises in two sectors: manufacturing and information and communication technology (ICT). In addition, only two ICT sectors were considered: International Standard Industrial Classification of All Economic Activities (ISIC) 62: Computer programming, consultancy, and related activities and ISIC 63: Information service activities (activities relating to the manufacture of devices and components, such as semi-conductors, used in data and information processing and communication were not included under “ICT”). Table 1.1 shows the size and sectoral distribution of respondent enterprises per country. Only active users of AI were considered. The next section explains why these sectoral and enterprise size classes were chosen.
The survey has an exploratory character. Budgets permitting, it could eventually be implemented, with possible revisions, across a wider set of countries, sectors, and enterprises, while also using a sampling frame and probabilistic method allowing generalisation to national populations of enterprises. Doing so would strengthen cross-country and cross-firm statistical analyses.
Reasons for focusing on manufacturing and ICT
Manufacturing is a priority sector for many OECD and non-OECD countries: national initiatives for advanced manufacturing have proliferated in recent years. Some examples are Germany’s Industry 4.0 programme, the National Network for Manufacturing Innovation in the United States, Japan’s Robot Strategy, and the People’s Republic of China’s Made in China 2025 and Internet Plus initiatives. Indeed, over 30 countries have developed national programmes for Industry 4.0, while many more have prepared manufacturing foresight studies and strategies, as well as in-depth roadmaps for manufacturing technologies deemed strategic. Manufacturing has also grown as an area of emphasis in recent national research and innovation strategies.
In addition, there are many uses of AI in manufacturing, and the potential for new applications is large. Early forms of AI – known as expert systems – have, in fact, been a part of manufacturing for over 40 years. However, their use was limited to just a few applications, such as process scheduling. Newer types of AI, which learn and make predictions from data, now have a role in many business processes, including:
Industrial research: A compelling example comes from Boeing, which wished to mass-produce 3D-printed metal parts for jets. However, most useful metal alloys are not printable because the different powder grains do not stack well. Boeing turned to an AI system belonging to Citrine Informatics. The AI trawled through decades of experiments, scanning 10 million possible recipes for alloy powders. Citrine Informatics wrote software so the AI could even scan data from old reference books and handwritten notebooks. The process of discovery of materials, which usually takes years, was shortened to days (WIRED Chen, 2017[8]).
Product design: AI-driven design software, combined with 3D printing, is revolutionising industrial design. Such software can generate vast numbers of potential designs, selecting those best suited to requirements. A novel aircraft bulwark partition was developed and incorporated using such a system in Airbus’ A320 aircraft. The new partition was stronger than the one it replaced and 45% lighter (Airbus, 2016[9]).
Fabrication and assembly: Another important use of AI is in quality control. In the semi-conductor industry, for example, defects in computer chips can appear as irregular shapes on otherwise regular circuit patterns. The irregular shapes attract feature detectors driven by AI.
Process control: “Digital twins” are computer models of a machine, or system of machines, based on real-time data from sensors in machinery. Aided by AI, the computer models help to monitor production, optimise key parameters and predict maintenance needs.
Supply chain management: BMW (Bayerische Motoren Werke AG) has set a goal of knowing the real-time status of all major production equipment at each company producing key components for its vehicles. Information of this sort can extend upstream to the supply of production inputs and downstream to distribution and retail. AI can help integrate and improve supply chains, for instance, by predicting fluctuations in customer demand and efficiently scheduling distribution.
Training and cognitive support: In aerospace, when building its A350 aircraft, Airbus deployed AI to analyse process disruptions. If a worker encounters an unfamiliar problem, the AI can suggest solutions by analysing a mass of contextual data on similar problems from other shifts or processes. The AI cut time lost to disruptions by a third (Ransbotham et al., 2017[10]). AI is also enhancing workforce training (using virtual reality) and cognitive assistance (using augmented reality). For example, a technician might see suggested solutions to production problems projected on a safety visor.
AI can also support generic business functions that matter to manufacturers and firms in other sectors. An example is digital security. Digital security incidents appear to be increasing in sophistication, frequency and impact and intensified during the COVID‑19 pandemic (OECD, 2020[11]). In one incident, hackers broke into the computers of a German steel mill and overrode the shut-off mechanisms on the steel mill’s blast furnace. Among other advances, AI systems can recognise when text is likely part of a password, helping to avoid accidental online dissemination of passwords. As occurs in many sectors, AI in manufacturing can also be applied in customer-facing processes such as business-to-business (B2B) marketing and pricing.
Unlike sectors such as insurance and finance, which for many years have led in using big data, manufacturing has traditionally been product-led, with a less prevalent culture of and familiarity with data analytics. However, manufacturing is one of the most data-intensive parts of the economy. As the understanding of how to create value from industrial data grows and as AI and data analytics practices spread within manufacturing, productivity gains could be significant (Atkinson and Ezell, 2019[12]). The automotive sector is an example of where AI in manufacturing is becoming transformative (Box 1.1).
AI is fundamentally changing the automotive industry, both in production and in the vehicles themselves. In production, among other applications, AI is helping manufacturers significantly reduce the time needed for design approval and authorisation. Combined with the Internet of Things (IoT), AI enables predictive maintenance and transforms quality control. For instance, with the help of computer vision and machine learning, manufacturers can detect even minor imperfections in vehicle components. Globally, the automotive sector has led advances in industrial robotics. AI is helping make such robots more autonomous, for instance, in efficiently and accurately picking parts. In a particularly advanced application, AI-enhanced robots have been used to co-ordinate workflows in combined human-machine teams.
AI is also revolutionising automotive logistics. By ensuring the right parts are available at the right locations and times, companies can minimise inventory holding costs and reduce the need for expedited shipments. As a result, some automotive companies have achieved up to a 30% reduction in logistics-related costs. AI is also used to identify when customers might be willing to purchase an upgraded product or service or related products or services.
Vehicles themselves are evolving due to AI. Using cutting-edge AI algorithms, vehicles can analyse extensive sensor data – from cameras, lidar and radar – to perceive their surroundings and make smart decisions. Over 80 companies in the United States alone are currently testing self-driving cars. Cars are also becoming software platforms, facilitated by AI. Many automakers deliver over-the-air software updates to vehicles, and some cars transmit enormous volumes of data back to manufacturers. AI-based car occupant monitoring systems are set to increase passenger safety, for instance, by monitoring the interiors of cars and ensuring driver attentiveness. AI-powered security systems, such as lane departure warning and autonomous emergency braking, are already enhancing safety, while seamless communication between cars will help maintain safe distances on the road. In addition, AI will affect automotive insurance, for instance, by gathering incident data to complete claim forms.
The industry faces several challenges related to data, technical limitations and regulation. For example, obtaining sufficient clean data on rare events is a problem. By facilitating large-scale data collection, connected fleets will help address this limitation, but other approaches are also needed. As AI systems require access to sensitive vehicle and user data, manufacturers must also employ secure data-handling practices, including encryption and access control, to safeguard data during transmission, storage and processing.
Source: Based on analysis from Eugene Hayden, Boston Consulting Group, drawing on Appinventiv, “AI in the automotive industry”, https://appinventiv.com/blog/ai-in-automotive-industry/; Boston Consulting Group, “Auto.AI research”, http://www.bcg.com/beyond-consulting/bcg-gamma/auto-ai; TaskUs, “Future trends in autonomous vehicles”, http://www.taskus.com/insights/future-trends-autonomous-vehicles/.
Compared with many other parts of the business sector, enterprises in ICT are major users and developers of AI, especially enterprises in the ICT subsectors examined in this survey (to recall, these are: ISIC 62: Computer programming, consultancy, and related activities; and ISIC 63: Information service activities). Table 1.2. shows data on the share of enterprises that use AI in manufacturing and ICT drawn from four national surveys. Precise comparison across countries is hindered because the included areas of economic activity do not overlap perfectly (the Swedish survey included a wide range of publishing activities). Nevertheless, in all the surveys the share of AI-using enterprises in ICT is far higher than in manufacturing; in Canada it is more than ten times higher.
Note: Sectoral classification for Canada: Manufacturing = NAICS 2017 31-33; ICT = NAICS 2017 541512
Sectoral classification for Sweden: Manufacturing = SNI_2007: 10-33; ICT = SNI_2007: 58-63
Sectoral classification for Germany: Manufacturing = NACE_Rev2: 10-12, 14-15, 31-32, 13, 16-18, 22-23, 20-21, 24-30, 33; ICT = NACE_Rev2: 61-63.
Source: Rammer, C., D. Czarnitzki, and G.P. Fernández, “Artificial Intelligence and Industrial Innovation: Evidence from Firm-Level Data (2021)”, https://ssrn.com/abstract=3829822; Destatis, “IKT-Indikatoren für Unternehmen: Deutschland, Jahre, Wirtschaftszweige (ICT indicators for companies: Germany, years, economic sectors)”, https://www-genesis.destatis.de/datenbank/online/statistics, accessed on 3 June 2024.
