2018 World AI Industry Development Blue Book 2018 World AI Industry Development Blue Book

Preface 3 As the driving force of a new round of technology Committed to taking the lead in China’s reform In-depth Report on the and opening-up and innovative developments, Development of the World AI and industrial revolution, artificial intelligence Industry, 2018 (AI) is wielding a profound impact on the world Shanghai has been working closely with economy, social progress and people’s daily life. related national ministries and commissions to Overview and Description Therefore, the 2018 World AI Conference will be jointly organize the 2018 World AI Conference. held in Shanghai from September 17 to 19 this year, Hosting this event becomes a priority in the Industrial Development jointly by the National Development and Reform process of Shanghai constructing the “five Environment Commission (NDRC), the Ministry of Science and centers” - international and economic center, Technology (MST), the Ministry of Industry and finance center, trade center, shipping center Technical Environment Information Technology (MIIT), the Cyberspace and technical innovation center, and building Administration of China (CAC), the Chinese the “four brands” of Shanghai - services, World AI Enterprises Academy of Sciences (CAS), the Chinese Academy manufacturing, shopping and culture. Also, of Engineering (CAE) and the Shanghai Municipal it is the practical measure to foster the reform World Investment and People’s Government, with the approval of the State and opening-up and optimize the business Financing Council, in order to accelerate the development environment in Shanghai. This meeting of a new generation of AI, complying with the provides the opportunity for Shanghai to Industrial Development development rules and grasping the future trends. accelerate industrial innovation and the It is organized as a mobilization conference to extensive application of AI and take efforts 40 make people join in planning for AI technologies to build national “AI highland”, looking to Hype Cycle for Artificial and industrial developments. It also works as become the innovation source of AI, the Intelligence, 2018 an important measure to promote the deep application demonstration zone, and the integration of the Internet, big data and AI with the industrial and talents clusters. real economy. As an important achievement of the conference, Themed “A NEW ERA EMPOWERED BY ARTIFICIAL this Blue Book is released by China Academy of INTELLIGENCE”, this conference is featuring Information and Communications Technology “Globalization, Sophistication, Specialization and (CAICT) and featuring research from Gartner, Marketization”. It boasts of being the world’s top the world’s leading information technology AI platform for cooperation and communication, research and advisory company. With the where get together the most influential scientists, opportunity of this conference, the Blue Book entrepreneurs and investors from the AI industry all is expected to open a window for Chinese AI around the world, as well as related government enterprises to communicate with global AI leaders and city administrators who will deliver giants. And it makes a comprehensive and in- speeches and engage in various high-level depth analysis of the developments of global dialogues surrounding the leading-edge AI industry and the trends of technologies technologies, industrial trends and hot issues in routes. Hopefully, the practitioners and the current AI field. This meeting is expected to researchers who are working on AI and other share authoritative opinions and consensus, and readers who are interested in AI could get exhibit new technologies, products, applications many inspirations from this report and use it and ideas, in an effort to bring out the “Chinese as a reference to learn about AI. Solutions” and the “Global Wisdom” in order to The Organizing Committee of WAIC address the common challenges facing human beings and build a better life for people.

2018 World AI Industry Development Blue Book is published by CAICT. Editorial supplied by CAICT is independent of Gartner analysis. All Gartner research is © 2018 by Gartner, Inc. All rights reserved. All Gartner materials are used with Gartner’s permission. The use or publication of Gartner research does not indicate Gartner’s endorsement of CAICT’s products and/or strategies. Reproduction or distribution of this publication in any form without prior written permission is forbidden. The information contained herein has been obtained from sources believed to be reliable. Gartner disclaims all warranties as to the accuracy, completeness or adequacy of such information. Gartner shall have no liability for errors, omissions or inadequacies in the information contained herein or for interpretations thereof. The opinions expressed herein are subject to change without notice. Although Gartner research may include a discussion of related legal issues, Gartner does not provide legal advice or services and its research should not be construed or used as such. Gartner is a public company, and its shareholders may include firms and funds that have financial interests in entities covered in Gartner research. Gartner’s Board of Directors may include senior managers of these firms or funds. Gartner research is produced independently by its research organization without input or influence from these firms, funds or their managers. For further information on the independence and integrity of Gartner research, see “Guiding Principles on Independence and Objectivity” on its website.

2 3 Research from CAICT In-depth Report on the Development of the World AI Industry, 2018

1 Overview and Description and technologies, the developments of the products (visual products, voice terminals, 1.1 Report Overview global AI enterprises, as well as the industrial robots, intelligent cars, drones). applications in AI fields, which helps to reflect This Blue Book will share the research results the situations of the current developments in 2) The Map of Industrial Distribution and practical experiences in the field of AI industry more clearly. This section has mapped out the distribution, AI, comprehensively analyzing the current scale and name of the main enterprises in the trend of the industrial developments and 1.2 Report Description country where the AI industry is leading the technologies in the world’s AI industrial world, as well as in various Chinese provinces powers, and elaborating on the technological The world AI industry map includes the and cities. research, industrial inputs, application following sections: (1) the map of industrial services in the AI field, in order to provide chain (2) the map of industry distribution 3) Industrial Research Institutions and Policy basic information and guidance for the (3) the industry research institution and This section has mapped out the developments of AI technology and industry. supporting policy. mainstream AI industrial research This Blue Book has been compiled by institutions and alliances in China and a group of editors together, by sorting 1) The Map of Industrial Chain across the world, and the industrial policy out the AI technologies and the artificial The section of AI industrial chain falls into documents. intelligence developments in industry and three layers: basic supporting layer, software application, and analyzing the technology algorithm layer and industry application layer. The structure of this report is divided fads, enterprise conditions, industrial into five parts: 1) industrial development applications and future trend of artificial The map of basic layer depicts the typical environment 2) technical environment 3) intelligence, based on thorough and enterprises engaged in the basic industry in The situations of world AI enterprise 4) extensive investigation in the developments the world. It is drawn in three dimensions: The situations of world investment and of the world AI industries, in accordance with computing hardware (cloud training, cloud financial 5) The situations of industrial the research data and relevant materials inference, device inference, intelligent chips, developments provided by authoritative departments. intelligent sensors) computing system technology (cloud computing, big data, 5G 1) Industrial Development Environment This report mainly discusses on the world communication and Internet of Things) and This part has comprehensively elaborated AI industry map, which is jointly formulated data (data acquisition, labeling and analysis). on the foundation and environment of the by the Research Institute of Information and AI industrial developments , discussing Industrialization and the Data Research The map of software algorithm layer the development progress of the artificial Center of CAICT. This map has systematically describes the typical players working on intelligence, the AI related policies in various analyzed and drawn the world AI industry the software algorithm industry across countries all around the word, and the from the underlying technology to the vertical the world mainly on the level of algorithm conditions of AI developments. application, from the industrial structure to theory (machine learning algorithm, brain- the industrial distribution, from the research like algorithm, and knowledge graph), 2) Technical Environment institutions to the government policy, development platform (basic open source This part has derived the current technical reviewing the trend of AI developments, in framework, technology open platform) and environments in the field of artificial order to promote the AI industry and provide application technology (computer vision, intelligence in terms of patents, papers, fundamental decision references for the natural language processing and human research environments and technical development of the AI industry. Meanwhile, computer interaction). competitions all around the world. East China Branch of China Academy of Information and Communications Technology The map of industry application layer 3) The Situations of World AI Enterprises has further expanded the contents of the pictures the main enterprises working on This part describes the situations of World AI industrial map by designing the overall the application-oriented industry across enterprises in the aspects of structure, scale framework of this report, and expounding two dimensions: industrial solutions (“AI+” and regional distribution. the environments of industrial development industrial vertical application) and typical

3 4) The Situations of World Investments 3) Patent Data 2 Industrial Development This section presents the distribution, The patent data in this report are derived Environment scale and rounds of world investment from the research results of the Intellectual 2.1 AI Development Milestones and financing in the field of AI, analyzing Property Center of CAICT. The Intellectual The history of AI can be roughly divided into the developments of developments of AI Property Center conducts search and three stages: the first stage (1956-1980) AI industry in AI industry, in the perspective of statistics on AI patents worldwide based on was born; the second stage (1980-2000) AI investment and financing, and the scale of professional databases such as PatSnap. stepped into industrialization; the third stage the industry. 4) Paper Data (2000-present) AI ushers in the explosion. 5) The Situations of Industrial Development The paper data in this report are derived The first stage (1956-1980) AI was born This section expounds the current situation from the core collection of Web of Science, and trend of technologies and applications which is based on the results of the data in the global AI industry, as well as several retrieval center’s AI keyword vocabulary. Date Iconic event application cases of model AI enterprises. 1956 The Dartmouth Conference in 5) Industrial Application Data the US gathered the first batch The specific research scope and data The industrial application data in this report of researchers to determine sources of this report are as follows: come from Qixin which is the product of CCi the name and mission of AI, Intelligence Co., Ltd. Besides, the data are also which was called the birth of 1) AI Enterprise from the related forecasts of major websites AI. The information of AI enterprises mentioned and research institutions such as CAICT,PwC, 1957 Frank Rosenblatt, an in this report came from the Monitoring MarketsandMarkets, Grand View Research, experimental psychologist Platform in the Data Research Center of IFR, Roland Berger, ASKCI Consulting, and at Cornell University, CAICT, in which the AI enterprises refer to Forward Industry Research Institute. implemented a neural network the enterprises that can provide AI products, “perceptron”. services and related solutions. Enterprises 6) Data Unit Description can be divided into two dimensions: The default currency for all market data in 1969 The International Federation technical level and product or solution level. this report is RMB, if not specified. of Artificial Intelligence was The technology dimension includes providers established and the first and manufacturers of general technologies This Blue Book does not seek to cover meeting was held in Seattle, such as algorithmic platform, basic all aspects; it only gives an analysis and Washington, US hardware, speech and vision. The product or description over the current industrial and solution dimensions include manufacturers technological development environment, and solution providers of various types of AI world enterprise situation and industrial products, as well as solution providers for application in the field of AI. It does not vertical industries. give much viewpoint statements on the AI field, trying to describe in a language and 2) Investment and Financing manner that is easy to understand. For the The data of investment and financing in main contents of this Blue Book, experts and this report come from the investment and scholars from various fields are welcome to financing websites such as CB insights, make suggestions. We are glad to listen to IT oranges, and sprouts. It is the result of the opinions of experts from all sides and matching and collating statistics based on continue to make improvements. the AI business directory.

4 5

The second stage (1980-2000) AI stepped into industrialization

Date Iconic event 1980 Carnegie Mellon University designed an expert system called XCON for DEC, which was a huge success, and at that time it saved the enterprise USD 40 million each year. 1982 Japan planned to invest USD 850 million to develop AI computers (the fifth-generation computers), aiming to create machines that can talk to people, translate languages, interpret images, and reason like humans. 1986 Multi-layer neural networks and BP back-propagation algorithms have emerged to improve the accuracy of automatic recognition. 1988 The German Research Centre for Artificial Intelligence was established and is currently the world’s largest non-profit AI research institution. 1997 Deep Blue, a chess-playing computer developed by IBM, defeated the world chess champion, a milestone event in the history of AI; under the influence of Moore’s Law, computing performance began to increase dramatically.

