This chapter describes the economic characteristics of artificial intelligence (AI) as an emerging general-purpose technology with the potential to lower the cost of prediction and enable better decisions. Through less expensive and more accurate predictions, recommendations or decisions, AI promises to generate productivity gains, improve well-being and help address complex challenges. Speed of adoption varies across companies and industries, since leveraging AI requires complementary investments in data, skills, digitalisation of workflows and the capacity to adapt organisational processes. Additionally, AI has been a growing target area for investment and business development. Private equity investment in AI start-ups has accelerated since 2016, doubling from 2016 to 2017 to reach USD 16 billion. AI start-ups attracted 12% of worldwide private equity investments in the first half of 2018, a significant increase from just 3% in 2011. Investment in AI technologies is expected to continue its upward trend as these technologies mature.
Artificial Intelligence in Society
2. The economic landscape
Abstract
The statistical data for Israel are supplied by and under the responsibility of the relevant Israeli authorities. The use of such data by the OECD is without prejudice to the status of the Golan Heights, East Jerusalem and Israeli settlements in the West Bank under the terms of international law.
Economic characteristics of artificial intelligence
Artificial intelligence enables more readily available prediction
From an economic point of view, recent advances in artificial intelligence (AI) either decrease the cost of prediction or improve the quality of predictions available at the same cost. Many aspects of decision making are separate from prediction. However, improved, inexpensive and widely accessible AI prediction could be transformative because prediction is an input into much of human activity.
As the cost of AI prediction has decreased, more opportunities to use prediction have emerged, as with computers in the past. The first AI applications were long-recognised as prediction problems. For example, machine learning (ML) predicts loan defaults and insurance risk. As their cost decreases, some human activities are being reframed as prediction issues. In medical diagnosis, for example, a doctor uses data about a patient’s symptoms and fills in missing information about the cause of those symptoms. The process of using data to complete missing information is a prediction. Object classification is also a prediction issue: a human’s eyes take in data in the form of light signals and the brain fills in the missing information of a label.
AI, through less expensive prediction, has a large number of applications because prediction is a key input into decision making. In other words, prediction helps make decisions, and decision making is everywhere. Managers make important decisions around hiring, investments and strategy, and less important decisions around which meetings to attend and what to say during these meetings. Judges make important decisions about guilt or innocence, procedures and sentencing, and smaller decisions about a specific paragraph or motion. Similarly, individuals make decisions constantly – from whether to marry to what to eat or what song to play. A key challenge in decision making is dealing with uncertainty. Because prediction reduces uncertainty, it is an input into all these decisions and can lead to new opportunities.
Machine prediction is a substitute for human prediction
Another relevant economic concept is substitution. When the price of a commodity (such as coffee) falls, not only do people buy more of it, they also buy fewer substitute products (such as tea). Thus, as machine prediction becomes less expensive, machines will substitute for humans in prediction tasks. This means that reduction of labour related to prediction will be a key impact of AI on human work.
Just as computers meant that few people now perform arithmetic as part of their work, AI will mean that fewer people will have prediction tasks. For example, transcription – conversion of spoken words into text – is prediction in that it fills in missing information on the set of symbols that match the spoken words. AI is already quicker and more accurate than many humans whose work involves transcription.
Data, action and judgment complement machine prediction
As the price of a commodity (e.g. coffee) falls, people buy more of its complements (e.g. cream and sugar). Identifying the complements to prediction, then, is a key challenge with respect to recent advances in AI. While prediction is a key input into decision making, a prediction is not a decision in itself. The other aspects of a decision are complements to AI: data, action and judgment.
Data is the information that goes into a prediction. Many recent developments in AI depend on large quantities of digital data for AI systems to predict based on past examples. In general, the more past examples, the more accurate the predictions. Thus, access to large quantities of data is a more valuable asset to organisations because of AI. The strategic value of data is subtle since it depends on whether the data are useful to predict something important to an organisation. Value also depends on whether the data are only available historically or whether an organisation can collect continued feedback over time. The ability to continue to learn through new data can generate sustained competitive advantage (Agrawal, Gans and Goldfarb, 2018[1]).