Enterprises in ICT also use AI in a wide range of applications. This is particularly so for businesses in software development and programming, as well as those in data processing, hosting and online platforms. Such enterprises often use AI in the same generic business functions that a manufacturer might. Other uses can include:
Process automation: AI is simplifying the work of programmers. AI tools can write working code, build on billions of lines of public code, learn to arrange code fragments and aid code completion. Some AI tools provide programmers with feedback as they type, suggesting alternative code that programmers can edit as they please. AI can also learn from and adapt to programmers’ coding habits and preferences.
Quality control: AI helps software developers test code and identify defects more efficiently and quickly than human inspection alone. AI can also process data on how software is used across devices, users in different population groups, and locations.
Social media analysis: AI can gather and process large volumes of data from social media and use the results to help predict market demand and customer behaviour.
After-sales services: Service desks can be highly automated. For example, AI can draw on historical data from across a company to help provide users with solutions to problems. An AI might interpret a user’s service-related questions, seek similar service questions and answers from company records, and propose response options most likely to be correct. Overall, service management can be made faster, cheaper and more effective.
With the above considerations in mind, Annex B shows the business sectors targeted by the survey. Because countries use different classificatory systems, Annex B also maps the codes from the ISIC to the national classification systems used in individual G7 countries and Brazil.
Reasons for focusing on medium-sized and large enterprises
The survey targeted enterprises in two size classes: medium-sized enterprises (50‑249 employees) and large enterprises (250 or more employees). Small enterprises (with 0‑49 employees) were not sampled. Analytic and budgetary reasons motivated the choice of targeted size classes.
As described in Chapter 2, in all OECD countries, AI use rates are very low in core business processes in small enterprises (i.e. with fewer than 50 employees). This raises the administrative cost of a survey aimed at small enterprises that are AI users. Moreover, it was judged that for several reasons, less policy analytic value might come from focusing on small enterprises. In addition to AI being much less prevalent than in larger enterprises, smaller enterprises tend to use AI in less sophisticated ways. Moreover, the reasons for their non- or limited use of AI are relatively well understood and often stem from inherent problems of scale, irrespective of technology. For instance, smaller firms tend to be more financially constrained when considering investment in AI. However, it is a common feature of small enterprises that many of the investment decisions they face – not just with respect to AI – involve financial challenges coming from, for instance, the difficulties of managing cash flow when product lines are few. Similarly, on account of scale, smaller enterprises generally have a more limited internal division of labour, meaning that functional specialisations such as research and development (R&D) are less developed. Consequently, an alternative focus on medium-sized and large enterprises, which are more likely to have adopted AI in core business processes, and in more far-reaching ways, could provide richer insights for policy makers regarding the adoption process. Such insights would become more relevant over time as today’s numerically larger group of small firms seek to adopt AI.
A reason for including medium-sized enterprises in the sample, rather than focusing exclusively on large enterprises, is that they are often a target of government policy to accelerate the diffusion of new technology. Tracking over 700 national AI policy initiatives from 60 countries, the OECD AI Observatory shows that around 15% of AI-related policies specifically target small and medium-sized enterprises (SMEs) (with some 4% aiming at large firms) (OECD, 2024[15]). It is unclear how many of those initiatives balance towards smaller enterprises, but supporting medium-sized enterprises is clearly a goal. In addition, a reason for focusing on medium-sized manufacturers is the possibility that in the relatively recent process of developing national AI strategies, the specific needs of this group of enterprises may have received insufficient attention (Bergeret, 2020[16]).
With respect to the classification of enterprises as medium-sized, the definition used in the European Union relies on two variables: the enterprise should employ between 50 and 249 persons and have an annual turnover not exceeding EUR 50 million and/or an annual balance sheet total not exceeding EUR 43 million. The OECD also uses the EU definition. However, the definition of an SME used by national authorities differs across non-European countries. In Japan, a country included in this survey, an enterprise is considered an SME if it employs from 4 to 299 employees. In the United States, SMEs include firms with fewer than 500 employees. Nevertheless, selection criteria based on enterprise turnover were not used in the OECD/BCG/INSEAD survey, as requests for such information could dissuade some potential respondents from participating (by contrast, surveys that request financial information often entail a legal obligation to respond when conducted by NSOs).
Novel content in the 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises and areas of overlap with prior surveys
The choice of topics covered in the survey reflects extensive consultation with independent experts, NSOs and delegates in OECD policy committees and working groups. As noted earlier, the aim is not to gauge the overall incidence of AI use in enterprises in G7 economies but to explore the policy and institutional conditions that affect the application and development of AI.
Annex C contains the full survey questionnaire. The questionnaire begins with a series of screening questions. These establish the enterprise’s location, size, and sector, as well as whether it uses AI, whether the use of AI is active or passive, and how AI is applied. Enterprises that meet the screening criteria then proceed to 19 questions, each with multiple response options. These 19 questions examine the following topics:
How important is AI to the enterprise: A question gauges if AI is critical to the enterprise, one among several important considerations, or of minor importance.
Data infrastructures: Questions explore sources of data acquisition for AI, enterprise-level data maturity, the types of professional roles in the enterprise relevant to data and information processing, the number of employees working in those roles, and the reasons for limited or non-use of cloud computing.
Building AI capabilities: Questions examine practices used to adopt AI, including partnerships with universities and public research organisations, the use of public services to help adopt AI, obstacles experienced in adopting AI, recent experience recruiting graduates, and understanding of required skill sets.
National policies and regulation: Questions consider the usefulness that enterprises accord to different public measures to strengthen staff skills, the usefulness for adopting AI of different public policies and support services, and awareness of and views on the utility of various types of regulation.
Research and innovation: A question examines the share of the enterprise’s R&D related to AI and whether the enterprise invests in R&D at all.
Relative to prior cross-country work, as well as most national surveys, the 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises asks novel – or infrequently put – questions, including on the following subjects:
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, 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 can be used in staff recruitment and human resource management, training and cognitive support for workers, customer-facing services and other functions. Thus, enterprises may exploit economies of scope associated with AI technologies, using them in several business functions once they are introduced.
Not just the roles of skills as a barrier to adoption but the less frequently addressed question of whether enterprises fully understand skills needs and if formal academic qualifications fully and clearly align with job requirements.
The incidence of use of different types of public services useful to adopting AI.
Enterprises’ preferences and needs concerning the types of services provided by the wide range of public programmes in all OECD countries to accelerate the diffusion of technology in firms (see Chapter 4).
The use of, and types of, collaboration with students and faculty in tertiary education institutions and partnerships with public research organisations.
Enterprises’ prioritisation of public policies in areas ranging from data to the regulatory environment.
The reasons for non- or limited use of cloud computing (rather than the frequently surveyed use of cloud computing).
Firms’ allocation of their R&D spending to AI.
The survey also includes questions on topics explored in prior firm-level surveys. Such questions cover essential characteristics of enterprises’ use of AI and the challenges faced in developing AI applications.
Implementation and statistical features of the 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises
The available budget determined the survey’s sample size and sampling process. During implementation – between November 2022 and January 2023 – the survey respondents were identified as AI-using enterprises from a pool of enterprises with a high probability of being AI users (drawing on a roster held by a survey administration company). The survey consists of 840 AI-using enterprises, a significantly higher number of observations than most studies focusing on AI-using enterprises.
However, with 120 observations per country and using a sampling approach that relied on a survey provider database with non-representative statistical characteristics and a search procedure selecting AI users, the findings and the underlying sample are not representative of the population of enterprises in each country. In other words, the results relate to averages among the surveyed enterprises and are not directly generalisable to the respective population of enterprises within a given country. However, the survey in Brazil, reported in Chapter 6, used essentially the same questionnaire but was conducted using a probabilistic sampling procedure, yielding results statistically representative of 2 561 enterprises in the State of São Paolo. A fuller discussion of how the survey was implemented and the statistical character of the results is given in Annex D.
According to the literature, business uptake of AI applications is still relatively low and mainly occurs in larger firms and in the ICT, finance, and insurance sectors
Chapter 2 discusses the findings of prior survey-based research on the diffusion of AI in firms. The focus is on AI in manufacturing and in ICT services, the same sectors covered by the 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises.
The prior literature shows that AI adoption is generally the exception across firms, with single-digit adoption rates being common in several sectors in many countries. For example, a major 2020 study in the United States across all AI-related technologies and all firm types found that the overall adoption rate was just 6.6% (Zolas et al., 2020[17]). In the European Union, a 2023 study showed Denmark and Finland have the highest share of enterprises utilising at least one AI technology, both at around 15%. The European Union average was 8%. Italy and France had shares of 5% and 6%, respectively (Eurostat, 2023[18]). A separate study in 2019, covering all firms in Germany, found that only 5.8% used AI (Rammer, Fernández and Czarnitzki, 2022[13]). In Brazil, a 2021 survey showed that 13% of companies use some type of AI (Brazilian Network Information Center, Brazilian Internet Steering Committee, 2022[19]).