The third stage (2000-present) AI ushers in the explosion Date Iconic event 2006 Geoffrey Hinton proposed a training algorithm in “Science” based on Deep Belief Networks (DBN) that can use unsupervised learning, making deep learning continue to heat up in academia. 2011 The IBM Watson system won at the US game show Jeopardy! against human players. 2012 The deep learning algorithm became well-known after the ImageNet Challenge, and was thereby widely used. 2016 AlphaGo developed by DeepMind defeated former World Go champion Lee Sedol.

2.2 AI Policies of Major Countries The rapid development of AI will profoundly change the face of human society and the world. In order to seize the strategic opportunities of AI development, more and more countries and organizations have rushed to set up the national-level development plans.

US Date Organization Policy released 1998 The Networking and Information Technology Research Next Generation Internet Research Act (P.L. 105-305) and Development Program (NITRD) 2013 The White House National Robot Project: A Roadmap for U.S. Robotics from Internet to Robotics 2013 Edition 2013.4 The White House Brain Research through Advancing Innovative Neurotechnology 2015.10 National Economic Council (NEC)/Office of Science and The new edition of the US National Innovation Technology Policy (OSTP) Strategy 2015.11 Center for Strategic and International Studies (CSIS) Defense 2045 2016.10 National Science and Technology Council (NSTC) / National Artificial Intelligence Research and Networking and Information Technology Research and Development Strategic Plan Development Subcommittee (NITRD) 2016.10 National Science and Technology Council (NSTC) Preparing for the Future of Artificial Intelligence 2017.9 US Congress SELF DRIVE ACT AV START ACT 2017.10 US Information Industry Council AI Policy Principles

5 China

Date Organization Policy released 2015.7 State Council Guiding Opinions on Actively Promoting the “Internet +” Action Plan 2016.3 State Council Outline of the Thirteenth Five-Year Plan for National Economic and Social Development 2016.4 Ministry of Industry and Information Technology Robot Industry Development Plan (2016-2020) (MIIT)/ National Development and Reform Commission (NDRC)/ Ministry of Finance (MOF) 2016.5 CPC Central Committee and the State Council National Innovation Driven Development Strategy Outline 2016.5 National Development and Reform Commission The Three-year Implementation Program for “Internet (NDRC)/Ministry of Science and Technology (MST)/ +” Artificial Intelligence Ministry of Industry and Information Technology (MIIT)/Cyberspace Administration of China (CAC) 2016.7 State Council 13th Five-year Plan on National Scientific and Technological Innovation 2017.3 State Council Report on the Work of the Government 2017.7 State Council Development Planning for a New Generation of Artificial Intelligence 2017.12 Ministry of Industry and Information Technology (MIIT) Three-Year Action Plan for Promoting the Development of a New Generation of Artificial Intelligence Industry (2018-2020) 2018.4 Ministry of Education (MOE) Innovative Action Plan for Artificial Intelligence in Colleges and Universities

Other countries and international organizations

Date Organization Policy released 2013.6 Japanese Cabinet Japan Revitalization Strategy 2015.1 Ministry of Economy, Trade and Industry New Strategy of Japan for Robots 2016.5 Japanese Cabinet Science and Technology Innovation Comprehensive Strategy 2016 2016.6 Headquarters for Japan’s Economic Revitalization Japan Rejuvenation Strategy 2017 Japanese Government Promotion Strategy for Next Generation Artificial Intelligence 2017.5 Ministry of Economy, Trade and Industry Vision of New Industrial Structure 2017.6 Japanese Cabinet Comprehensive Strategy on Science, Technology and Innovation for 2017 2018.6 Japanese Cabinet Strategy for Comprehensive Innovation 2018.6 Japanese Cabinet Future Investment Strategy 2018 2013 Korea Electronics and Telecommunications Research Exobrain Plan Institute

6 7

Date Organization Policy released 2014.7 Ministry of Trade, Industry and Energy, Republic of The second Master Plan to develop intelligent robots Korea (2014-2018) 2016.8 Korean Government Nine National Strategic Projects 2017.7 Korean Parliament Robotic Basic Act 2018.5 The Fourth Industrial Revolution Committee Artificial Intelligence Research and Development (R&D) Strategy 2013 The UK Government The Eight Great Technology 2016.12 The UK Government Office for Science Artificial intelligence: opportunities and implications for the future of decision making 2017.1 The UK Government building our industrial strategy 2017.10 The UK Government Growing the Artificial Intelligence Industry in the UK 2010.7 German Government Ideas. Innovation. Prosperity. High-Tech Strategy 2020 for Germany 2011.11 German Government “Industry 4.0” as the Strategic Focus 2013.4 German Ministry of Education / Research “Industry Recommendations for Implementing the Strategic 4.0 Working Group” Initiative INDUSTRIE 4.0-- Securing the Future of German Manufacturing Industry 2017.6 German Federal Ministry of Transport and Digital Automated and Connected Driving Report Infrastructure 2013 French Government French Robot Development Plan 2017.3 French Ministry for the Economy and Finance/ Artificial Intelligence Strategy Ministry of National Education 2018.5 French Government Make Sense to Artificial Intelligence 2017.5 Singapore National Research Foundation “AI.SG” National Artificial Intelligence Program 2013.1 European Union (EU) Human Brain Project 2013.12 European Commission / European Robotics SPARC Program Association 2015.12 EU SPARC Robotics 2020 Multi-Annual Roadmap 2016.6 European Commission Proposed AI legislative motion 2016.10 European Parliament’s Committee on Legal Affairs European Civil Law Rules in Robotics (JURI) 2017.10 European Union (EU) Horizon 2020 2016.8 World Commission on the Ethics of Scientific Draft Preliminary Report on Robot Ethics Knowledge and Technology (COMEST) 2016.12 Institute of Electrical and Electronics Engineers (IEEE) Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems (1st Edition) 2017.12 Institute of Electrical and Electronics Engineers (IEEE) Ethical Guidelines for Artificial Intelligence Design (2nd Edition)

7 2.3 Development Conditions Table 2-1 Deep Learning Framework 2.3.1 Evolution of the Algorithms Framework Organization Language(s) supported Introduction The development of AI algorithms has continued to innovate and the learning level TensorFlow Google Python/C++/Go/… Open-source neural network has been increasing. The early research library in the academic field focused on symbolic Caffe UC Berkeley C++/Python Open-source convolutional neural computing. The artificial neural network was network framework completely negated in the early stage of AI PaddlePaddle Baidu Python/C++ Open-source deep learning development, then gradually recognized, platform and became a large-scale algorithm that CNTK Microsoft C++ Deep learning computing leads the development trend of AI today, network toolkit showing strong vitality. The current popular algorithms of machine learning and deep Torch Facebook Lua Machine learning algorithm open learning are actually further extensions of source framework the theory of Symbolism, Evolutionism, and Keras Google Python Modular neural network library Connectionism. API Theano University of Python Deep learning library Machine learning algorithms and deep Montreal learning algorithms are two hot points in AI. The open source framework has become DL4J Skymind Java/Scala Distributed deep learning library the focus of the overall development of MXNet DMLC C++/Python/R/… Open-source deep learning community library

Fig. 2-1 AI Setting off a New Wave of Technological Development

Source: CAICT (2018)

8 9

technology giants. The open source deep AI is still evolving rapidly and changing 2.3.2 Increase in Computing Power learning platform is an important driving people’s lives, and more AI algorithms are The implementation of AI algorithms require force for the development of AI technology. waiting for computer scientists to explore. strong computing power support, especially The open source deep learning platform Due to the long technology investment cycle, the large-scale use of deep learning allows the public to use, copy and modify most AI enterprises in China still lack original algorithms, which put higher demands on the source code. It has the characteristics algorithms, and they still need to plan ahead computing power. AI has ushered in a real of fast update and strong expansion, which and pay attention to the talent reserve of AI explosion since 2015, which is largely related can greatly reduce the development cost algorithm level; combine academic research to the widespread use of GPUs. Prior to this, of enterprises and the purchase cost of and industrial application scenarios, hardware computing power could not meet customers. These platforms are widely encourage innovation, actively explore the needs of AI computing power. When used by enterprises to quickly build a talents in AI algorithm, and let AI researchers GPU and AI were combined, AI ushered in a deep learning technology development with strong potentials truly enter the industry. real high-speed development. Therefore, the environment and promote the accelerated improvement of hardware computing power iteration and maturity of their own is one of the important factors for the rapid technology, so as to ultimately achieve the development of AI. application of the products. Fig. 2-2 Development of AI Computing Power

Source: CAICT (2018)

Fig. 2-3 AI High Performance Computing Unit

Source: CAICT, Internet

9 In recent years, the new high-performance Fig. 2-4 Tesla V100 Training and Reasoning Performance Comparison computing architecture has become the catalyst for the evolution of AI technology. With the emergence of the deep learning boom in the AI field, the architecture of computing chips have gradually evolved toward the trend of deep learning application optimization, from the traditional Intel processor that takes CPU as the main and GPU as the auxiliary to a structure with GPU as the main and CPU as the auxiliary. In 2017, NVIDIA launched Tesla V100, the new generation of graphics processing Source: NVIDIA official website chip, which is mainly used to study AI that is based on deep learning. For TensorFlow, the Google open source deep learning requirements of dense linear algebra and the dedicated chip ASIC for neural network framework, Google launched a customized massive data in terms of basic capabilities. algorithms is being launched one after another TPU for machine learning. It is urgent to have high performance and by Intel, Google, NVIDIA and many startups versatility of the cloud, as well as high energy and is expected to replace the current general- The development of AI needs urgently efficiency and low latency of the terminal. purpose chip and serve as the main force of AI core hardware upgrades and computing chips in the next few years. innovation becomes the focus of From the development stage of AI chips, deployment. The existing chip products general-purpose chips such as CPU, GPU and cannot meet the high-throughput FPGA are the main chips in the AI field, and

Fig. 2-5 AI Chip Industry Map

Source: CAICT (2018)

10 11

Fig. 2-6 Heating up Battle for AI Chips

Source: CAICT (2018)

2.3.3 Data Support of Internet users worldwide has exceeded Public data sets built by academic and From the software age to the Internet, 4.1 billion, and the penetration rate of global research institutions are keeping enriched to today’s big data era, the volume and unique mobile users has reached 67% of the and promote the growth of start-ups. Public complexity of data have undergone a total population. data sets are generally used in algorithmic quantitative to qualitative change, and big testing and competency competitions, data has led the development of AI into an Big data is a booster for the development with high quality, providing quality data for important strategic window. of AI, because some AI technologies use innovation, entrepreneurship and industry statistical models to calculate the probability competition, and bringing essential resources Data is the cornerstone for the development of data, such as images, text or voice, by to start-ups. of AI, and the core of AI lies in data support. exposing these models to the ocean of data, From the status quo of development, AI so that they are constantly optimized, or The industry data set is the core technology has made rapid progress thanks “trained”. With the support of big data, the competitiveness of the enterprises. The to a good big data foundation, and massive output of the deep learning algorithm will industry data set is closely integrated with data provides raw materials for training be more accurate as the amount of data the industry. The self-built data sets of each AI. According to the 2018 Q3 global digital processing increases. enterprise belong to its core competitiveness. statistics report of We Are Social, the number The data service sector has developed rapidly, including data set construction, data cleaning, and data labeling.