More new tasks come from the other elements of a decision: action and judgment. Some actions are inherently more valuable when performed by a human rather than a machine (e.g. actions by professional athletes, child carers or salespeople). Perhaps most important is the concept of judgment: the process of determining the reward to a particular action in a particular environment. When AI is used for predictions, a human must decide what to predict and what to do with the predictions.
Implementing AI in organisations requires complementary investments and process changes
Like computing, electrification and the steam engine, AI can be seen as a general-purpose technology (Bresnahan and Trajtenberg, 1992[2]; Brynjolfsson, Rock and Syverson, 2017[3]). This means it has potential to substantially increase productivity in a wider variety of sectors. At the same time, the effect of AI requires investment in a number of complementary inputs. It may lead an organisation to change its overall strategy.
In the AI context, organisations need to make a number of complementary investments before AI has a significant impact on productivity. These investments involve infrastructure for the continued collection of data, specialised workers that know how to use data, and changes in processes that take advantage of new opportunities arising from reduced uncertainty.
Many processes in every organisation exist to make the best of a situation in the face of uncertainty rather than to serve customers in the best way possible. Airport lounges, for example, make customers comfortable while they wait for their plane. If passengers had accurate predictions of how long it would take to get to the airport and through security, lounges might not be needed.
The scope of opportunities offered by better predictions is expected to vary across companies and industries. Google, Baidu and other large digital platform companies are well-positioned to benefit from major investments in AI. On the supply side, they already have systems in place to collect data. On the demand side, having enough customers to justify the high fixed costs of investment in the technology is in its early stages. Many other businesses have not fully digitised their workflows, and cannot yet apply AI tools directly into existing processes. As costs fall over time, however, these businesses will recognise the opportunities that are possible by reducing uncertainty. Driven by their needs, they will follow industry leaders and invest in AI.
Private equity investments in AI start-ups
AI investment as a whole is growing fast and AI already has significant business impact. MGI (2017[4]) estimated that USD 26 billion to USD 39 billion had been invested in AI worldwide in 2016. Of this amount, internal corporate investment represented about 70%, investment in AI start-ups some 20% and AI acquisitions represented some 10% (Dilda, 2017[5]). Large technology companies made three-quarters of these investments. Outside the technology sector, AI adoption is at an early stage; few firms have deployed AI solutions at scale. Large companies in other digitally mature sectors with data to leverage, notably in the financial and automotive sectors, are also adopting AI.
Large technology companies are acquiring AI start-ups at a rapid pace. According to CBI (2018[6]), the companies that have acquired the most AI start-ups since 2010 include Google, Apple, Baidu, Facebook, Amazon, Intel, Microsoft, Twitter and Salesforce. Several AI cybersecurity start-ups were acquired in 2017 and early 2018. For example, Amazon and Oracle purchased Sqrrl and Zenedge, respectively.
AI start-ups are also acquisition targets for companies in more traditional industries. These include, notably, automotive companies; healthcare companies such as Roche Holding or Athena Health; and insurance and retail companies.
After five years of steady increases, private equity investment in AI start-ups has accelerated since 2016. The amount of private equity invested doubled between 2016 and 2017 (Figure 2.1). It is estimated that more than USD 50 billion was invested in AI start-ups between 2011 and mid-2018 (Box 2.1).
Figure 2.1. Total estimated investments in AI start-ups, 2011-17 and first semester 2018
Note: Estimates for 2018 may be conservative, as they do not account for a likely lag in reporting (see Box 2.1. Methodological note).
Source: OECD estimation, based on Crunchbase (July 2018), www.crunchbase.com.
Box 2.1. Methodological note
This section estimates private equity investments in AI start-ups based on Crunchbase (July 2018 version). Crunchbase is a commercial database on innovative companies created in 2007 that contains information on more than 500 000 entities in 199 countries. Breschi, Lassébie and Menon (2018[7]) benchmark Crunchbase with other aggregate data sources. They find consistent patterns for a broad range of countries, including most OECD countries (with the exception of Japan and Korea). Consistent patterns were also found for Brazil, the Russian Federation, India, the People’s Republic of China (hereafter “China”) and South Africa. Companies in Crunchbase are classified into one or several technological areas, taken from a list of 45 groups.
Caveats to using Crunchbase include the broadly defined scope of the database, the reliability of self-reported information and sample selection issues. In particular, input of new deals into the database likely takes time and the delay may vary across countries. Start-ups may also increasingly self-categorise as AI start-ups because of investors’ growing interest in AI.