Most surveys reveal a strong positive correlation between firm size and AI adoption. For instance, a 2022 study in the United Kingdom found that 15% of small firms were adopting AI, compared to 34% of medium-sized firms (Evans and Heimann, 2022[20]). Only 2% of small firms, 5% of medium-sized firms, and 9% of large firms were piloting AI. The same study showed that larger firms that adopt AI are also more likely to adopt multiple AI technologies. In Japan, in a survey from 2021, the rate of use of AI in firms in the 100‑299 employee size class was half that of firms with 1 000 to 1 999 employees (10% and 22% respectively) (MIC, 2021[21]). Similarly, the incidence of AI adoption in firms with 1 000 to 1 999 employees was less than half that of firms with 2 000 or more employees (at 22% and 48%, respectively).
Numerous factors could explain why large firms generally lead in adopting AI. For instance, large firms often serve larger markets, which allows them to spread the fixed costs of using AI in production over more sales, thereby lowering unit production costs. Similarly, large firms often offer superior workplace conditions, making attracting and retaining talent easier. However, even in large firms, adoption can be surprisingly limited. For example, Chapter 2 describes a 2019 survey of 60 manufacturers in the United States with annual sales of between USD 500 million and USD 10 billion. Over half indicated they were only at the initial stages of manufacturing digitalisation. At the time, just 5% of the companies had mapped where AI opportunities exist and developed a clear strategy for sourcing the data that AI requires. Furthermore, 56% had no plans to do so (Atkinson and Ezell, 2019[12]).
Most surveys concur that AI is more prevalent in some sectors than others, with the highest uptake rates generally being in ICT, finance and insurance, legal, and other professional and technical services (such as engineering, advertising, design and consulting).
While the existing literature shows use to be low overall, the figures on adoption rates cited above, and elsewhere in Chapter 2, vary considerably across countries. This points to a need to better understand the comparability of surveys across countries, as different methodologies may lead to discrepancies. Potential sources of discrepancy include the units of observation used (e.g. enterprises or establishments), survey sample sizes, choice of industries examined, and the size of sampled firms. Another source of variation could come from differences in the questionnaires themselves, such as how AI is defined, as well as differences in the wording and scope of the questions asked (Montagnier P. and Ek I., 2021[22]).
Chapter 2 also discusses why the adoption of AI is often limited. Repeatedly identified obstacles include a lack of digital and data readiness (including interoperability between equipment, which affects the data integration necessary for AI applications), the challenge of creating new business models, the cost of implementing AI, uncertainty about how to use AI solutions to solve specific challenges, uncertainty over the returns on investment (ROI) in AI, how to implement effective change management strategies, and lack of access to suitable and specialised vendors of AI solutions.
Acquiring or developing a skilled workforce is a common problem. About 85% of firms responding to an EU-wide survey indicated that hiring new staff with the right skills was the principal barrier to AI adoption (Kazakova et al., 2020[23]). Even among a sample of large US manufacturers with ample recruitment budgets, 47% lacked the skills necessary to implement AI technologies (Atkinson and Ezell, 2019[12]). These issues are discussed in greater detail later in this chapter and in Chapters 3 and 4.
The OECD/BCG/INSEAD survey covers enterprises that actively use AI and reveals the main characteristics of their AI adoption patterns, the challenges they face and their assessments of the role of public support
Data maturity
The surveyed enterprises were relatively data mature. Most – 78% – used at least one data management solution. Smaller enterprises were a little less likely than others to use a data management solution. In addition to data generated internally, between 51% and 61% of enterprises used external data, whether from private data providers (such as organisations dedicated to producing and selling data), a partner enterprise, or the public sector.
Before adopting AI, firms often need to implement digital technologies that systematically gather data, whether from business processes or interactions with customers and suppliers. High-quality and sufficiently voluminous data are essential to create, test, evaluate and validate AI models. As discussed in Chapter 4, agencies like Canada’s Vector Institute, France’s Cap Digital, Japan’s New Energy and Industrial Technology Development Organization, and others can support the uptake of AI in business. Box 1.2 reflects the views of such organisations on how enterprises gather and manage data and how this affects the adoption of AI.
In their day-to-day work, diffusion institutions develop a deep familiarity with how the characteristics and behaviours of firms affect their attempts to adopt AI. A recurrent problem that diffusion institutions note is that many firms often do not possess data with sufficient quantity, quality, cleanness and structure. They frequently lack an adequate understanding of what information needs to be gathered systematically. Consequently, they may not have the necessary data collection mechanisms in place, or if they do, they struggle to assess how appropriate their data are for a given AI use case.
In addition to collecting the necessary data, firms face data management challenges. They often have to integrate data from different sources, such as software, machines, business areas within the firm, and data provided by third parties. Data sources can vary in periodicity (e.g. weekly, daily, hourly), type (e.g. quantitative or qualitative) and format (e.g. Excel spreadsheets, MySQL databases). Data can be unstructured, unlabelled and disorganised, making it challenging to integrate. Time and expertise are required to prepare data to build an AI model.
Data transfer and exchange also raise challenges. For instance, for fear of losing the value of the data they collect, companies are sometimes unwilling to sell it or to enter collaborative projects that exploit it. Data security (i.e. avoiding data breaches) and regulatory compliance are further concerns. Some enterprises are reticent to publish data or other results from work with diffusion institutions or from publicly funded applied research. This aversion can make partnering with research institutions difficult, as academics often want to publish their research.
Source: OECD/BCG/INSEAD interviews with senior staff from technology diffusion institutions.
The most frequent and infrequent uses of AI
Among the uses of AI examined in the survey, AI for R&D was the most likely and consistently used application across enterprises in any given industry and across industries overall. AI was least likely to be used in human resources management. The low frequency of use in human resource functions is perhaps unsurprising, as many enterprises have concerns about inadvertent misapplication of AI in recruitment, a possibility often raised in public discussion of AI.
How enterprises adopt AI
Enterprises were asked about the practices they use to adopt and develop AI. Most use several mechanisms. More than 70% of enterprises in both manufacturing and ICT report that they carry out R&D on AI technologies for their own use. Nearly three-quarters of enterprises in both sectors rely on employee training. Large enterprises in ICT are the most likely to train employees and hire staff to develop AI. In addition, more than 60% of the sampled enterprises hire new staff to help develop AI technologies. Between 53% and 64% of enterprises use customised systems built by third parties or purchase off-the-shelf software or hardware. About every second enterprise has institutionalised AI development by creating a senior management role or a team with responsibilities for AI. Establishing such senior functions and responsibilities can help promote understanding of AI across a business and help to implement some of the systemic changes that adopting AI can require (Box 1.3). Finally, many enterprises speed up the uptake of AI through partnerships with national or international enterprises that have AI capabilities.
As described in Chapter 4, even when firms have some familiarity with AI, managers often do not have a sufficient grasp of what AI is, what adoption entails, or what their businesses can gain from it. Many managers have a plug-and-play conception of adoption, expecting AI to be a commodity technology they can easily integrate into core business processes.
Compared to adopting other digital technologies, adopting AI can require a significantly larger company-level transformation involving changes to business operations across various departments. If managers lack AI literacy, they may fail to foresee or be unprepared to make the shifts in organisational structure, business processes and culture needed to adopt AI solutions. Many companies also run AI pilots without a strong vision or business plan to expand and integrate them more widely.
In addition, companies often fail to understand the extent of continuing investments required for AI quality management. Keeping AI models performing well over time requires constant assessment, retraining (with the most recent data) and redeployment.
Source: OECD/BCG/INSEAD interviews with enterprises.
In the survey sample, and perhaps unsurprisingly, spending on R&D for AI as a share of all R&D spending was positively related to how critical enterprises deem AI to be. 38% of enterprises that allocate between 0-10% of their R&D spending to AI considered AI to be critically important to their core business processes. By comparison, among enterprises that spend more than 30% of their R&D on AI, 87% considered AI critical to the business.
Collaboration with universities and public research organisations
Many enterprises in the sample collaborate with universities, public research organisations and other partners to support the use and development of AI. More than half have worked with university faculty, PhD, or postdoctoral students over the past 12 months. Roughly one-third work with undergraduate students.
Perhaps unsurprisingly, enterprises that spend more of their R&D on AI are also more likely to establish collaborations on AI with researchers in public research organisations. Between 60% and 65% of enterprises that spend more than 11% of their R&D on AI 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 encouraging and directing this form of investment. Educational and research institutions also possess a range of tools to facilitate R&D and related collaborations.