11 Table 2-2 Partial Global AI Public Data Set Type Data set name Characteristics WikiText Wikipedia corpus Natural Language SQuAD The Stanford question answering dataset Processing Common Crawl PB level web crawler data Billion Words Common language modeling database VoxForge Accented corpus Speech Recognition TIMIT Acoustic-Phonetic continuous speech corpus CHIME Speech recognition data set containing ambient noise SVHN Image dataset in Google street view Machine vision ImageNet Commonly used image dataset based on wordnet composition Labeled Faces in the Wild Face area image data set for face recognition training

Fig. 2-7 Industry database classification

Source: CAICT (2018)

Fig. 3-1 Global AI Related Patent Application Trend 3 Technical Environment 3.1 Patent From 1999 to 2017, the number of invention applications and authorized patents in the AI field, such as image recognition, biometrics, speech recognition, speech synthesis, natural language understanding and machine learning, had exceeded 100,000. In China, the number of patent applications and authorizations for AI has increased year by year since 2010, and has achieved rapid growth since 2014.

Source: CAICT (2018)

12 The number of patent applications in China, the US and Japan is From the perspective of the patentee, the US, Japan, Korea’s ahead in the world, and China has overtaken the US with the highest technology giants in the field of artificial intelligence patent patent applications in the field of artificial intelligence. The total accumulation has a leading edge, Microsoft patent applications in number of patent applications of China, the US and Japan accounted the world’s first, followed by IBM and Google. for 75% of the global AI patents. Fig. 3-4 Main Applicants for Global AI Related Patents Fig. 3-2 Geographical Distribution of Global AI Patent Applications

Source: CAICT (2018)

Source: CAICT (2018)

3.2 Paper China’s AI patent applications mainly come from five provinces and cities, namely, Beijing, Guangdong, Jiangsu, Shanghai and Zhejiang. Between 1998 and 2018, the number of papers and journals in the Among them, the number of AI patents applied and authorized by AI field increased significantly in the world, with the total number Beijing exceeds 10,000. exceeding 630,000 and a CAGR of 11.59%. In 2017, the core collection of Web of Science had a total of 65,100 AI papers and academic publications such as academic publications in journals. Among them, AI papers and academic publications such as academic publications in Fig. 3-3 Geographical Distribution of AI Patent Applications in China journals of China (including Hong Kong, Macao and Taiwan) reached 17,300, with a CAGR of 24.32%. The global ratio of artificial intelligence in China increased from 5.52% in 1998 to 26.63% in 2017, indicating that China’s strength in the field of AI research has increased dramatically.

In terms of global growth trend, 1998-2017 was largely sustained, with an average annual growth rate of more than 10% per cent in both 2001- 2007 and 2012-2016, and a decline in 2008-2011 and a decline in 2010.

Source: CAICT (2018)

13 Fig. 3-5 Global/China AI Paper Publication Trend The US has always been a leader in the world of artificial intelligence research, with a higher number of academic studies than in other countries. However, the number of academic publications on AI in China has grown rapidly, and in 2009 and 2014 years later, China surpassed the US in the first place. In addition, India has developed rapidly in the AI field research since 2013. In 2014, its number of AI papers published exceeded that of the UK, ranking third in the world.

Fig. 3-7 Trend of AI Paper Output in Major Countries

Source: CAICT (2018)

From 1998 to 2018, the largest number of papers in the AI field in the world was 149,100 in the US, and China ranked second in terms of 141,800. The UK, Germany, and India ranked three to five.

Fig. 3-6 TOP 10 Countries for AI Paper Output Source: CAICT (2018)

In China, the most productive institutions for AI papers are the Chinese Academy of Sciences, Tsinghua University and Harbin Institute of Technology ranking the second and third. In terms of number, with more than 10,000 papers, the Chinese Academy of Sciences is far ahead of the second-ranked Tsinghua University (more than 4,500) and other institutions.

Fig. 3-8 Trend of AI Paper Output in China

Source: CAICT (2018)

Source: CAICT (2018)

14 Among the highly cited literature included in Web of Science, the Scientific research institutions represented by universities have strong number and proportion of China’s AI related literature has increased scientific research strength in the AI field, and stand at the forefront rapidly since 2012, from less than 15% in 2008 to 47% in 2017, of the development of AI theory and technology. Scientific research indicating that the quality of scientific research in the AI field in China institutions with a certain scale of scientific research team, with more has improved considerably. advanced laboratories and research facilities, and other institutions such as enterprises have a unique comparative advantage. Scientific Fig. 3-9 Trend of Highly Cited AI Paper Output in China research institutions can also continuously cultivate high-end artificial intelligence talents for the sustainable development of the industry to provide back-up forces.

The Digital Bibliography & Library Project (DBLP), jointly maintained by the University of Trier and the Schloss Dagstuhl – Leibniz Information Center, contains a large number of journals and literature in the computer field and provides literature search service based on metadata from the scientific literature in the computer field. In the past five years, DBLP has included 300,000-400,000 articles per year. As of now, there are more than 4.2 million DBLP index documents. Based on this, statistics can be made for the research institutions and scholars who have published literature on the AI field collected by DBLP every year, so as to understand the degree of attention paid by the academic community to the AI field.

Source: CAICT (2018) According to the literature data of the DBLP index about the number of academic institutions and scholars who have literature published 3.3 Research Environment in the AI field in 2013-2017, in many countries, the attention paid to AI improved significantly in 2015; after cooling down in 2016, it reached Global graduates of science, technology, engineering and a high point in 2017. See the figure below for the number of research Mathematics (STEM) are growing every year, and China ranks first in institutions and scholars who have published papers in the DBLP the world. According to the World Economic Forum, 4.7 million Chinese index literature in recent years. graduates were from the STEM field in 2016. In addition, China has a

total of 30,000 STEM PhD graduates each year. After China, the second- ranked country is India. In 2016, India had 2.6 million STEM graduates, Fig. 3-11 Attention to AI Paid by the Scientific Research Community including local and overseas graduates. India is second only to China in the world not only in the total number of STEM graduates, but also in the number of students studying abroad, with 26% of these students majoring in computer science and mathematics.

Fig. 3-10 Number of STEM Graduates in the Countries (2016)

Source: CS Rankings, CAICT

Source: The World Economic Forum, CAICT

15 The US ranks first in AI technology in the world, and its leading 3.4 Technical Competition advantage is obvious. The number of both the AI research institutions The technological development of AI is not only reflected in scientific and scholars account for nearly half of the world. The number of research works such as patent papers, but also in various technology Chinese scientific research institutions is comparable to that of the challenges held around the world. According to the competition data UK and India. Its number of scholars is less than that of the US, but released on Kaggle, there are currently 19 competitions in progress, is significantly higher than that of other countries and has shown an and 275 competitions have been completed. The content of the overall upward trend in recent years, with 2017 increasing by about competition involves image recognition, speech recognition, object 30% over 2013. detection, classification, as well as predictive problems in a variety of scenarios. The number of participating teams in different competitions Fig. 3-12 TOP10 Countries for Number of Scientific Research Institutions is also different. Some popular competition teams can reach tens of with Publications thousands, and the prize pool can reach millions of US dollars.

ImageNet’s annual Large-Scale Visual Identity Challenge (ILSVRC) is an early and influential event in the field of computer vision. It is held once a year since 2010 and the competition process will classify and detect objects and scenes. In terms of the accuracy of classification of objects, the human level is 95%. Since 2015, the best AI system has outperformed the humans. In the 2017 ILSVRC competition, the classification accuracy of the AI system has reached 97.5%.

Fig. 3-14 ILSVRC Best Classification Accuracy

Source: CS Rankings, CAICT

Fig. 3-13 TOP10 Countries for Number of Scholars with Publications

Source: CAICT (2018)

The more influential competition in the field of NLP is the system quiz contest based on the Stanford Question Answering Dataset (SQuAD). SQuAD is a data set of reading comprehension, which is composed by crowdsourcing workers who ask questions from many Wikipedia articles. The answer to each question is a combination of text and cross-paragraph content in the corresponding reading paragraph, or the problem itself has no solution. Since the release of SQuAD 1.0, the community has made great strides, and the optimal models have been comparable to human performance. The following is the ExactMatch (EM) of the optimal model evaluated on the v1.1 test set. Source: CS Rankings, CAICT

16 Fig. 3-15 Evolution Trend of SQuAD Optimal Model In recent years, Chinese enterprises have also begun to pay attention to and actively organize and participate in AI challenge. At the CVPR conference this year, Baidu Apollo and the University of California at Berkeley jointly organized an autopilot seminar and defined a number of challenging tasks based on the ApolloScape large-scale data set, in which Megvii Technology, the unicorn enterprise in the visual field, defeated DeepMind and won the first place. In addition, in the video behavior recognition challenge, Chinese enterprises also performed well, taking the top three. In addition to the CVPR Challenge, in a growing number of top international challenges, Chinese enterprises and teams are doing better and better, repeatedly winning the title, indicating that China’s AI technology is moving to the forefront of the world.

4 World AI Enterprises 4.1 Enterprises Structure Globally, AI enterprises focus on AI+ (all vertical areas), Big Data & Data Services, vision and intelligent robotics. Among them, AI+ enterprises are mainly concentrated in business (mainly in the field of Source: CAICT (2018) marketing and customer management), healthcare, and finance.

For China, the vertical areas of AI enterprises are also concentrated. Among all kinds of vertical industries, AI penetrates more in the fields of healthcare, finance, commerce, education and security. Among them, the healthcare field accounts for the largest proportion of 22%, followed by finance and business intelligence, accounting for 14% and 11%, respectively.

Fig. 4-1 Global AI Enterprise Structure

Source: CAICT (2018)

17 Fig. 4-2 China AI+ Enterprise Structure

Source: CAICT (2018)

Fig. 4-3 Distribution of Global AI Enterprises 4.2 Business Scale According to the real-time monitoring of the Data Research Center of the global ICT monitoring platform of CAICT, 4,998 AI enterprises were monitored worldwide as of the first half of 2018. Among them, the number of AI enterprises in the US was 2,039, ranking first in the world, followed by 1,040 in China (excluding Hong Kong, Macao and Taiwan), 392 in the UK, 287 in Canada and 152 in India. In addition, the number of artificial intelligence enterprises in Israel, France and Germany has also surpassed 100.

Source: CAICT (2018)

18 From the perspective of start-up time of enterprises, the global trend of Fig. 4-5 TOP 20 Cities in Terms of Number of AI Enterprises AI entrepreneurship was concentrated in 2014 to 2016, with 2015 seeing the largest number of new AI enterprises in 2015, both globally and in China. During 2015, the number of newly established AI enterprises in the world reached 847, including 238 in China. Since 2016, the number of new startups in the world has begun to decrease, and the pace of entrepreneurship has slowed down. There were 738 new start-ups in the world, and by 2017 this number had dropped to 324.