In this report, AI start-ups correspond to those companies founded after 2000 and categorised in the “artificial intelligence” technological area of Crunchbase (2 436 companies). They also include companies that used AI keywords in their short description of their activities (an additional 689 companies). Three types of keywords are considered to be AI-related. The first type is generic AI keywords, notably “artificial intelligence”, “AI”, “machine learning” and “machine intelligence”. The second type of keywords pertains to AI techniques, notably “neural network”, “deep learning” and “reinforcement learning”. The third type refers to fields of AI applications, notably “computer vision”, “predictive analytics”, “natural language processing”, “autonomous vehicles”, “intelligent systems” and “virtual assistant”.
More than one-quarter (26%) of investment deals in AI start-ups included in the database do not report investments by venture capitalists. This analysis estimates the amounts of these deals by using the average amount invested in smaller deals (considering only deals of less than USD 10 million) for the same period and the same country. The rationale for excluding larger deals is that their amounts are more likely to be public information. The estimated value of non-disclosed deals represents about 6% of the total value from 2011 to mid-2018, which may be conservative. The numbers for the first half of 2018 are likely to be conservative because reporting is often not immediate.
AI now represents over 12% of private equity investments in start-ups
AI start-ups attracted around 12% of all worldwide private equity investments in the first half of 2018, a steep increase from just 3% in 2011 (Figure 2.2). All countries analysed increased their share of investments in start-ups focused on AI. About 13% of investments in start-ups in the United States and China were in AI start-ups in the first half of 2018. Most dramatically, Israel has seen the share of investments in AI start-ups jump from 5% to 25% between 2011 and the first half of 2018; autonomous vehicles (AVs) represented 50% of the investments in 2017.
Figure 2.2. AI as a share of private equity investments in start-ups, 2011 to 2017 and first semester 2018
Note: The percentage for 2018 only covers the first half of the year (see Box 2.1. Methodological note).
Source: OECD estimation, based on Crunchbase (July 2018), www.crunchbase.com.
The United States and China account for most AI start-up investments
Start-ups operating in the United States account for most AI start-up equity investments worldwide. This is true for the number of investment transactions (“deals”) and amounts invested, which represents two-thirds of the total value invested since 2011 (Figure 2.1). These facts are unsurprising, considering that the United States accounts for 70-80% of global venture capital investments across all technologies (Breschi, Lassébie and Menon, 2018[7]).
China has seen a dramatic upsurge in AI start-up investment since 2016. It now appears to be the second player globally in terms of the value of AI equity investments received. From just 3% in 2015, Chinese companies attracted 36% of global AI private equity investment in 2017. They maintained an average of 21% from 2011 through to mid-2018.
The European Union accounted for 8% of global AI equity investment in 2017. This represents an important increase for the region as a whole, which accounted for just 1% of this investment in 2013. However, member states varied widely in terms of investment levels. Start-ups in the United Kingdom received 55% of the European Union total investment between 2011 and mid-2018, followed by German (14%) and French ventures (13%). This means the remaining 25 countries shared less than 20% of all private AI equity investments received in the European Union (Figure 2.3).
Figure 2.3. Private equity investments in AI start-ups based in the European Union, 2011 to mid-2018
Note: The percentage for 2018 only covers the first half of the year.
Source: OECD estimation, based on Crunchbase (July 2018), www.crunchbase.com.
Together, the United States, China and the European Union represent over 93% of total AI private equity investment from 2011 to mid-2018. Beyond these leaders, start-ups in Israel (3%) and Canada (1.6%) also played a role.
The volume of AI deals grew until 2017, but so did their size
The number of investment transactions grew globally, from fewer than 200 investment deals to more than 1 400 over 2011-17. This represents a 35% compound annual growth rate from 2011 to the first half of 2018 (Figure 2.4). Start-ups based in the United States attracted a significant portion of all investment deals, rising from 130 to about 800 over 2011-17. The European Union has also seen an increase in the number of deals, from about 30 to about 350 during the same period.
China-based start-ups signed fewer deals than companies in the United States or the European Union, going from none to about 60 over 2011-17. However, the high total value of investment in China implies the average value of these deals was considerably higher than in the European Union.