Enterprises often partner with universities to gain access to skilled graduates. A significant portion (76%) of enterprises involved in such collaborations had hired AI graduates within the past year. A further indication of the importance of AI skills is the share of enterprises that consider government investment in university education and vocational training related to AI to be “very helpful” or “helpful”. Even among businesses that do not prioritise AI in their core operations, 73% view such public investments as either "very helpful" or "helpful". Collaborations with researchers in public research organisations are particularly widespread among smaller manufacturers (64%).
Obstacles to using and adopting AI
Workforce skills
Identifying and implementing AI applications requires a mix of technical and domain expertise, generally involving employees with MSc or PhD diplomas. In addition, the presence of AI-skilled staff is often a prerequisite for venture capital funds to invest in firms developing AI applications. Access to AI talent can be very problematic, especially for SMEs. Smaller firms compete with large companies for limited AI specialists and data engineers with postgraduate education. Competitors for talent include tech giants such as Amazon, Google and Microsoft, which can offer more attractive salaries and work conditions. SMEs may also have more limited access to on-the-job training opportunities to help staff to build AI skills. In addition, countries often compete for talent at the postgraduate level, for instance, by offering higher PhD salaries.
The 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises reiterates these findings and observations. 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. In a context where AI skills are scarce almost everywhere, enterprises collaborate with universities to secure access to talented graduates. Indeed, a high share (76%) of enterprises collaborating with universities recruited graduates in AI in the previous 12 months.
Many enterprises do not understand which skills they need
An under-examined question is whether firms fully understand their AI skills needs and whether formal academic qualifications provide sufficient information to employers making recruitment decisions. The survey asked if enterprises had experienced difficulties during the preceding 12 months in understanding the skills to look for in potential AI recruits. Almost 19% of respondents acknowledged having this problem. Indeed, 86% of enterprises that place a high value on public support for partnerships with educational and vocational institutions also consider the development of new qualification frameworks to be either “very useful” or “moderately useful”. All told, many enterprises in search of increased AI skills feel they need a better practical understanding of how to identify and use the skills in question. Updated qualification frameworks could help recruiters better assess how different AI skills can be applied in specific corporate contexts.
Other obstacles to using AI
Respondents were asked to indicate which conditions, if any, had limited their enterprise in using AI over the preceding year. The most frequently experienced obstacle was 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 the experience of agencies across the G7 countries charged with accelerating the spread of digital and other technologies in firms (Box 1.4 and Chapter 4).
AI projects involve a degree of experimentation where the ROI is inherently uncertain. This happens even for well-established use cases. Interviews with diffusion institutions indicate that many firms – particularly SMEs – are uncertain about what they can gain financially from implementing AI. They may find it challenging to define and delimit the business case for adoption. Finding reliable estimates of the ROI can be difficult, even when applications are narrowly defined. For example, an AI system might be able to notify users about potential machinery failures, allowing a firm to conduct preventive maintenance on equipment. However, verifying the necessity of this intervention and confirming that it, in fact, prevented a breakdown (and its associated expenses) might not be straightforward. Having a documented history of breakdowns could assist in calculating the ROI for implementing such an AI system, but such data might not be easily accessible. In addition, the process of gathering reliable data incurs expenses that must also be considered in the ROI assessment.
While estimating an AI system's contribution to cost savings and efficiency gains can sometimes be relatively uncomplicated, calculating the ROI for new AI-enabled products, services, or business models can be more challenging. Service providers selling AI solutions also face ROI-related problems, as the right revenue model can be unclear (e.g. subscription, licence or charging per task, as some cloud computing companies do).
Source: OECD/BCG/INSEAD interviews with enterprises.
Over 40% of manufacturing and ICT enterprises reported difficulty finding AI system vendors that provide customised solutions. This issue has prompted some public sector agencies – such as AI Singapore – to adopt a process for recommending vendors with proven track records, with the goal of reducing search costs, particularly for small businesses (see Chapter 4).
Approximately 40% of enterprises lack clarity around the possible 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). About 40% of businesses state that insufficient external funding for investment hindered their use of AI in the previous year. However, as might be expected, 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).
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.
Manufacturers experience almost all obstacles to AI adoption more frequently than enterprises in ICT. This might have several causes. For instance, manufacturing has historically been product rather than data-led and has less of a tradition of working with big data (although differences exist within the manufacturing sector, especially regarding continuous flow manufacturing, for instance, of petrochemicals, and manufacturing of discrete products, such as cars).
Obstacles to using cloud computing
Prior research on AI adoption has revealed a pattern where organisations that are early adopters of websites and computer systems tend to be early adopters of cloud services as well, with AI adoption following suit. Survey participants were asked to specify the obstacles they encountered, if any, in using cloud services. The cost of retooling systems was the most frequently cited obstacle, both in manufacturing (60%) and ICT (56%). Approximately every second enterprise in both sectors had concerns about customisation of applications, corporate compliance or network stability. Roughly one-third reported 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.
Public services to support the adoption of AI
A main feature of the survey is its assessment of the extent to which enterprises use and value public sector services to support the adoption of AI. A key finding is that a significant share of enterprises use such services. The most frequently used services in ICT and manufacturing are those providing access to information or advice (75% in manufacturing, 69% in ICT). Initiatives to develop human capital are also among the most widely used and highly valued. Roughly 58% of enterprises make use of training services in some way supported by the public sector. In addition, 42% use public programmes that promote access to finance, such as tax credits on R&D spending, grants or credit guarantees.
Public sector services are most used by manufacturers with 50‑250 employees. Some 85% of such enterprises use some form of information or advisory service, compared with roughly 68% for other groups of 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 the surveyed enterprises in the United States use services promoting access to finance, compared to 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 types of support to increase staff skills in AI: partnerships with educational and vocational institutions; tax allowances or tax credits for training in AI; and support to develop qualification frameworks for graduates in the field of AI. Regardless of size, most enterprises indicate that one or more of these forms of public support would help 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”. In addition, 67% of enterprises indicate that tax allowances or tax credits for training in AI would be “very useful” or “moderately useful”. As noted earlier in this section, most enterprises state that they would value support to develop qualification frameworks for graduates in the field of AI.
Across the survey sample, just over 50% of enterprises use AI to facilitate training or to provide cognitive support for workers. Applying AI for cognitive support is a relatively advanced use of the technology. Such applications frequently combine AI with other technologies, such as augmented and virtual reality.
Information services provided by the public sector to assist in the adoption of 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 they would be at least “helpful”. Fully 83% of enterprises judged that having more information on current or forthcoming regulations around data or AI or on expected ROIs in AI would be either “helpful” or “very helpful”. It is striking that even though many of the sampled enterprises use AI in quite advanced ways, they still seek additional information on various domains of AI. This suggests that such information may be even more important for firms that do not use AI already. 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.
Other public sector initiatives to support the uptake of AI
Looking to the future, enterprises were surveyed on the possible value of a wider set of public initiatives to foster the use of AI beyond information services, namely:
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.
The responses can be interpreted as enterprises’ wish for more of the above initiatives. Most enterprises in the sample considered all the listed public sector initiatives “helpful” or “very helpful”. Reiterating answers to previous questions, among the most widely and highly valued initiatives were those to develop human capital. Some 86% of enterprises considered 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 considered public investments in university education and vocational training in fields related to AI to be “helpful” or “very helpful”. In addition, but slightly less prevalent, the surveyed enterprises thought enhancing government officials' understanding of AI was important.
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. Public initiatives 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 adopting AI. This finding emphasises the potential benefits of making administrative public datasets (more) accessible to firms.
Enterprises that use more AI applications are more likely to use all three of the following categories of public support: information and advice; training services; and measures that improve access to finance. Enterprises that report more obstacles to using cloud computing and AI are more likely to use public sources of information and advice but not training services or access to finance and subsidies. One possible way to understand these findings is that information scarcity is the primary barrier to surmount when adopting AI applications, whereas assistance for training and financial resources becomes pertinent only after adopting an AI application.
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 to using AI. The generally positive view of possible public sector initiatives varies little in terms of industry and firm size.
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. One main message is that enterprises seek clarity with respect to accountability for the safe use of AI. While the desire for clear accountability is unsurprising, these findings underscore the need for policy makers to examine regulations for possible ambiguities and to assess how best to communicate regulatory information to firms.
Structured interviews and case studies reveal diverse approaches used by institutions supporting AI diffusion in firms
Institutions for technology diffusion are public or quasi-public bodies that facilitate the spread and use of knowledge and methods to assist firms in adopting technologies. Some are well known, such as Germany’s Fraunhofer IAO/IPA, the United Kingdom’s Digital Catapult and the United States’ Manufacturing Extension Partnership programme. Such institutions work to assist the adoption of many, frequently digital, technologies. Chapter 4 examines the types of support provided by diffusion institutions in respect of AI. All the institutions in question have established dedicated services to support the uptake of AI. The literature on AI adoption has barely explored the role of institutions in technology diffusion. Chapter 4 is based on evidence gathered through desk research, structured interviews and written contributions from 19 institutions supporting AI uptake in firms in G7 countries plus Singapore. AI Singapore was invited to participate as this organisation has employed several novel approaches to diffusion, from which valuable lessons can be drawn.