Fig. 4-4 Time of Establishment of AI Enterprises

Source: CAICT (2018)

In China, AI enterprises are mainly concentrated in Beijing, Shanghai Source: CAICT (2018) and Guangdong Province. Beijing ranks first, followed by Shanghai and Guangdong. In addition, Zhejiang Province and Jiangsu Province have also gathered a number of AI enterprises.

4.3 Business Area Fig. 4-6 Number of AI Enterprises in Major Provinces of China Among the top 20 cities in terms of number of AI enterprises, 9 are in the US, 4 in China, 3 in Canada, and 1 respectively in the UK, Germany, France and Israel. Among them, Beijing has the largest number of AI enterprises (totally 412) in the world. The second is San Francisco and London, with 289 and 275 AI enterprises respectively. The number of AI enterprises in Shanghai, Shenzhen and Hangzhou has also entered the global Top 20.

Source: CAICT (2018)

19 5 World Investment and Financing Fig. 5-1 Distribution of Numbers of Global AI Investment and Financing Projects (2013-2018 Q1) 5.1 Distribution of Investment and Financing In the past five years, the global investment in AI industry focused on AI+ (vertical industry), vision, big data, data service and intelligent robotics. Among the various AI+ vertical industries, the most favored areas for capital are business intelligence, healthcare and finance.

The quarterly investment and financing data show that the healthcare, vision, business intelligence and intelligent robotics sectors have maintained a high investment intensity since the second quarter of 2017, especially when it peaked in the third quarter, and then followed by a slight fall. The big data and data services field, as an evergreen tree in Source: CAICT (2018) the AI field financing, compared with other areas in recent years, has attracted more capital investment. In general, it is also the field with the highest number of investments.

Fig. 5-2 Distribution of Global Investment Intensity in Various AI Fields (2013-2018Q1)

Source: CAICT (2018)

20 21

Fig. 5-1 Distribution of Numbers of Global AI Investment and Financing Projects (2013-2018 Q1) By comparing the number of financing Fig. 5-3 Distribution of Number of AI Financing Projects in Major projects in various AI fields in China, the US, Countries (2013-2018Q1) the UK, Canada and other major countries, it is found that in China the distribution of fields is more balanced, while other major countries have more emphasis on vertical industry applications (AI+).

5.2 Investment and Financing Scale Since 2013, the global enthusiasm for investment and financing in the AI field has continued to rise. In 2017, the total scale of global AI investment and financing reached 39.5 billion USD, and China’s investment and financing scale reached 27.7 billion USD, accounting for 70%, making it the world’s most capital-absorbing country in the AI field. In contrast, the US accounts for 41% of the number of investment and financing projects, surpassing China, as the most active country in terms of investment and financing.

Source: CAICT (2018)

Fig. 5-4 Global (including China)/China AI Investment and Financing Trend

Source: CAICT (2018)

21 Fig. 5-5 Geographical Distribution of Global AI Investment and Financing (2013-2018 Q1)

Source: CAICT (2018)

5.3 Distribution of World Rounds of Fig. 5-6 Numbers of Global AI Investment and Financing Projects Investment and Financing In the past five years, globally, the number of financing rounds in the AI field showed a year-on-year growth trend, except for a slight fluctuation in the E and F rounds. In 2017, the annual growth rate was more than 40% for each of the A, B, C, D, and E rounds.

Source: CAICT (2018)

22 Early investment in the AI field in the world continued to be active, By comparing the trend of number investment projects in the seed with the seed/angel round and A round investment being the round/angel round/A round in China, the US, the UK, Canada and highest, accounting for about 70% in total. With the development and India, it is found that the proportion of early investment in all these maturity of the AI industry and technology, the proportion of B, C, and countries decreased in 2017, and China has declined obviously since D rounds of financing has increased year by year, reaching about 2016. These trends indicate that China’s AI industry has experienced a 20% in 2017. period of entrepreneurial outbreaks and has begun to move toward a more mature stage of development.

Fig. 5-7 Proportion of Number of Global AI Financing Projects of the Rounds

Source: CAICT (2018)

Fig. 5-8 Total Proportion of Number of AI Financing Projects of Angel/Seed/A Round of Countries

Source: CAICT (2018)

23 6 Industrial Development global AI chip market is expected to reach USD 10.8 billion by 2023, 6.1 Industrial Development Technology with a CAGR of 53.6% during the forecast period (2017-2023). The Development Planning for a New Generation of Artificial Intelligence 6.1.1 Intelligent Hardware predicts that by 2020, China’s intelligent computing chip market will Intelligent sensors and intelligent chips are an important part of reach RMB 10 billion. Intelligent hardware. If the intelligent chip is the central brain of AI, then the intelligent sensor belongs to the nerves with nerve In the global intelligent hardware market, international giants such endings. Different from traditional hardware, the intelligent sensor as Honeywell, BOSCH and ABB have comprehensively deployed a is a relatively independent intelligent processing unit with primary variety of types of intelligent sensors product; in China, there are also sensing processing capability that integrates traditional sensors, fingerprint sensors from Goodix and force sensors from ColliHigh, microprocessors and related circuits. The intelligent chip has high- but the product layout is relatively simple. In terms of intelligent chips performance parallel computing capability and supports mainstream around the world, there are NVIDIA’s GPU, Google’s TPU, Intel’s NNP artificial neural network algorithms. Intelligent sensors mainly and VPU, IBM’s True North, ARM’s DynamIQ, Qualcomm’s Snapdragon include touch, visual, ultrasonic, temperature, distance sensors, etc.; series, Imagination’s GPU Power VR and products of other mainstream intelligent chips mainly include GPU, FPGA, ASIC and brain-like chips. enterprises; in China, there are Huawei HiSilicon’s Kirin series, Cambrian’s NPU, Horizon’s BPU, Westwell Lab’s Deepsouth and The ResearchAndMarkets report shows that the global market value Deepwell, Unisound’s UniOne, and Alibaba DAMO Academy’s Ali-NPU. of intelligent sensors in 2017 was USD 26.906 billion. It is estimated that the total market size will reach USD 70.617 billion by 2023, and the CAGR will be 17.45% within the forecast period (2018-2023). The

Fig. 6-1 Typical Players for Intelligent Hardware

Source: CAICT (2018)

24 6.1.2 Machine Vision Technology With the continuous integration of AI technology and the real industry, Compared with the traditional visual technology, AI empowers the image recognition capability of computer vision algorithms is machine with the vision technology, which makes it have a visual getting stronger and stronger, and a large number of excellent perception and cognitive mechanism similar to human classification computer vision enterprises have emerged in many countries. In recognition of image features, with a series of advantages, such as the US, a number of multinational technology enterprises such as fast speed, high precision and high accuracy. Amazon, Google, Microsoft and Facebook present the characteristics of an industry-wide layout covering the basic layer, the technical layer From the point of view of technical ability, it mainly realizes the and the application layer; there are also some startups that focus on requirements of identification, classification, positioning, detection local application areas, e.g., Cape Analytics realizes smart valuation and image segmentation of object/scene within the picture or video based on aerial photography of residence, and Photor, Steam, Oculus and other functions in industrial applications, so it is widely used to Home and Viveport serve as the three major VR content distribution realize video surveillance, automatic driving, vehicle/face recognition, platforms. In China, the technical experts of some top computer vision medical imaging diagnosis, autonomous robot navigation, industrial enterprises are mostly well-educated and well experienced, and the automation systems, aviation and remote sensing measurement. related industries get accumulation for many years. For example, According to the report of MarketsandMarkets, the global market for SenseTime is currently providing AI+ shooting, AR special effects and AI-based computer vision in 2017 was USD 2.37 billion, and is expected AI identity verification services for major smartphone manufacturers; to reach USD 25.32 billion in 2023. During the forecast period (2018- DeepGlint is deeply concentrating on visual algorithm technology and 2023), the CAGR is 47.54%. According to the report of Forward Industry embedded hardware R&D technology; Yi+ is more about providing Research Institute, in 2017, the size of China’s computer vision market intelligent analysis and recommendation services for commercial was RMB 6.8 billion. It is estimated that this figure will reach RMB 78 visual content, while Cloudwalk Technology, Megvii Technology and billion in 2020, with an average CAGR of 125.5%. YITU Technology have also different layouts.

Fig. 6-2 Typical Players in the Machine Vision Field

Source: CAICT (2018)

25 6.1.3 Intelligent Speech Technology RMB 10.57 billion, an increase of 70% compared with 2016. With Intelligent speech technology is a technology that can realize the the expansion of the intelligent speech application industry and intelligent transformation of text or command and speech signals. increasing market demand, it is expected that the scale of China’s It mainly includes speech recognition and speech synthesis. Speech intelligent speech market will further increase in 2018, reaching RMB recognition is like a “machine’s auditory system” that transforms a 15.97 billion. speech signal into a corresponding text or command by recognizing and understanding. Speech synthesis is like a “machine’s phonetic At present, the application of intelligent speech technology on user system”, in which a machine can read the appropriate text or command terminals is the hottest. Many Internet enterprises have invested in and turn it into a personalized voice signal. The intelligent speech human and financial resources to carry out research and application technology has been widely used in smart speakers, voice assistants in this area, with the aim of rapidly capturing the customer base and other fields because of its functions of human-computer speech through the novel and convenient mode of voice interaction. In the interaction, speech control, voice-pattern recognition and so on. US, Apple’s Siri, Microsoft’s PC-based Cortana, Microsoft’s mobile- based Xiaobing, Google’s Google Now, and Amazon’s Echo are According to data from ASKCI Consulting, in 2017, the global smart widely-known household applications; in China, iFLYTEK, AISpeech, voice market amounted to USD 11.03 billion, a year-on-year increase Unisound, as well as the Internet giant BAT all involve themselves of 30%. In 2017, the scale of China’s intelligent speech market reached deeply in this area for the business layout.

Fig. 6-3 Typical Players in the Intelligent Voice Field

Source: CAICT (2018)

26 6.1.4 Natural Language Processing of which the NLP market was RMB 4.977 billion, accounting for 21%. Natural Language Processing encompasses a wide variety of At present, there are many related mature technology application research directions, including natural language understanding and products. For example, Amazon and Facebook of the US, and natural language generation. To be clear, the former is to make the Toutiao of China use natural language technology to implement computer “understand” the idea or intention of the natural language product reviews, social networking or news platform product review, text, while the latter is to make the computer “express” the thought community review, and news article topic classification and sentiment or intention with the natural language text. From the perspective analysis over their own shopping website, social platform or news of application, NLP includes machine translation, public opinion platform; Google, Baidu and Youdao are well-versed in and are monitoring, automatic summary, extraction of point of view, subtitle constantly upgrading their online translation services; Logbar from generation, text classification, question & answer (Q&A), text semantic Japan, iFLYTEK and Sogou from China have developed their own comparison, etc. multilingual translation machine. In terms of the basic platform, there are Kore.ai, Linguamatics, etc. in the US, as well as Baidu Cloud, According to the data of MarketsandMarkets, the global NLP market Tencent Wenzhi and LTP-Cloud in China. The application of the public is expected to grow from USD 7.63 billion in 2016 to USD 16.07 billion opinion monitoring system includes iAcuity of Xalted in the US, the in 2021, a CAGR of 16.1%. According to China’s AI Development Report Wom-Monitor of Chaowen Tianxia in China, and Benguo public 2018, the scale of China’s AI market reached RMB 23.7 billion in 2017, opinion monitoring of APEX Technologies.