The large average size of investments in China is in line with a general trend of deals with an increase in per-investment value. In 2012 and 2013, close to nine out of ten reported investment deals were worth less than USD 10 million. Only one out of ten deals was worth between USD 10-100 million. No deals were worth more than USD 100 million. By 2017, more than two deals out of ten were larger than USD 10 million and close to 3% were larger than USD 100 million. The trend accentuated in the first half of 2018, with 40% of reported deals worth more than USD 10 million and 4.4% worth over USD 100 million.
In terms of value, “mega-deals” (those larger than USD 100 million) represented 66% of the total amount invested in AI start-ups in the first half of 2018. These figures reflect the maturing of AI technologies and investor strategies, with larger investments focused on fewer AI companies. For example, the Chinese start-up Toutiao attracted the largest investment in 2017 (USD 3 billion). The company is an AI-powered content recommendation system based on data mining that suggests relevant, personalised information to users in China through social network analysis.
Figure 2.4. Number of private equity investments in AI start-ups, by start-up location
Note: Estimates for 2018 may be conservative, as they do not account for a likely lag in reporting (see Box 2.1. Methodological note).
Source: OECD estimation, based on Crunchbase (July 2018), www.crunchbase.com.
Since 2016, Israel (Voyager Labs), Switzerland (Mindmaze), Canada (LeddarTech and Element AI) and the United Kingdom (Oaknorth and Benevolent AI) have all seen deals worth USD 100 million or more. This highlights dynamic AI activity beyond the United States and China.
Investment patterns vary across countries and regions
The total amount invested and the global number of deals have increased greatly since 2011, but with wide variations in investment profiles between countries and regions.
In particular, the profile of investments in Chinese start-ups appears different from that of the rest of the world. Individual private equity investments in Chinese AI start-ups registered in Crunchbase were worth an average of USD 150 million in 2017 and in the first half of 2018. By comparison, the average investment size in 2017 in other countries was just one-tenth of that amount.
Overall, three patterns can be observed. First, there are few Chinese start-ups, but they have large investments. Second, EU start-ups have a steadily increasing number of smaller investments. The average per investment increased from USD 3.2 million in 2016 to USD 5.5 million in 2017 to USD 8.5 million in the first half of 2018. Third, the United States has a steadily increasing number of larger investments. The average per investment increased from USD 9.5 million in 2016 to USD 13.2 million in 2017 to USD 32 million in the first half of 2018. These differences in investment profiles remain notable even when deals over USD 100 million are excluded from the sample (Table 2.1 and Table 2.2).
The investment profiles described above are not limited to AI start-ups. They are, instead, true across industries. In 2017, Chinese start-ups across all industries raised USD 200 million on average per investment round. Meanwhile, start-ups in the United States and the European Union raised on average USD 22 million and USD 10 million, respectively.
Figure 2.5. Size of investment transactions, 2012-17 and first semester 2018
Note: The percentages for 2018 only cover the first half of the year.
Source: OECD estimation, based on Crunchbase (July 2018), www.crunchbase.com.
Table 2.1. Average amount raised per deal, for deals up to USD 100 million
USD million
Source: OECD estimation, based on Crunchbase (April 2018), www.crunchbase.com.
Table 2.2. Average amount raised per deal, for all AI deals
USD million
Source: OECD estimation, based on Crunchbase (April 2018), www.crunchbase.com.
Autonomous vehicle start-ups are receiving significant funding
Levels of private equity investment in AI vary widely by field of application. AVs represent an increasing share of private equity investments in AI start-ups. Until 2015, AVs represented less than 5% of total investments in AI start-ups. By 2017, AVs represented 23% of the total, growing to 30% by mid-2018. The bulk of venture capital investment in AV start-ups went to US-based start-ups (80% between 2017 and mid-2018). This was followed by AV start-ups based in China (15%), Israel (3%) and the European Union (2%). The growth is due to a dramatic increase in the per-investment amount; the actual number of investments remained fairly constant (87 in 2016 and 95 in 2017). In the United States, the average amount per investment in this sector increased ten-fold from USD 20 million to close to USD 200 million between 2016 and the first half of 2018. This was in large part due to Softbank’s USD 3.35 billion investment in Cruise Automation. This self-driving car company owned by General Motors develops autopilot systems for existing cars. In 2017, Ford invested USD 1 billion in AV company Argo AI.