The interviews first aimed to characterise how each diffusion institution supports AI adoption. They then explored each institution’s experiences and understanding of the main barriers to AI adoption in firms. Finally, interviewees were asked to describe the institutions' views on the most effective ways to support AI adoption.
Mechanisms used by diffusion institutions
Chapter 4 identifies seven main mechanisms that diffusion institutions use to assist firms in adopting AI:
1. technology extension services, which can help firms define business problems to be solved and develop proofs-of-concept of how AI can help
2. grants for business R&D and public research, which can help mitigate some of the risks associated with developing AI
3. business advisory services, which give non-technical support to managers to improve their understanding of their firm's AI readiness, opportunities and challenges
4. networking and collaborative platforms, which aid in the development of public and private AI ecosystems, create demonstration effects and facilitate knowledge transfer
5. on-the-job training
6. information services
7. open-source code to help firms increase their AI capabilities.
Many diffusion institutions blend these mechanisms. Table 1.3 matches each of the 19 diffusion institutions in question to the services they provide.
Effective ways to support AI adoption identified by institutions for the diffusion of technology
Institutions that support the diffusion of technology in business usually choose to work with firms certain initial capabilities and where AI is or can be part of the company's core business. To work with potential AI adopters, staff in diffusion institutions usually commence with an evaluation of firms’ digital and AI proficiency. This evaluation may be conducted when assessing eligibility for grants for business R&D, during technical visits, and in workshops offering business advice. AI Singapore uses self-assessment tools to help firms gauge their capabilities and identify the assistance they require. For companies that are not sufficiently digitally mature, many governments have separate policy instruments offering dedicated support for digitalisation.
Some diffusion institutions only select AI projects with a clear path to increases in performance, product or service quality, or cost reduction. Interviewees explained that this increases the likelihood that a proof-of-concept will achieve tangible impacts, which helps to convince firms to scale up investments. However, other institutions consider that to begin, firms should not focus just on the ROI but should also value experimentation that may lead to subsequent breakthroughs.
Diffusion institutions broadly agree that preparing catalogues of applications, use cases, and success stories can help firms understand the possible gains from AI. Such catalogues help establish a record of success. They can document positive and negative experiences that other firms can learn from. Specifically, case studies that quantify the economic impact of investments (such as sales increases and cost reductions) can help to estimate the ROI. Catalogues or directories of this sort can also help managers better understand the opportunities, challenges and constraints posed by AI. Several of the diffusion institutions described in Chapter 4 compile such catalogues, including the Digital Manufacturing and Cybersecurity Institute (MxD), Plattform Lernende Systeme, the Forum IA Québec, Fraunhofer IAO/IPA and the NHS AI Lab.
Each of the following sections considers a specific mechanism that diffusion institutions use, i.e. technology extension services, grants for business R&D, business advisory services, funding for applied research, networking and collaborative platforms, on-the-job training, and data platforms and open-source services. Each section sets out the main observations on policy and institutional practices for the mechanism concerned.
Technology extension services
In implementing technology extension services, interviewees suggested that diffusion institutions should work with firms in a sequence of steps: 1) establishing one or more business cases describing how to apply AI (for instance, clarifying how autonomous forecasting, decision support or decision making would help); 2) scoping possible AI solutions and assessing data maturity (for example, assessing if the business is gathering and processing the correct data); and 3) developing pathways to implementation. Several recommendations are evident for each of these steps:
1. Generic information on use cases can help advance a base understanding of AI in firms. However, to establish the business case for AI adoption, diffusion institutions need to obtain as much operational data from the firm as possible, mapping possible AI applications to firm-specific goals.
2. The staff of diffusion institutions should spend time at the firm to assess its digital maturity and simulate what an AI solution could do. Developing proofs-of-concept should begin by tackling more straightforward problems using readily available data. Staff can also help estimate the ROI for a more extensive AI project and help firms decide whether to invest in it. To this end, diffusion institutions highlight the need to have an economist join data engineers and other technical experts in technology extension projects.
3. An implementation roadmap should describe in detail what deploying a fully integrated AI solution across the organisation would entail. AI solutions can significantly impact various business processes and departments (e.g. accounting, purchasing and production). The roadmap should also describe Top of Form.How to ensure AI models perform well over time. The implementation plan should be co-developed with the firm’s staff from the outset to secure co-operation and draw on employees' collective knowledge.
Technology extension services reportedly work best when firms assign their own staff and contribute in-kind resources. Projects can also involve other actors, such as universities and research institutes. Such collaborations can be particularly valuable in projects involving pre-commercial AI applications.
Business advisory services
According to interviewees, business advisory services can be particularly effective in three main ways. First, they can help firms make initial estimates of the ROI using scenario analysis without necessarily going into the technicalities of AI. For instance, advisors can help managers estimate the downtime of machines or production lines and the financial savings to be made using predictive maintenance. Secondly, diffusion institutions can help raise awareness and understanding of any public support for AI adoption offered at national and international levels (e.g. EU calls). Firms are often unaware of such support, including funding opportunities. Thirdly, diffusion institutions can offer business advisory workshops to raise AI literacy among managers. They can also provide advice on ethics and regulation.
Networking and collaborative platforms
Companies often have similar business problems and ways of using AI to solve them. Seminars and conferences can facilitate valuable exchanges between business executives and help raise understanding of the opportunities that AI presents and the types of transformation firms need to make. Seminars and conferences also facilitate networking between managers, researchers, trade associations, diffusion institutions, AI solution providers and other actors. Such events can help AI reach business sectors where adoption tends to be lower. They can also be used to gather the views of stakeholders in order to inform and shape policies and regulations for AI.
Grants for business R&D and applied public research
Financial support can reduce the risks entailed in developing proofs-of-concept and exploring theoretical applications. As part of their allocation criteria, some grant schemes ask firms to indicate the expected ROI, or the cost reduction, expected of an AI system. Financial support can also help firms build a digital infrastructure for collecting, managing and processing data for AI, e.g. support for deploying IoT technologies. Some business sectors, such as fintech, already use AI intensively. However, when used to help acquire third-party AI applications, grants can encourage firms in other business sectors to work with AI solution providers. According to the interviewed diffusion institutions, grants that deliver the best outcomes require beneficiaries to match public support with their own resources (financial or in-kind). Similarly, publicly funded research projects reportedly produce the best results when companies assign their staff to the research team.
On-the-job training
Training courses are essential for existing employees to gain the technical knowledge required for AI adoption. Tools for self-assessment of digital maturity, like AI Singapore's AI Readiness Index, can also be used to help managers, venture capitalists and solution providers identify use cases and design business models for AI solutions. Managers and technicians can also be trained in information governance, regulations and ethical issues. Such training can help tackle compliance and AI assurance concerns that often stop firms from using their data or prevent them from engaging with AI altogether. While on-the-job training can help firms address the scarcity of workforce skills in AI in the short term, various diffusion institutions consider that countries need to embed AI across tertiary education.
Information services and open-source code
Open-source tools make AI methods and resources accessible to a broad audience beyond AI specialists and computer scientists. It is easier for statisticians, data engineers, physicists and other professionals with varied backgrounds to work with such tools than to develop algorithms from scratch. Diffusion institutions such as AI Singapore and NHS AI Lab use open-source resources together with other mechanisms, such as on-the-job training and technology extension services.
Particularly helpful for SMEs are publicly funded infrastructures that subsidise computing resources (e.g. hardware and cloud computing) and provide real or synthetic training data for free or at low cost. Such resources also need to be combined with other forms of support, such as business advice. For example, Digital Catapult's Machine Intelligence Garage gives SMEs access to computational resources in combination with mentorship and fundraising opportunities. By verifying the parties' identities and ensuring the integrity of data transfer, digital platforms and online marketplaces can also provide a trustworthy channel for secure data transfers. In addition, many firms underestimate the opportunities to establish data partnerships to tackle common problems, especially those involving competitors.
Findings from interviews with enterprises offer new insights on the role of public services for AI adoption in firms
Chapter 5 reports the findings of interviews with senior staff working in firms in the two sectors addressed in the survey. The interviews aimed to elicit qualitative information to better interpret the quantitative data gathered through the survey.
Various of the interviewees came from enterprises that responded to the survey. Others were identified from a pool of over 600 candidate interviewees. The 15 interviewed experts hold positions such as chief information officer, chief technology officer, head of digital business, head of R&D, and chief technology officer, among others. There was an approximately equal representation among G7 countries and between the two surveyed sectors.