Fig. 6-4 Typical Players in the Field of NLP

Source: CAICT (2018)

27 6.2 Industrial Development Application healthcare industry will reach a total of USD 25.4 billion, accounting 6.2.1 AI+ Healthcare for about one-fifth of the global AI market. China is at the tuyere of medical AI. According to data from Forward Industry Research AI technology empowers the healthcare field, which significantly Institute, the scale of China’s medical AI market exceeded RMB 13 improves the efficiency of medical institutions and personnel, billion in 2017, and it is expected to reach RMB 20 billion in 2018. The significantly reduces medical costs, and enables people to achieve space for medical AI is broad. scientific and effective daily health check-up and prevention and better manage their own health. Currently, in medical research and pharmaceutical R&D, BergHealth and Numerate use data to drive drug discovery, vion and HBI According to data from the ICT monitoring platform of CAICT, in recent Solution provide patient disease prediction and risk analysis for years, AI+ healthcare ranks among the hottest areas of AI+ vertical medical institutions; in the field of intelligent diagnosis, IBM Watson application. From the perspective of application, intelligent healthcare focuses deeply on the field of cancer, and through acquisition and mainly includes medical research, pharmaceutical research and cooperation, continues to accumulate medical data resources and development, intelligent diagnosis and treatment, and family health expand capabilities in various fields. Alibaba’s “Doctor You” series, management. From the perspective of technical subdivision, it mainly Tencent’s Miying, YITU Technology’s “care.ai™” and PereDoc’s intelligent includes the use of machine learning technology to achieve drug image-assisted diagnosis and treatment platform realizes medical performance prediction, crystal form prediction, gene sequencing image-assisted diagnosis and treatment, and Fourier Intelligence’s prediction, etc.; use of intelligent speech and NLP technology Fourier X1 gives birth to China’s first exoskeleton robot. In the field of to achieve electronic health records, intelligent inquiry, guiding family health management, WellTok pays more attention to personal consultation, etc.; use of machine vision technology to realize medical health management and lifestyle improvement, AiCure is committed to image recognition, lesion identification, skin disease self-test, etc. helping users to use drugs on time, and iCarbonX address itself to the According to McKinsey’s forecast, by 2025, the global intelligent building of a digital life health management platform.

Fig. 6-5 Typical Players in the Smart Healthcare Field

Source: CAICT (2018)

28 6.2.2 AI+Finance is expected to grow from USD 1.338 billion in 2017 to USD 7.306 billion in AI technology empowers the financial field. From an application 2022, a CAGR of 40.4%. The White Paper on the Development of a New perspective, it mainly includes robo-advisor, credit/risk control, Generation of Artificial Intelligence (2017) predicts that China’s intelligent financial search engine, insurance, identity verification and intelligent financial industry will reach USD 800 million by 2020. customer service. Finance is one of the most data-dependent industries. AI technology is integrated with the financial industry At present, robo-advisor enterprises are mainly transformed from to drive intelligent upgrade of financial technology through the AI licensed securities, fund or asset management operators, e.g., technology based on big data. In the foreground, it can be used to Wealthfront and Betterment of the US, as well as Koudai Fortune and provide users with more comfortable, convenient and safer services; JD.com robo-advisor of China. In the financial intelligent customer in the middleground, it can provide decision-making support function service field, Digital Genius, Qiyukf, sobot and other enterprises focus for transactions, credits and analysis in financial services; in the on improving the user experience; in the field of credit/risk control, background, it can improve the financial system for various risks. In the most of the players rely on government information, enterprise background, it can improve the ability of the financial system to identify, information or personal information to form industry barriers based warn, prevent and control all kinds of risks. All in all, AI technology will on big data intelligence analysis, such as Zest Finance and Affirm, deeply reconstruct the current financial industry’s ecological structure, both being the US financial technology enterprises with consumer making financial services (banking, insurance, wealth management, finance and mobile payment data, and Qixin, which has enterprise lending, investment, etc.) more humane and intelligent. multi-dimensional real-time dynamic and full-scale business data credit platform; other enterprise applications, such as Rong According to the PwC 2017 Global Digital IQ Survey, the information 360, Data.GOV and DBpedia focus on financial search engines; utilization rate in the global financial service sector is only 26%, which SenseTime, Cloudwalk Technology, YITU Technology and Megvii is a low level compared with other industries. According to the report of Technology have entered the identity authentication market by MarketsandMarkets, the global market size of AI in financial technology relying on their leading face recognition core technology.

Fig. 6-6 Typical Players in the Global Smart Finance Field

Source: CAICT (2018)

29 6.2.3 AI+Retail Nowadays, the global intelligent retail industry participants are mainly AI technology empowers the retail industry. Intelligent retail drives e-commerce giants and startups. As for the business application the new market retail business pattern with big data and intelligent scenarios, it is still mainly on the sales side. For example, in the technologies, optimizing resource allocation and efficiency across US, for the unmanned retail store, there are Standard Cognition, the entire industry chain from production to distribution to sales, the unmanned convenience store, and Amazon’s Amazon Go. In thus enabling intelligent upgrades of industry services and China, there are Alibaba’s Taocafe, as well as JD.com’s X unmanned performance. Its commercial applications include intelligent marketing supermarkets; there are also related products of famous startups recommendation, intelligent payment system, intelligent customer such as DeepBlue Technology, F5 Future Store, and Bingo Box. In service, unmanned storehouse/unmanned vehicle, unmanned store, the aspect of customer service robots, China has Cheetah Mobile’s intelligent distribution and more. Fanbot retail robots, Keenon Robotics’s Peanut guiding robots, and SIASUN’s Sunbot-I sales promotion and guiding robot, which have The MarketandMarkets report shows that the global intelligent retail been applied in various application scenarios. In the smart retail market is expected to grow from USD 13.07 billion in 2018 to USD supply chain scenario, UPS in the US tested UAV distribution in Florida; 38.51 billion in 2023, with a CAGR of 24.12% over the forecast period Wal-Mart’s Pickup Tower is being deployed across the US. In China, (2018-2023). In China, according to the National Bureau of Statistics, Meituan has launched an unmanned distribution open platform; at the end of 2017, the total retail sales of consumer goods in China JD.com is building an all-environment intelligent retail logistics with amounted to RMB36.6262 trillion, an increase of 10.2%. Roland Berger unmanned distribution stations, unmanned warehouse “Asian One” predicts that by 2030, AI technology will bring about RMB 420 billion in and large cargo drone “Jinghong”. cost reduction and gain increase to the Chinese retail industry.

Fig. 6-7 Typical Players in the Smart Retail Sector

Source: CAICT (2018)

30 6.2.4 AI+ Education CAGR during the forecast period (2018-2023) is 47.0%. Global market AI technology empowers education. It pays attention to the Insights also released a new research report that predicts that the individualized education of students, helps teachers to teach students value of the AI industry in the education market will exceed USD 6 according to their aptitude, enhances the quality of teaching and billion in 2024. Among them, the Asia-Pacific intelligent education learning, and promotes equalization and affordability of education. market, including China, will have a CAGR of more than 51%, making At present, the intelligent education comprehensively covers the it the most profitable region. industrial chain of “teaching, learning, examination, evaluation and management”, and has accelerated in the subdivision market, such At present, in the field of AI adaptive learning, whether it is Knewton as preschool education, K12, higher education, vocational education in the US, Century Tech in the UK, Smart Sparrow in Australia, and online education. From the perspective of application, intelligent and Yixue Education and TAL Education of China, the AI education education can be divided into four parts: learning management, platform is used to help students quickly grasp knowledge points learning evaluation, teaching guidance, and teaching cognitive and improve their learning result. In the field of intelligent evaluation, thinking. From the perspective of subdivision, it includes educational China’s Xuebajun and iFLYTEK and other enterprises have launched evaluation, photo-taking & answering, intelligent teaching, intelligent intelligent scoring system; in the field of teaching guidance, China’s education, intelligent scoring, AI adaptive learning and other Singsound and Liulishuo and other enterprises have introduced application scenarios. language tutoring system, Tabtor, Carnegie Learning, Front Row and other enterprises of the US have launched the intelligent tutor According to the report of MarketsandMarkets, in 2017, the global system to simulate one-on-one tutoring, which forms the effect of AI technology market size in the education industry was USD 373.1 close expert guidance. LightSail and Newsela of the US have made million, and is expected to reach USD 3,683.5 million by 2023. The personalized and intelligent recommendations for students’ reading, and developed students’ reading ability and interest.

Fig. 6-8 Typical Players in the Intelligent Education Field

Source: CAICT (2018)

31 6.2.5 AI+ Home In recent years, smart home has shown strong vitality on a global AI technology empowers the home, helping the home ecosystem scale. As the largest market for smart home, the US pays attention to develop from perception to cognition, making home life safer, to smart home with smart speakers as the central control, including more comfortable, more energy efficient, more efficient and more Amazon Echo, Google Home and other hot sale products. In China, convenient. In the future, smart home will gradually implement on the one hand, major players in the market have released various adaptive learning and control functions to meet the individual needs types of smart speaker products, such as Alibaba’s “Tmall Genie of different families. Smart home is an IoT-based home ecosystem Smart Speaker”, Xiaomi’s “Mi AI Speaker”, iFLYTEK-JD.com’s “DingDong that includes the intelligent lighting system, intelligent energy Smart Speaker”, Baidu’s “Xiaodu Smart Speaker”, Tencent’s “Tingting management system, intelligent audiovisual system, and intelligent Smart Speaker”, Rokid’s “Rokid Smart Speaker”, and Himalaya’s security system. “Xiaoya Smart Speaker”. On the other hand, major players are also actively building smart home with IoT platforms, such as Xiaomi’s According to the latest Strategy Analytics report, the global smart MioT, Huawei’s HiLink, and Haier’s U+. home market reached USD 84 billion in 2017, up 16% from 2016’s USD 72 billion, and is expected to reach USD 96 billion in 2018. According to data from ibaogao.com, the size of China’s smart home market in 2017 was RMB 91.66 billion, and it is expected to expand to RMB 139.6 billion in 2018.