Broader trends in development and diffusion of AI
Efforts to develop empirical measures of AI are underway, but are challenged by definitional issues, among other concerns. Clear definitions are critical to compile accurate and comparable measures. Joint experimental work by the OECD and the Max Planck Institute for Innovation and Competition (MPI) proposes a three-pronged approach aimed at measuring i) AI developments in science, as captured by scientific publications; ii) technological developments in AI, as proxied by patents; and iii) AI software developments, in particular open-source software. The approach entails using expert advice to identify documents (publications, patents and software) that are unambiguously AI-related. These documents are then used as a benchmark to assess the degree of AI-relatedness of other documents (Baruffaldi et al., forthcoming[8]).
Scientific publications have long been used to proxy the outcome of research efforts and of advancements in science. The OECD uses bibliometric data from Scopus, a large abstracts and citations database of peer-reviewed literature and conference proceedings. Conference proceedings are particularly important in the case of emerging fields such as AI. They help provide a timely picture of new developments discussed at peer-reviewed conferences prior to being published. By establishing a list of AI-related keywords and validating them with AI experts, the approach aims to identify AI-related documents in any scientific domain.
The patent-based approach developed by the OECD and MPI Patent identifies and maps AI-related inventions and other technological developments that embed AI-related components in any technological domain. It uses a number of methods to identify AI inventions, including keyword search in patents’ abstracts or claims; analysis of the patent portfolio of AI start-ups; and analysis of patents that cite AI-related scientific documents. This approach has been refined through work under the aegis of the OECD-led Intellectual Property (IP) Statistics Task Force.1
Data from GitHub – the largest open-source software hosting platform – are used to help identify AI developments. AI codes are divided into different topics with topic modelling analysis to show key AI fields. General fields comprise ML (including deep learning), statistics, mathematics and computational methods. Specific fields and applications include text mining, image recognition or biology.
References
[1] Agrawal, A., J. Gans and A. Goldfarb (2018), Prediction Machines: The Simple Economics of Artificial Intelligence, Harvard Business School Press, Brighton, MA.
[8] Baruffaldi, S. et al. (forthcoming), “Identifying and measuring developments in artificial intelligence”, OECD Science, Technology and Industry Working Papers, OECD Publishing, Paris.
[7] Breschi, S., J. Lassébie and C. Menon (2018), “A portrait of innovative start-ups across countries”, OECD Science, Technology and Industry Working Papers, No. 2018/2, OECD Publishing, Paris, http://dx.doi.org/10.1787/f9ff02f4-en.
[2] Bresnahan, T. and M. Trajtenberg (1992), “General purpose technologies: ’Engines of growth?’”, NBER Working Paper, No. 4148, http://dx.doi.org/10.3386/w4148.
[3] Brynjolfsson, E., D. Rock and C. Syverson (2017), “Artificial intelligence and the modern productivity paradox: A clash of expectations and statistics”, NBER Working Paper, No. 24001, http://dx.doi.org/10.3386/w24001.
[6] CBI (2018), “The race for AI: Google, Intel, Apple in a rush to grab artificial intelligence startups”, CBI Insights, 27 February, https://www.cbinsights.com/research/top-acquirers-ai-startups-ma-timeline/.
[5] Dilda, V. (2017), AI: Perspectives and Opportunities, presentation at "AI: Intelligent Machines, Smart Policies" conference, Paris, 26-27 October, http://www.oecd.org/going-digital/ai-intelligent-machines-smart-policies/conference-agenda/ai-intelligent-machines-smart-policies-dilda.pdf.
[4] MGI (2017), “Artificial Intelligence: The Next Digital Frontier?”, Discussion Paper, McKinsey Global Institute, June, https://www.mckinsey.com/~/media/McKinsey/Industries/Advanced%20Electronics/Our%20Insights/How%20artificial%20intelligence%20can%20deliver%20real%20value%20to%20companies/MGI-Artificial-Intelligence-Discussion-paper.ashx.
Note
← 1. This took place with the advice of experts and patent examiners from the Australian IP Office, the Canadian Intellectual Property Office, the European Patent Office, the Israel Patent Office, the Italian Patent and Trademark Office, the National Institute for Industrial Property of Chile, the United Kingdom Intellectual Property Office and the United States Patent and Trademark Office.