Enterprises and data acquisition
The interviews reveal that while many enterprises acquire data from research institutes and the public sector, they rely most on private data sources. The most common type of data acquired from public sources is generic data, such as demographic information, public company records, labour statistics and weather data. More specific and commercially valuable data sets from public administrations are rare. Private data sources are the preferred choice for most firms, as they offer more specialised and proprietary data that can provide a competitive advantage. In addition, more opportunities exist for providing feedback on data quality to private data sources than to public sources.
Procedural complexities in acquiring public data can impede data-driven decision making. These complexities often exist for legitimate reasons, such as maintaining data integrity and security, but can create inefficiency. Multiple layers of approvals, reviews and checks can lead to prolonged waiting periods, which can also render data obsolete when accessed.
In addition, data in public repositories are often too old for real-time applications. Many businesses must invest time and effort to validate the currency of public data. Policy makers need to ensure that data remain relevant and actionable. Using publicly sourced datasets can also be problematic due to vague terminologies and other shortcomings. The absence of comprehensive documentation can leave users to grapple with the data's true meaning and context. For example, a business might come across a large CSV or Excel file from a public source and encounter columns filled with terminology that is not easily understandable while lacking accompanying documentation for clarification. A common problem the interviewees reported is the data quality itself. For example, it is not uncommon to encounter discrepancies, conflicting information and missing data. Overall, policy makers need to ensure that shortcomings of the above sorts are addressed.
Interviewees generally considered that a centralised platform for public sector data access could streamline the search and retrieval process. A centralised hub could facilitate seamless transitions between databases, enhancing users' ability to access and link to specific studies or datasets.
Interviewees emphasised that the legal frameworks governing cross-border data flows could be made more compatible. International data sharing can be an intricate process, particularly if navigating diverse data-sharing laws. Different countries have distinct data protection and privacy legislation. This can pose challenges for companies that operate in multiple jurisdictions and need to comply with each region's specific and sometimes complex regulations.
Vendor certification is common across industries and could be adapted for data vendors. One interviewee noted that such certification would help to provide assurance and confidence in the data's authenticity and reliability. Especially for SMEs, checklists of the most important criteria to consider in vendor search and selection would be helpful.
Public services that support the adoption of AI
Access to information or advice concerning the adoption of AI
Most interviewed experts affirm that the insights derived from public sector sources help to make informed decisions and shape business strategies. Access to information such as economic data, regulatory updates and compliance guidance is considered valuable. Such information is crucial in planning, analytics, market sizing, go-to-market strategies and understanding market dynamics.
Interviewees drew attention to a lack of consolidated information on private-sector AI software or services. Companies frequently receive use case solicitations from vendors, presented in marketing language. Governments might help by providing such information in more neutral ways. Several interviewees suggested that governments could provide guidelines or a framework to aid SMEs in navigating the vendor selection process, advising them on, for instance, the top ten considerations to consider when choosing an AI vendor.
Interviewees also pointed to challenges in accessing public sector information to facilitate AI development. They highlighted the frequent lack of clear pathways to specific public agencies. The absence of a one-stop interface and streamlined processes and the occasional fragmentation of channels to public services create challenges in identifying the right agency or programme to consult. Policy makers could help by establishing a consolidated platform or resource hub that streamlines access to AI-related public information, guidance and advice. Especially for SMEs, guidelines outlining agency roles and expertise and mechanisms for companies to communicate their needs would also enable more targeted and efficient exchanges.
Publicly provided or supported training services
Regardless of sector, all interviewees reported challenges in finding specialised AI talent. In the OECD/BCG/INSEAD survey, 58% of enterprises use public sector training services to help adopt AI. Interviewees who expressed a reluctance to use public sector training programmes emphasised the need for more specificity in the training offered. For instance, instead of generic AI training, they found greater value in programmes tailored to industry or business-specific needs. For example, manufacturers may prefer training in AI that focuses on optimising supply chain management.
Additionally, the experts highlighted the value of hands-on training oriented towards real-world projects. Workshops in which participants use AI tools and datasets in practical exercises related to their industry can significantly improve AI readiness.
Various interviewees asserted that public sector providers should collaborate with industry to design targeted training. Inviting industry professionals to share their experiences and insights can help develop practical training materials that resonate with private companies.
Reiterating the survey findings (Chapter 3), the interviewees observed that AI is often perceived as a broad, all-encompassing term, overlooking the existence of distinct subfields within it. Academic certifications may not provide the comprehensive information that employers seek. In this rapidly evolving field, there is a growing need for new qualification frameworks that effectively communicate precise and relevant information regarding candidates' capabilities and competencies to employers.
The interviewed experts agreed on the need for new AI curricula to meet the growing demand for skilled AI professionals. AI degree programmes often lack sufficient focus on industry-specific applications and practical skills. For example, a healthcare organisation may seek AI graduates who are well-versed in medical image analysis and diagnosis. Companies often seek AI professionals who can quickly apply their knowledge in the workplace. Consequently, curricula that incorporate practical components, such as internships or industry placements, are highly valued by employers.
External collaboration to develop AI
Engaging with universities and public research institutions to develop AI
The 2023 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises showed widespread collaboration with universities and public research institutions. More than half of the responding enterprises collaborate with university faculty members, PhD candidates, or postdoctoral students to advance AI development.
The interviewees reiterated the value of such collaborative partnerships. Firms can share their industry insights, practical experience and real-world datasets, enriching academic research. In turn, academic institutions can share their latest research, methodologies and theoretical advances, helping firms utilise cutting-edge research. Such partnerships can also provide access to advanced computing infrastructure and dedicated R&D teams, enabling firms to undertake more ambitious and resource-intensive AI projects. Firms can also enjoy opportunities for talent acquisition and development.
Collaborations between research institutions and firms, especially those in the ICT sector, frequently yield intellectual property (IP) and the associated IP rights. The interviewees noted that striking a balance between the interests of both parties regarding the ownership, usage and commercialisation of IP can be complex and may give rise to disagreements. Indeed, naturally, firms often focus on commercialisation and ROI, while academic institutions prioritise scientific discovery, publication and academic recognition. These differing goals and incentives can lead to conflicts regarding confidentiality and data sharing. One interviewed expert highlighted the potential benefits of developing a framework or model non-disclosure agreements to facilitate collaboration between firms and universities.
Most interviewees drew attention to the complexity of managing the distinct cultures, priorities and operational structures characteristic of corporate and academic environments. These diverse institutions typically have different approaches to decision making and timelines. Academic institutions often operate on longer-term research cycles, while firms operate in faster-paced, market-driven environments.
An obstacle mentioned in some interviews was the lack of transparency in how universities use the funding that firms provide, how other developments within universities might affect a project (such as a turnover in postdocs), and overall project governance. Delays, misunderstandings, and even conflict can arise without clear guidelines, transparent processes, and well-defined project governance structures.
One interviewee highlighted that centres of AI research predominantly focus on collaborations with medium and large-size enterprises. Dedicated programmes could help address specific challenges faced by SMEs, such as overall resource constraints and more limited access to AI talent.
Interviewees held that some public financial support for collaborations could help mitigate risks for firms. Public financial support might be limited to enterprises’ first collaborative experience. Companies would also like less complex processes when applying for public funds that support AI research in collaboration with universities. Interviewees stressed the importance of enhancing transparency throughout the process. Clear guidelines, well-defined evaluation criteria, practical examples of successful applications and accessible information about funding opportunities would all help. Additionally, interviewees advocated for feedback loops to facilitate communication between funding agencies and applicants.
Chapter 6: Implementing the OECD/BCG/INSEAD survey in Brazil: key findings
Chapter 6 reports the results of a survey on the use of AI in enterprises in the State of São Paulo, Brazil. São Paulo is the most populous State in Brazil and the largest economically. It also hosts an innovation ecosystem that includes Brazil’s main universities and research centres, as well as many businesses in high-tech sectors.
The survey was conducted by SEADE, the official statistics and data production organisation of the State of São Paulo, in partnership with the Regional Center for Studies on the Development of the Information Society (Cetic.br), from the Brazilian Network Information Center (NIC.br).
The survey instrument was adapted from the OECD/BCG/INSEAD questionnaire, permitting comparison with the survey in G7 countries. The São Paulo survey had the same target populations of medium- and large-sized manufacturing and ICT enterprises. The survey adopted a probabilistic approach, meaning that it aimed to obtain results statistically representative of the entire population of enterprises in the state.
How enterprises in the State of São Paolo use AI
The survey findings indicate that the use of AI among large and medium-sized enterprises in the manufacturing and ICT sectors in the State of São Paulo is relatively incipient. From a sample of 2 561 enterprises, only 167 (6.5%) were found to use AI actively. This corroborates previous research in Brazil, such as (Brazilian Network Information Center, Brazilian Internet Steering Committee, 2022[19]), which highlighted low rates of use of AI across enterprises of all sizes in all sectors. There is considerable room for expanding the use of AI, transitioning from point solutions to more integrated adoption, such as incorporating customer relationship management systems.