Fig. 6-9 Typical Players in the Smart Home Sector

Source: CAICT (2018)

32 6.2.6 AI+Agriculture As the world’s first agricultural power, the US has always led the AI technology empowers the agricultural sector, enabling agriculture development of intelligent agriculture industry. In precision farming, to effectively cope with the effects of extreme weather, reduce enterprises such as Prospera, Arable and Trimble use cameras, resource consumption, optimize resource allocation, reduce costs, sensors, micrometeorological data or positioning technology to and optimize time and resource allocation for maximum yield and monitor and analyze crops; in the field of agricultural intelligent efficiency. From the perspective of application, intelligent agriculture devices, Blue River has created “LettuceBot” robots, “see-and-spray” mainly includes agricultural robots, precision farming, drone system and the unmanned aerial system. At the same time, the analytics, and livestock monitoring. intelligence of China’s agricultural industry is also accelerating its transformation. For example, McFly’s intelligent agricultural According to the report of MarketsandMarkets, the global intelligent monitoring drone, UniStrong’s “Huinong” Beidou navigation agriculture market amounted to USD 6.7 billion in 2017. It is expected agricultural automatic driving system, and GAGO’s large-scale to reach USD 7.53 billion in 2018 and USD 13.5 billion in 2023. The application of AI technology in farming and pig raising are typical CAGR in the forecast period (2018-2023) is 12.39%. Among them, the cases. At the same time, some technology giants have also begun value of AI technology in the agricultural market in 2016 was USD to make deployment in the intelligent agriculture sector. In April 432.2 million, and the value is expected to be USD 2.6285 billion by 2018, JD.com’s “JingDong Farm” made its debut; in June, Alibaba 2025. The CAGR during the forecast period (2017-2025) is 22.5%. Cloud’s ET agricultural brain came out.

Fig. 6-10 Typical Players in the Intelligent Agriculture Sector

Source: CAICT (2018)

33 6.2.7 AI+ Manufacturing In recent years, intelligent manufacturing has become the main AI technology empowers the manufacturing sector to significantly battlefield for industrial upgrading of the countries, and some optimize the manufacturing cycle and efficiency, improve product developed countries have gone far in this respect. For example, the quality, and reduce labor cost. The intelligent manufacturing industry intelligent workshop in Phoenix, Germany, the AWS cloud ecosystem chain has a wide range of scenarios. Its typical application scenarios of C3 IoT in the US, and the intelligent manufacturing unit of Harley- include intelligent product and equipment; intelligent factory, Davidson in the US. In the field of smart device monitoring, there workshop and production line; intelligent management and service; are also enterprises such as KONUX in Germany, Scortex in France, intelligent supply chain and physics; intelligent software development and BrainsTechnology in Japan. Some Chinese enterprises are also and integration; intelligent monitoring and decision-making, etc. building smart factories and increasing the intensity of enterprise Data from Market Research shows that the global intelligent transformation and upgrading. For example, Eston’s industrial robot manufacturing market reached USD 202.82 billion in 2017 and is intelligent factory established in Nanjing, Trumpchi’s smart factory expected to reach approximately USD 479.01 billion in 2023. The in Hangzhou, and the digital factory of CRRC Nanjing Puzhen. In the CAGR during the forecast period (2018-2023) is approximately 15.4%. traditional household appliance manufacturing industry, Midea, According to data from Forward Industry Research Institute, in 2017, Haier, Gree and other enterprises are actively shifting toward the the output value of China’s intelligent manufacturing industry has intelligent manufacturing model. reached about RMB 1.5 trillion. It is expected that in the next few years, China’s intelligent manufacturing industry will maintain an average CAGR of around 11%. By 2023, the industry market will reach RMB 2.81 trillion, and the industry has huge room for growth.

Fig. 6-11 Typical Players in the Intelligent Manufacturing Sector

Source: CAICT (2018)

34 6.2.8 AI+ Cybersecurity estimated global AI-based cybersecurity market will grow at a CAGR AI technology empowers the entire network, helping manufacturers, of over 29% during 2018-2022. enterprises, and individuals to effectively address the growing number of cybersecurity issues such as cyber fraud and malicious Governments attach great importance to cyber security. For example, attacks. AI has some unique advantages in the field of network IBM and CrowdStrike of the US, and RepKnight of the UK provide defense, which make it a breakthrough for AI cybersecurity protection. cybersecurity protection service such as blocking malware and The main applications of intelligent cybersecurity include: network detecting phishing and data leakage by providing a cybersecurity monitoring and prevention (including real-time identification, response protection platform. Fraud and risk detection are another important and defense against network attacks, and security vulnerabilities application in the field of cyber security. Enterprises such as DataVisor and system failure prediction, cloud security, etc.); prevention of the and Drawbridge from the US, and Feedzai from Portugal have execution of malware and files; improvement of operational efficiency their related business layouts. In recent years, China’s intelligent of security operation center; network traffic anomaly detection; cybersecurity enterprises have seen rapid growth. Among them, application security detection; network risk assessment. TuringSec, 360, NetEase Cloud and other enterprises provide network defense and anti-virus service, while Eagle Eye Tech, Maxent, Ahi According to Technavio’s latest market research report, the global Fintech, Tongdun Technology and other enterprises provide data intelligent cybersecurity market was USD 4.96 billion in 2017, and the monitoring and anti-fraud service.

Fig. 6-12 Typical Players in the Field of Intelligent Cybersecurity

Source: CAICT (2018)

35 6.2.9 AI+ HR The international HR recruitment sector attaches great importance to AI technology empowers the HR field, which helps HR to automate the the application of AI. The international HR recruitment sector attaches process of management process, greatly improve work efficiency and great importance to the application of AI, with the US playing a compliance, as well as reduce staff recruitment and management leading role. IBM’s Watson contains three frameworks of HR, which, costs and personal bias. The main application contents include pre- through the deep learning of models, finds professionals, seek insight recruitment talent channel maintenance, talent forecast analysis, job into the needs of competitors, and evaluate job experience and skill matching, resume screening, AI chat support, etc.; appointment for perception. interview, interview result inquiry, handling employment procedures, etc. during the recruitment process; new employee training, Q&A Ultimate Software offers a range of HR solutions from recruitment to interaction, knowledge learning and career planning support; retirement. Japan’s HR giant Recruit is known for its talent appraisal employee behavior and efficiency analysis, salary analysis, mental technology. Its business model consists of two parts: Career Adviser health analysis, team culture analysis, etc. after the recruitment. (CA), which serves for personal consultation for job seekers, and According to Grand View Research, the global HR management Recruit Adviser (RA), which serves for the recruiting enterprise. market size was valued at USD 12.6 billion in 2016. It is forecast to Many Chinese enterprises have applied AI technology to the daily reach USD 30 billion by 2025, and the CAGR during the forecast HR work. For example, Emotibot’s intelligent HR assistant robot, period (2017-2025) is 10.4%. According to ASKCI Consulting, the scale Beichoo Technology’s HR super assistant “Xiaobei”, and the “MOBOT” of China’s HR service market increased from approximately RMB 158.4 developed by Mozi AI can share most of the work of recruiters. billion in 2013 to RMB 343.6 billion in 2017, with a CAGR of 21.9% and is Ifchange uses AI to provide scientific judgment and decision-making expected to increase to RMB 842.7 billion in 2022. basis, and helps enterprises to fully increase their human capital.

Fig. 6-13 Typical Players in the Field of Intelligent HR

Source: CAICT (2018)

36 6.2.10 AI+ Public Safety in 2017, but the output value of all AI security products is less than RMB AI technology empowers the security field, filling the gap that 2 billion, indicating a less than 1% penetration rate of AI in the security traditional security is unable to meet the industry’s requirements industry. for the accuracy, breadth and efficiency of security systems. Intelligent security is the area that gets earliest large-scale application of AI and The establishment of intelligent security systems is inseparable from continues to generate commercial value. Its functions put into application the integration of software algorithm and hardware system. In terms are mainly in object tracking, detection and abnormal behavior analysis, of system hardware, in the world, there are Axis Communications video quality diagnosis and summary analysis, face recognition and AB, the video processing chip vendor in Sweden, as well as ADT and feature extraction analysis, vehicle identification and feature extraction OPTEX, the security hardware providers in the US. In China, Hikvision, analysis, etc. Dahua, NetPosa and other enterprises are the market leaders in the relevant fields. In terms of software algorithms, Israel’s Agent According to analysis of Mordor Intelligence, the market size of Video Intelligence, Canada’s Genetec, the US’ Google, Facebook global video surveillance system was USD 34.9623 billion in 2017, and Microsoft, and China’s SenseTime, Megvii Technology, YITU and is expected to reach USD 82.6153 billion in 2023. The CAGR in the Technology, Cloudwalk Technology and other enterprises have forecast period (2018-2023) is 15.41%. According to CAIJING.COM.CN, excellent image analysis algorithms. the output value of China’s security industry reached RMB 450 billion

Fig. 6-14 Typical Players in the Field of Intelligent Public Safety

Source: CAICT (2018)

37 6.2.11 Intelligent Driving 68.1 billion in 2017. It is expected to reach RMB 89.3 billion in 2018, with AI technology empowers traditional driving, which can effectively an annual growth rate of 31.1%. improve production and transportation efficiency, alleviate labor shortage, achieve safety, environmental protection and high Globally, autonomous driving mainly includes mainframe vendors efficiency, and lead the overall upgrading and reshaping of industrial (BMW, GM, Audi, etc.), suppliers (Autoliv’s Veoneer, Bosch, etc.), ecology and business models. Intelligent driving is a complex industry technology enterprises (Google’s Waymo, Baidu’s Apollo, Israel’s chain that covers chips, software algorithms, high-definition maps, Mobileye, etc.) and travel enterprises (Tusimple, Pony.ai, JingChi. security controls, and more. ai, etc.). Mobileye, BMW, Volvo and Ford have all announced that they will achieve commercial application of at least L4-class self- By the end of 2017, China’s per capita car ownership was about 0.156, driving cars in 2021. However, due to factors such as maturity of less than 1/5 of the US, and the market has huge room for growth. relevant technologies, improvement of laws and regulations, and The policies of countries in the world and the industrial layout of infrastructure support, there are still many uncertainties, and some major enterprises have made intelligent driving a hot research field. relatively closed driving scenarios with low complexity and few At present, the auto-driving cars are generally divided into L0 to L5, L4 external interference factors are considered to be the most hopeful and L5 can be collectively referred to as “Unmanned Driving”, and the scenario in which the driverless car application will be first achieved. current autonomous driving technology is basically at the L2-L3 level. At present, the Tesla model series (L2-L3 car) and Audi a8 ( L3 car ) The US IHS Automotive report predicts that by 2025, the global have been mass-produced. In March 2018, autowise.ai announced autonomous vehicle sales number will be close to 600,000, and in the trial operation of the world’s first self-driving clean-up team in 2035 it will reach 21 million. The market will maintain a CAGR of 48% Shanghai. In July,Baidu announced mass-production of the world’s during the forecast period (2025-2035). According to the report of first L4 self-driving bus “Apolong” manufactured in cooperation with Sootoo, the size of China’s intelligent driving market had reached RMB Jinlong Bus.