Among enterprises that actively use AI, 49% use it in customer-oriented services. The second most frequent application is in process control, automation and optimisation of production (44%), including such uses as predictive maintenance and automated support for programmers. These results broadly align with the findings from G7 countries. However, most enterprises surveyed in the State of São Paulo use only a few AI applications (58% of enterprises with just one or two AI applications compared to 4% in G7 countries).
Most enterprises in São Paulo procure solutions externally and exhibit a relatively low level of internal development. Some 28% of the surveyed enterprises use AI for R&D, considerably lower than in most G7 countries. A higher share of enterprises in São Paulo also considers AI of minor importance to main business processes (20%) than in G7 countries (8%). Moreover, managerial positions related to AI are still rare, even in enterprises that use AI.
Practices and partnerships to adopt and develop AI
Particularly salient is the limited extent of partnerships with researchers. Only 6% of enterprises collaborate with undergraduate students, faculty, doctoral students or postdoctoral researchers, while partnerships with researchers outside of universities occur in only 5% of AI-using enterprises. By contrast, among G7 countries, more than 50% of enterprises have collaborated with university faculty, PhD, or postdoctoral students.
Expanding the uptake of AI and the role of the public sector
Public authorities in Brazil have created many initiatives to support business innovation. However, none to date specifically target AI. As in G7 countries, most enterprises in the State of São Paulo would welcome one or another form of public support to help strengthen staff skills in AI. For instance, 64% of respondents assert that help to establish partnerships with educational and professional training institutions would be “very useful”. Regarding broader public sector initiatives to support the adoption of AI, investment in university education and professional training in AI is considered particularly important. Fully 75% of the enterprises declare that such policies would be “very useful”.
As in G7 countries, most enterprises judge that public information services could be “helpful” or “very helpful”. Some 62% of respondents consider that information on current or forthcoming regulations about data or AI would be “very useful”.
It is widely understood that IT infrastructure and connectivity problems in some regions of Brazil require public sector initiatives to be fully resolved. It is perhaps unsurprising that 73% of enterprises cite upgrading IT infrastructure, such as high-speed broadband, as “very useful” for adopting AI.
The survey findings suggest that benefits could come from creating support instruments that encourage partnerships around AI. Benefits could also be had from examining the suitability and current designs of innovation support instruments, with a view to identifying opportunities where adjustments might facilitate enterprises’ efforts to adopt and innovate with AI. Developing and/or strengthening a variety of possible information services might be a low-cost but relatively high-impact first step.
This section summarises the study’s main policy-relevant take-aways. A key methodological caveat is that this report comprises a mix of cross-sectional survey data and information gleaned from interviews. The data analyses are correlational and cannot offer evidence on which sorts of policies are likely to be most cost-effective. Several results merit further examination using other methods. For instance, it would be helpful to better understand causal relationships associated with public sector support to AI diffusion in business. 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 leadership in an enterprise will both adopt AI more actively and seek external assistance more actively?
An overarching observation, evident from the studies reported in Chapter 2, is that productivity benefits could come from accelerating rates of adoption of AI. In all OECD countries (plus Brazil) rates of use are low in core business processes, particularly in smaller enterprises. Increasing uptake of AI matters because of the possible positive impacts on labour productivity. However, other benefits could also accrue, for instance with respect to environmental outcomes. AI can, for example, lower defect rates in production, reducing the need for material inputs. AI can also make many processes more energy efficient, for instance, by optimising logistics.
Policy insights from this and comparable studies will likely become more important as AI adoption rates increase. This is because the focus of this work has been on medium-sized and large enterprises as well as active users of AI. This focus provides insights that will be relevant to the numerically larger group of small firms that will aim to adopt AI. Indeed, the survey shows that enterprises that use more applications of AI are more likely to use a range of public support services. As the number of enterprises seeking to apply AI more widely grows, how public services and policies respond will become more consequential.
A main and somewhat unexpected finding is that a significant share of enterprises have, at some point in time, used and valued various public services that can aid the adoption of AI positively. The most frequently used services are those providing access to information or advice.
Addressing the need for skills
The 2022-23 OECD/BCG/INSEAD Survey of AI-Adopting Enterprises reiterates the findings of many previous surveys that a scarcity of skills – particularly specialised talent – hinders uptake. Even many large enterprises experience the same problem. Initiatives to develop human capital are also among the most widely used and highly valued.
Several results were obtained from the survey and interviews relevant to the content of curricula in academic institutions as well as the content of training services. All told, many enterprises in search of increased AI skills feel they need a better practical understanding of how to identify and use the right skills. Academic certifications may not provide the comprehensive information employers seek in this rapidly evolving field. Updated qualification frameworks could help recruiters better assess how the AI skills possessed by holders of different qualifications can be applied in their businesses.
Training programmes should be well tailored to industry or business-specific needs (such as using AI to optimise supply chain management). Particularly valuable is training on real-world projects using AI systems and datasets common in specific areas of business. Public sector providers should collaborate with industry to design training materials.
Curricula in some AI degree programmes might also incorporate a greater focus on industry-specific applications.
Policy makers should likewise examine where enhancing government officials' understanding of AI could have the greatest effect. The survey indicated that some enterprises consider that strengthening these skills would be beneficial, but this study was unable to explore the topic in more detail.
Public data and data policies
Policy makers need to be alert to the quality of data in public data repositories, ensuring, for example, that there is no conflicting information. Procedures facing enterprises seeking to acquire public data should be reviewed and, where possible, simplified. Irrespective of procedural complexity, data from public repositories is often too old for real-time use. Policy makers should ensure that data remains current and actionable.
Data in public datasets can also be problematic due to vague terminologies and other shortcomings. Documentation should be available such that users of public data can easily comprehend its meaning and context.
Centralised platforms for public sector data access could streamline the search and retrieval process.
Policy makers should seek to enhance the compatibility of legal frameworks governing cross-border data flows. International data sharing can be an intricate process and poses particularly acute challenges for companies that operate in multiple jurisdictions.1
Consideration could be given to developing checklists for SMEs to help them select the most suitable private data vendors. The aim would not be for public authorities to identify preferred vendors but to help SMEs evaluate which vendors best suit their needs.
Collaboration with universities and public research organisations
To aid the use and development of AI, collaborations with universities and public research institutions are widespread, valued and have several purposes. Some public financial support for collaborations could help mitigate risks for firms. However, public financial support might be limited to enterprises’ first collaborative experience. R&D tax credits might be used to encourage industry-university collaboration and research commercialisation. In several countries, R&D tax credits are designed to provide significant additional incentives for collaborative research, above and beyond the tax credit received when R&D is undertaken in-house. Collaborative R&D has been shown to positively affect the technological capacity of firms even after controlling for the possibility that more dynamic firms might also collaborate more often (Barajas, Huergo and Moreno, 2011[24]).
Processes for applying for public funds that support AI research in collaboration with universities should be simplified. Information could be made widely available to enterprises, clearly describing funding opportunities, evaluation criteria and examples of successful applications. Feedback loops to funding agencies could also help.
As seen in many studies, including those unrelated to AI, firms and research bodies typically have different goals, working practices, implementation schedules and interests. This perennial issue invites policy attention to the possible benefits of developing model framework agreements for the benefit of both parties to facilitate collaboration between firms and universities.
Universities and public research organisations themselves could seek to ensure transparency in key operational practices, including how the funding that firms provide is used, how other developments within universities might affect a project (such as a turnover in postdocs), and overall project governance.
Information and advisory services provided by the public sector to assist in the adoption of AI
A large majority of enterprises judge that information and advice provided by specialised public bodies could help their use of AI. Even though many of the sampled enterprises use AI in quite advanced ways, they still seek additional information on various domains of AI. This suggests that policy makers should look for cost-effective ways of delivering easily findable, accessible, current and specific information and advice, for instance, on regulatory updates, compliance guidance and evolving business use cases for AI.
Governments could provide guidelines or a framework to aid SMEs in navigating the vendor selection process, advising them on, for instance, the principal considerations to be aware of when choosing an AI vendor.
Policy makers could help by establishing a consolidated platform or resource hub that streamlines access to AI-related information, guidance and advice from public agencies. Especially for SMEs, guidelines outlining agency roles and expertise, along with mechanisms for companies to communicate their needs, would also help.
One main message is that enterprises seek clarity with respect to accountability for the safe use of AI. While the desire for clear accountability is unsurprising, these findings underscore the need for policy makers to examine regulations for possible ambiguities and to assess how best to communicate regulatory information to firms.
Other obstacles to using AI
The survey data suggest that IT infrastructure deficits, such as a lack of high-speed broadband, require examination and possibly a policy response.
Institutions supporting AI diffusion in firms
Institutions dedicated to facilitating the uptake of digital technologies, including AI, in firms are present in most, if not all, OECD countries. This study offers insights on good practice in the programmes they implement and may be of particular interest to policy makers looking to review, create or expand such organisations. The insights on operational practices are described in greater detail earlier in this chapter, and encompass:
Technology extension services, for instance, as concerns a sequence of steps that might be followed in implementing extension services, raising AI literacy among managers, helping estimate ROIs using scenario analysis, and providing advice on ethics and regulation.
Networking and collaborative platforms, for instance, in facilitating exchanges between business executives, helping increase understanding of the types of transformation that firms need to make, and gathering feedback for policy makers.
Grants for business R&D and applied public research, for instance, helping firms to estimate the costs and benefits of an AI system, and the role of resource sharing on the part of beneficiaries.
On-the-job training, for example, providing tools for self-assessment of digital maturity, training in information governance, regulations and ethical issues.
Information services and open-source code, for instance, in helping SMEs gain access to computing resources (e.g. hardware and cloud computing), as well as real or synthetic training data, for free or at low cost.
Improving the evidence base for policy
Policy makers should examine the size of the impact of diffusion institutions relative to the goal of accelerating the uptake of productivity-enhancing technology across the economy. Most institutions considered in this study do not work with large numbers of client firms. In one or two cases, they only engage with tens of enterprises a year. Most OECD countries now have national AI strategies. These often assert that the widespread adoption of AI is a strategic economic aim. However, the flagship institutions working to diffuse AI in business are small relative to the challenge. Policy making could benefit from more systematic economic evidence on the direct and indirect effects of these institutions. For instance, if they only work with a tiny percentage of the enterprise population, does their work nevertheless create wider demonstration or other secondary effects? If so, are those secondary effects large or small?
A policy-relevant contribution of this study is the development and cognitive testing of novel survey questions addressing such topics as enterprises’ use and assessment of services and policies relevant to AI uptake. As noted earlier, several NSOs helped shape the survey questionnaire, and various of the new questions it contains might be considered for inclusion in future surveys.
As described in Chapter 2, prior studies highlight the need to better understand the international comparability of surveys of AI in firms. Differences in methodology may have created measurement discrepancies. Greater assurance of comparability will help to better inform policy.
Regarding the OECD/BCG/INSEAD questionnaire, a future development could be to expand the scope of research to enterprises that do not currently use AI but intend to do so or are in the initial steps of implementing AI. This would help to better understand the difficulties experienced in using AI and how these difficulties manifest in the different phases of implementation, such as in decision making around investments, the organisation and management of data, equipment acquisition and staff hiring. Such a shift to broader themes on AI uptake would be especially important in contexts where the overall use of AI in the corporate sector is low.
Regarding data collection, in a future iteration of the survey, it could be helpful to identify in advance specifically qualified persons in the responding enterprises to answer the questionnaire. This is because the survey encompasses varied and specific topics, from implementation obstacles to insights into the most helpful support services for the enterprise. An alternative would be to consider having more than one respondent, as the topics addressed may be the responsibility of more than one team within the enterprise.
The enterprise survey has an exploratory character. Budgets permitting, it could eventually be implemented, with possible revisions, across a wider set of countries, sectors, and number of enterprises, and using a sampling frame and probabilistic method allowing generalisation to national populations of enterprises. Doing so would strengthen cross-country and cross-firm statistical analyses.
References
[9] Airbus (2016), “Pioneering bionic 3D printing”, Pioneering bionic 3D printin.
[4] 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.
[12] Atkinson, R. and S. Ezell (2019), The Manufacturing Evolution: How AI Will Transform Manufacturing and the Workforce of the Future, Information Technology and Innovation Foundation, Washington, DC, https://itif.org/publications/2019/08/06/manufacturing-evolution-how-ai-will-transform-manufacturing-and-workforce.
[24] Barajas, A., E. Huergo and L. Moreno (2011), “Measuring the economic impact of research joint ventures supported by the EU Framework Programme”, The Journal of Technology Transfer, Vol. 37/6, pp. 917-942, https://doi.org/10.1007/s10961-011-9222-y.
[16] Bergeret, B. (2020), “How private applied AI R&D labs can make AI work for mid-size businesses”, The AI Wonk, https://oecd.ai/wonk/private-applied-ai-rd-labs-for-mid-size-businesses.
[19] Brazilian Network Information Center, Brazilian Internet Steering Committee (2022), Survey on the Use of Information and Communication Technologies in Brazilian Enterprises, https://cetic.br/media/docs/publicacoes/2/20221121122540/tic_empresas_2021_livro_eletronico.pdf.
[6] Cazzaniga, M. et al. (2024), “Gen-AI: Artificial intelligence and the future of work”, International Monetary Fund, Washington, DC IMF, SDN2024/001, Staff Discussion Note, http://www.imf.org/en/Publications/Staff-Discussion-Notes/Issues/2024/01/14/Gen-AI-Artificial-Intelligence-and-the-Future-of-Work-542379.
[14] Destatis (2023), IKT-Indikatoren für Unternehmen: Deutschland, Jahre, Wirtschaftszweige (ICT indicators for companies: Germany, years, economic sectors), http://www-genesis.destatis.de/genesis//online?operation=table&code=52911-0002 (accessed on 3 June 2024).
[18] Eurostat (2023), “Artificial intelligence by size class of enterprise”, Eurostat Database, https://doi.org/10.2908/ISOC_EB_AI (accessed on 24 May 2024).
[20] Evans, A. and A. Heimann (2022), AI Activity in UK Businesses Report, Capital Economics and DCMS, January 2022, http://www.gov.uk/government/publications/ai-activity-in-uk-businesses.
[3] 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.
[23] Kazakova, S. et al. (2020), European Enterprise Survey on the Use of Technologies Based on Artificial Intelligence, Publications Office of the European Union, Luxembourg, https://op.europa.eu/en/publication-detail/-/publication/f089bbae-f0b0-11ea-991b-01aa75ed71a1.
[5] 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.
[1] LinkedIn Economic Graph (2019), Talent in the European Labour Market, https://economicgraph.linkedin.com/content/dam/me/economicgraph/enus/referencecards/research/2019/AI-Talent-in-the-European-Labour-Market.pdf.
[21] MIC (2021), 2020 Communication Usage Trend Survey, Ministry of Internal Affairs and Communications, http://www.soumu.go.jp/johotsusintokei/statistics/pdf/HR202000_002.pdf.
[22] Montagnier P. and Ek I. (2021), “AI measurement in ICT usage surveys: A review”, OECD Digital Economy Papers, No. 308, OECD Publishing, Paris, https://doi.org/10.1787/72cce754-en.
[15] OECD (2024), “OECD AI Policy Observatory”, https://oecd.ai/.
[25] OECD (2021), OECD Recommendation on Enhancing Access to and Sharing of Data, https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0463 (accessed on 6 March 2023).
[11] OECD (2020), Dealing with digital security risk during the Coronavirus (COVID-19) crisis, http://www.oecd.org/coronavirus/policy-responses/dealing-with-digital-security-risk-during-the-coronavirus-covid-19-crisis-c9d3fe8e/ (accessed on 18 June 2021).
[7] OECD (2019), Recommendation of the Council on Artificial Intelligence, https://legalinstruments.oecd.org/en/instruments/oecd-legal-0449 (accessed on 6 March 2023).
[13] Rammer, C., G. Fernández and D. Czarnitzki (2022), “Artificial intelligence and industrial innovation: Evidence from German firm-level data”, Research Policy, Vol. 51/7, p. 104555, https://doi.org/10.1016/j.respol.2022.104555.
[10] Ransbotham, S. et al. (2017), “Reshaping business with artificial intelligence: Closing the gap between ambition and action”, MIT Sloan Management Review, https://sloanreview.mit.edu/projects/reshaping-business-with-artificial-intelligence/.
[2] Tricot, R. (2021), “Venture capital investments in artificial intelligence: Analysing trends in VC in AI companies from 2012 through 2020”, OECD Digital Economy Papers, No. 319, OECD Publishing, Paris, https://doi.org/10.1787/f97beae7-en.
[8] WIRED Chen, S. (2017), , http://www.wired.com/story/the-ai-company-that-helps-boeing-cook-new-metals-for-jets.
[17] Zolas, N. et al. (2020), National Bureau of Economic Research Working Paper.
← 1. The OECD is undertaking a broad set of analytic and policy initiatives on cross-border data flows. A landmark achievement is the Recommendation of the Council on Enhancing Access to and Sharing of Data (OECD, 2021). Adopted on 6th October 2021, the Recommendation provides the first internationally agreed upon set of principles and policy guidance on how governments can maximise the cross-sectoral benefits of all types of data while effectively protecting stakeholders' rights (OECD, 2021[25]).