Fig. 6-15 Typical Players in the Field of Intelligent Driving

Source: CAICT (2018)

38 6.2.12 Intelligent Robotics integration market exceeded RMB 120 billion, a year-on-year increase AI technology empowers robots, enabling robots to have human-like of 25.4%. Among them, the industrial, service and special robot perception, coordination, decision-making and feedback capabilities. market has reached RMB 43.5 billion, and is expected to reach From the perspective of application, it mainly includes intelligent RMB 71.9 billion by 2020, with a CAGR of 18.24% over the forecast industrial robots, intelligent service robots and intelligent specialized period (2017-2020). The four international giants in the robotics robots. The current mainstream intelligent industrial robots generally field (ABB, Fanuc, Yaskawa and Kuka) have already occupied an have functions of packaging, positioning, sorting, assembly and absolute market share in the intelligent industrial robot industry. As detection; intelligent service robots generally have functions such an emerging industry, the intelligent service robot industry is more as family companion, business service, healthcare, retail sales, diversified in market demand as it directly faces the consumers, and and rehabilitation of disabled people; intelligent specialized robots the market competition is more regional. Home cleaning robots such generally have functions of reconnaissance, search & rescue, fire as iRobot, LEGO’s educational programming robots, CYBERDYNE’s fighting, decontamination, and demolition. medical assistant robots, Ninebot’s walking robots, and Roobo’s commercial service robots are used in different service areas such Data released by IFR show that the global robot market had reached as home, education, healthcare, travel, and business. In the field of USD 50 billion in 2017. In China, according to data from Analysis intelligent special robots, there are reconnaissance robots such as Report on Robot Industry in China 2018, robot ontology and system ReconRobotics and fire-fighting robots of CITIC HIC.

Fig. 6-16 Typical Players in the Field of Intelligent Robotic

Source: CAICT (2018)

39 Research from Gartner Hype Cycle for Artificial Intelligence, 2018

AI is almost a definition of hype. Yet, it is still To find short-term wins and hone the long- encourage developers to experiment early: New ideas will surface and some term vision for AI, CIOs, IT leaders and AI with AI developer toolkits and AI PaaS, as current ideas will not live up to expectations. champions should track major AI trends: well as plan developers’ upskilling to get This Hype Cycle will help CIOs and IT leaders this contingent ready for its new role in AI trace essential trends and innovations to • Conversational AI is on many corporate strategies. determine scope, state, value and risk in agendas, spurred by the worldwide their AI plans. success of Amazon Alexa, Google New Entrants Assistant and others represented by • AI Governance: Concerns about validity, Analysis VPA-enabled wireless speakers at the explainability and unintended bias of AI What You Need to Know pinnacle. To develop chatbot and voice- came to the foreground this year. Many enabling strategies, implementers should Now is the deciding time for the future of Gartner clients want to understand what pay attention to the time before plateau of AI. Only 4% of CIOs worldwide report they it takes to govern AI even before they start virtual assistants, chatbots, NLG, NLP and have AI projects in production. Every decision AI initiatives. speech recognition, among others. about AI influences AI’s long-term direction. AI gives hope and fulfills sci-fi fantasies; • AI Developer Toolkits: The need for AI is • Machine learning, and related DNNs, it is both utopian and dystopian. Anxiety massive, and data scientists are a small ensemble learning, and predictive and about implementing AI is increasing. The group compared to software developers. prescriptive analytics, are becoming term “artificial intelligence” is on gartner. Vendors make a concerted effort to a common capability, with new tools com’s top 10 lists for emerging searches, enable this very large group to perform and approaches pouring into the high-growth searches and most popular basic AI development functions via market. DNNs remain a focal point for searches. Data and analytics leaders across familiar concepts presented in developer implementers and scientists, but the key many industries are seeking a breakthrough, toolkits. is to find the right problems for deep which they should target in the long run. learning to solve. Meanwhile, in his However, the immediate impact of AI is • Knowledge Graphs: The rising role Test-of-Time award talk, a renowned within practical applications. of content and context for delivering AI researcher, Ali Rahimi, asked, “Has insights with AI technologies, as well as machine learning become alchemy?” The Hype Cycle recent knowledge graph offerings for in his call for greater rigor within the AI applications have pulled knowledge AI is overhyped as a socioeconomic machine learning community. graphs to the surface. phenomenon. The media, governments, corporations and individuals each have an • Compute infrastructure drives AI • AI PaaS: The AI PaaS hype is heating up, opinion about AI, mostly based on vague progress and is being tailored for AI. with the leading cloud service providers’ ideas of what it really is. This Hype Cycle It will frontier AI advancement. GPU competition using AI PaaS as a lure views AI as a pervasive paradigm and an accelerators, FPGA accelerators, deep to their clouds and as a tool to attract umbrella term for many innovations at the neural network ASICs and neuromorphic developers and data scientists. different stages of value creation. The traffic hardware showcase different compute jam at the Peak of Inflated Expectations is ideas, and more approaches are looming • Chatbots: Chatbots have increased in increasing, as early implementers grow in in the future. CIOs and IT leaders should hype and are up for major growth over numbers, but production implementations balance cost and performance for use- the next years, but also they are set up for remain scarce. A long line of high-promise case-driven capabilities in their compute a backlash once they reach the Trough of innovation profiles at the Innovation Trigger infrastructure strategies. Disillusionment. phase are approaching the traffic jam at the Peak of Inflated Expectations, indicating that • AI is coming to the masses of application • VPA-Enabled Wireless Speakers: AI hype the AI hype will continue. None of the profiles developers and software engineers, would be incomplete without Amazon in this Hype Cycle is Obsolete Before Plateau, most of whom don’t even suspect yet that Alexa, Google Assistant and the likes. but not all will not survive, and many will they will be the main AI implementation Although, these are just speakers, they morph into something different — this will force in two to five years. Although it sound and some even look intelligent. depend on the choices and decisions that is early, CIOs and IT leaders should customers are making today.

40 41

• Intelligent Applications: These • RPA Software: RPA has accelerated to Name Changes applications signify a trend of embedding the top 10 most popular searches on • Deep Neural Nets (Deep Learning), AI in enterprise applications, as well as Gartner.com. Hype Cycle readers should formerly Deep Learning: The second encapsulating AI in domain applications. be familiar with RPA capabilities and name change in two years reflects understand that AI is a small part of them. the vibrant innovation, disruption, and debate around these algorithms and frameworks.

FIGURE 1 Hype Cycle for Artificial Intelligence, 2018

Source: Gartner (July 2018)

41 The Priority Matrix business efficiencies and require ongoing The key to AI success is “narrow AI” — that is, Except for just three innovations, everything education. Transformational innovations not artificial general intelligence, but narrow else on this Hype Cycle is high-impact or are game changers — they call for new use cases with defined benefits. The priorities transformational. More than a half of the skills and present high risk and reward. AI in Figure 2 should help CIOs and IT leaders Hype Cycle entries could reach the Plateau techniques themselves, when applied to risk to identify profiles relevant to their plans and of Productivity (20% market penetration) in taking, make risks literally “calculated,” as to stage their commitments appropriately. less than five years. These are indeed big well as more accurate and more frequent, so Innovations at the peak are already useful if expectations! High-benefit innovations bring every step becomes smaller and failures are approached without inflated expectations. not so dramatic. For practical efficiency, start with the profiles approaching the Plateau of Productivity. FIGURE 2 Priority Matrix for Artificial Intelligence, 2018

Source: Gartner (July 2018)

42 43

For competitive advantage, start with the Other changes to this year’s Hype Cycle are: by AI against the need for appropriate profiles on the Innovation Trigger that will go oversight, risk management and investment through the Hype Cycle fast, such as chatbots, • Bots: Chatbots are included instead of management. As yet, there aren’t many natural-language generation, AI-related C&SI bots to more accurately represent the answers, but enterprise practitioners are services and AI developer toolkits. hype. already making steps toward establishing AI governance. New organizations — like It is still early for AI. Many innovations in • Algorithm marketplaces: The hype did the AI Now Institute and Partnership on AI this Hype Cycle are not yet fully understood. not live up to the expectations. This profile — are being formed to prevent AI-related For example, DNNs (which started the was absorbed into API marketplaces, an biases, discrimination and other negative current AI hype) are opaque, computer innovation profile in the “Hype Cycle for implications. vision algorithms are overfitted, and Data Science and Machine Learning.” autonomous vehicles cause debates about User Advice: AI governance is set on responsibilities in case something goes • Learning BPO: While AI requires three cornerstones — trust, transparency wrong. That is why AI governance, digital continuous learning, and it will also take and diversity. At its core, it builds upon the ethics and human in the loop draw the away routine in favor of creative tasks that principles of data and analytics governance, attention of more AI-advanced organizations require constant education, learning BPO but the fundamental difference of AI and should be on the radar of AI strategists. will be one of many upskilling and training governance is in the probabilistic nature of approaches. This innovation profile AI, and in how AI is used to drive advanced Off the Hype Cycle remains on the “Hype Cycle for Business forms of prediction. AI-based systems (often This year’s Hype Cycle does not include Process Services and Outsourcing.” using machine learning) have emergent lower-level cases that are represented by designs, while classical information higher-level concepts. These are: On the Rise systems have deliberate designs. Data and AI Governance analytics governance focuses on trust, but analytics governance puts an emphasis on • Virtual customer assistants (part of the Analysis By: Svetlana Sicular; Frank transparency. AI governance extends these virtual assistants innovation profile) Buytendijk concepts to trust, transparency and diversity of data, algorithms and the people in the • Cognitive expert advisors (part of the AI governance is the process of Definition: AI teams. AI governance favors diversity to cognitive computing innovation profile) creating policies, assigning decision rights counteract bias and predictive errors. and assuring organizational accountability • Level 3 and level 4 vehicle autonomy for risks and investment decisions for the Data and analytics leaders should: (part of the autonomous vehicles application and use of artificial intelligence innovation profile) in the context of predictive models and • Ensure trust in the current data sources to algorithms. AI governance is part of adaptive avoid one-sided information. • Deep reinforcement learning (part of the data and analytics governance. It addresses deep neural networks [deep learning] the perceptive, predictive and probabilistic • Demand new, different and even innovation profile) nature of AI. contradictory data to combine with what you already use to minimize risks of AI • Intelligent apps (part of the intelligent Position and Adoption Speed Justification: biases. applications innovation profile) Until recently, AI has been mostly in the domain of scientists and researchers, with • Identify transparency requirements for • Artificial intelligence for IT operations little focus on the practical implications of AI data sources and algorithms. (AIOps) platforms (part the of augmented upon the wider enterprise. With AI having analytics innovation profile that is now reached the perimeter of practical • Promote transparency of AI solutions and included in “Hype Cycle for Analytics and enterprise application, data and analytics communication around AI to minimize Business Intelligence” and “Hype Cycle for leaders are beginning to raise the question of different interpretations of AI results. Data Science and Machine Learning”) how AI governance will be conducted, before they start implementing AI. They are asking • Diversify algorithms to meet the how to balance the business value promised complexity of the problems that AI is solving.

43 • Challenge the expected outcomes. If the Maturity: Embryonic User Advice: Focus on business results outcomes are fully expected, the problem enabled by applications that exploit special- you are solving is too simple. Artificial General Intelligence purpose AI technologies, both leading-edge Analysis By: Tom Austin and older. • Create feedback loop, “guard rails” with “circuit breakers” and human oversight to Definition: Artificial general intelligence Leading-edge AI is enabling what are prevent AI mistakes. (AGI) — also known as “strong AI” and currently considered “amazing innovations,” “general-purpose machine intelligence” — including deep-learning tools and related • Maximize the benefit from AI by would handle a very broad range of use natural-language processing capabilities. establishing organizational roles and cases, if it existed. It does not, though it is a These innovations are doing what we responsibilities, starting with a center popular subject of science fiction. Current previously thought technology could not do. of excellence that allows to share skills, AI technologies do not deliver AGI. Despite They are, however, typically research tools resources and knowledge. appearing to have human-like powers of that are only just emerging from research learning, reasoning and adapting, they labs, undergoing turbulent changes in • Define governance process to evaluate lack commonsense, intelligence, and direction, and not fully understood in terms business returns to either pivot to another extensive means of self-maintenance and of engineering principles. Over time, we will AI project or iterate. Vary governance reproduction. Special-purpose AI — “weak learn their limitations and develop workable approaches: In some cases, no or little AI” — does exist, but only for specific, narrow engineering guidelines. As the amazement governance accelerate initial innovation, use cases. wears off and ennui sets in, we will treat but this should be a conscious decision them as “aging innovations.” within your AI governance framework. Position and Adoption Speed Justification: Tangible progress on AI has been limited Look for business results enabled by Business Impact: AI reflects what it learns: to weak AI. AGI’s position and adoption applications that exploit aging innovations It can be cool, or it can be creepy (or both, speed on this Hype Cycle therefore remain (including expert systems and other symbolic unintentionally). Typically, an AI team unchanged. (We changed this entry’s AI approaches, as well as simpler forms of decides what to learn, how to do it and how name from “general-purpose machine machine learning), amazing innovations to ensure the best outcomes. The goal of intelligence” in 2017 to reflect the popularity (typically more powerful but less understood governance is to ensure that such decisions of the term “AGI.”) technologies), or both. Examples of such maximize value and minimize risk. Some applications include autonomous means of “black box” models are acceptable if you Today’s AI technology cannot be proven to transportation, smart advisors and virtual can prove the validity of the outputs, while possess the equivalent of human intelligence assistants focused on various goals (such some other models require transparency (the lack of agreement about a test to prove as improved wealth management) and in order to meet regulations or preserve such intelligence is itself a problem). It responsibilities (such as sales or budget organizational reputation (for example, may, at some point, be possible to build a management). Most use both amazing and when it comes to decisions like credit, machine that approximates human cognitive aging innovations. employment or housing). Data sources for AI capabilities, but we are likely decades away often contain incomplete or unintentionally from completing the necessary research and Special-purpose AI will have a huge and biased information — this is the main engineering. disruptive impact on business and personal cause of erroneous AI outcomes. The life. End-user organizations should ignore awareness of AI risks is currently limited The subject of AGI often arises in discussions AGI, however, until researchers and to those publicized by the media, such as of “cognitive computing” — a term that advocates demonstrate significant progress. fake news and self-driving cars, while every means different things to different people. Until then, ignore any suppliers’ claims that industry may encounter its own AI problems. For some it denotes a set of AI capabilities, their offerings have AGI or artificial human Unfortunately, 95% currently neglect AI for others a specialized type of hardware intelligence — these are generally illusions governance, unaware of the potential risks. (as in neuromorphic or other highly parallel, created by programmers. short propagation path processors). It can Benefit Rating:High also describe the use of information and Business Impact: AGI is unlikely to emerge communication technology to enhance in the next 10 years, although research will Market Penetration: Less than 1% of target human cognition, which is how Gartner uses continue. When it does finally appear, it will audience the term. probably be the result of a combination of

44 45

many special-purpose AI technologies. • Device development kits (e.g., AWS • Abstract adopted vendor offerings Its benefits are likely to be enormous. But DeepLens and Microsoft Vision AI) where possible to minimize portability some of the economic, social and political constraints and lock-in. implications will be disruptive — and • AI serving SDKs (e.g., Apple’s CoreML and probably not all positive. Google ML Kit). • Avoid directing disproportional investments or effort in migrating There are currently no vendors of systems Across all categories, vendor offerings established applications to a new that exhibit AGI, but many companies are require distinct deployment considerations platform for a small set of differentiating engaged in basic research. Examples are and have varied feature coverage features. DeepMind (owned by Google), OpenAI and differences, but we expect greater Vicarious. consistency in the future. • Ensure deployed capabilities are aligned to direct end-user benefits that cannot be Benefit Rating:Transformational Cloud-based AIaaS platforms reduce data easily achieved without AI. science complexities for more developer- Market Penetration: Less than 1% of target friendly adoption, as compared with native • Understand that such offerings do not audience PaaS platforms. Model life cycle support enable software engineers to replace an varies widely by vendor across data experienced data scientist. Maturity: Embryonic preparation, feature engineering, model selection and training, hyperparameter • Leverage established information AI Developer Toolkits tuning and model deployment phases. management best practices for data Analysis By: Eric Hunter, Svetlana Sicular management and privacy. AI developer toolkits support a limited set of Definition: Artificial intelligence (AI) native use cases, such as image recognition • Adopt offerings in alignment with developer toolkits are applications and (including faces and landmarks), text larger organizational cloud and mobile software development kits (SDKs) that analytics and image labeling. Developers development standards and strategies. abstract data science platforms, frameworks can also deploy custom-built models and and analytic libraries to enable software optionally update those models from cloud Vendor offerings are being released at a engineers to deliver AI-enabled applications. services at model runtime. Although Core rapid pace in the market with a desire to They cover four maturing categories: cloud- ML and ML Kit have unique model formats, attract new development communities. based AI as a service (AIaaS); toolkits for numerous conversion utilities continue to be Early adopters should carefully evaluate virtual assistants (e.g., Apple Siri, Amazon released for models from numerous formats, and stress test employed offerings, along Alexa and Google Assistant), device including ONNX and MXNet. Commercial with fully understanding the going concern development kits; and AI serving SDKs. vendors have also introduced services (such support for each specific function. Software engineers use them to incorporate as IBM’s Watson Services for CoreML) to AI into new or existing applications. extend AI serving SDK support. Business Impact: CIOs, application development leaders, and data and Position and Adoption Speed Justification: Device development kits position custom analytics leaders should prepare for Vendors have worked aggressively to hardware devices (such as cameras) software developers to become a key deliver developer-oriented AI toolkits and with developer-friendly APIs and SDKs to contingent for AI development and SDKs during the past 18 to 24 months. encourage platform developer adoption. As implementation. The demand for AI is Representative offerings include: platform support is incorporated into broader significant and is increasing at a rate beyond market offerings, direct platform vendor kit which experienced data scientists can • Cloud-based AIaaS platforms (e.g., offerings will diminish. meet alone. Gartner notes that 60% of data Google AutoML, AWS SageMaker and science talent is concentrated in 50 cities Azure ML Studio) User Advice: Application development worldwide and, in those cities, there is a leaders must evaluate AI developer toolkits finite set of employers. • Toolkits for Virtual Assistants (e.g., and balance their present-day benefits Amazon Alexa Skills Kit, Apple SiriKit, and capabilities. IT leaders adopting these Adoption of developer toolkits will continue Baidu DuerOS Open Platform, Google offerings to incorporate AI capabilities and to increase. As these offerings continue to DialogFlow and Cortana Devices SDK) features into applications should: mature, Gartner expects offerings to:

45 • Expand support for edge and device- Position and Adoption Speed Justification: knowledge graph concepts to determine centric AI models through lightweight The rising role of content and context in how vendor solutions could benefit their runtime frameworks delivering insights through the use of AI digital business platform. technologies has pulled knowledge graphs • Mature into distinct categories in future to prominence Google’s Knowledge Graph For example, Microsoft and Google embed Hype Cycles and Microsoft Graph are examples of the knowledge graphs in their cloud office knowledge graph’s growing popularity environments — Office 365 and G Suite • Increase support for higher-level, focused due to its promise to enrich your data with respectively. By capturing signals from the AI use cases across specific business missing data. Specialist vendors are offering usage of these environments, their graphs verticals and consumer demands graph-based products to new markets and are able to ingest data about the use of well-known vendors are accommodating the applications, enabling working relationships • Continue to reduce adoption barriers technology in their platforms and products. between employees, as well as thematic in the deployment of AI capabilities for connections between digital assets to be software engineers and citizen data Knowledge graphs are ideally suited to gathered. This supports collaboration and scientists storing data extracted from the analysis of sharing, search and discovery, and the unstructured sources, such as documents, extraction of insights through analysis. Other • Increase user gravity and stickiness using natural-language processing (NLP) platforms and applications — such as text to broader, vendor-based cloud and and related text analysis techniques. They analytics and insight engines — also include platform offerings, including platform as a are also capable of storing structured data, the underlying graph technology upon which service (PaaS) including metadata that implicitly provides to build knowledge graphs and enhance structure and context. For this reason, graphs functionality. In contrast, stand-alone Benefit Rating:High enable the storage of data, the means to products are graph-based applications structure and contextualize this by building dedicated to the management of data using Market Penetration: Less than 1% of target relationships within the data, and the ability a graph-based approach in support of other audience to subject the information it encodes to products — see “Magic Quadrant for Data processing in support of varied use cases. Management Solutions for Analytics,” for Maturity: Emerging example. User Advice: Application leaders should Sample Vendors: Amazon; Apple; Baidu; employ knowledge graphs to connect Business Impact: Organizations can expect Google; IBM; Microsoft disparate concepts and enrich their data with significant value from knowledge graphs missing information. Using graph analysis, in many areas, with the following being Knowledge Graphs organic and dynamic relationships between prominent: Analysis By: Stephen Emmott; Svetlana digital assets, data sources, processes, Sicular; Alexander Linden people and interactions can be discovered • Collaboration/sharing — Interrelated and exploited automatically. A key aspect data is contextualized data, thereby Definition: Knowledge graphs encode in this respect is entity extraction, whereby aiding its discovery and findability via information (“knowledge”) as data arranged entities — people, events, etc. — can be implicit and indirect connections. in a network (“graph”) of nodes and links identified through analysis of unstructured (“edges”) rather than tables of rows and data prior to ingestion, and subsequent • Investigation and audit — With the columns. Nodes hold data or their labels; disambiguation within the knowledge graph capability to capture and disambiguate edges link nodes together, representing once contextualized. entities that map to entities in the real relationships between them. This results in world, relationships can be explored sequences of “triples” — i.e., node-edge- Knowledge graphs silently accrue “smart to identify fraud, supply chain risks or node, or Mary-manages-John — which can data” — i.e., data that can be easily read patterns of collaboration. accommodate, throughout the graph’s life and “understood” by AI systems. Although cycle, multiple and varied data schemas available as stand-alone products from niche • Analysis/reporting — Once structured without the need for redesign. Once vendors, the knowledge graphs’ benefits are in the form of a knowledge graph, encoded, information can be recalled, or typically realized through the wider platforms unstructured data can be queried, thereby synthesized, in response to queries. and applications they service. Application preprocessing it for analysis. leaders should evaluate how vendors apply

46 47

• Interoperability and automation Intel’s “Loihi” chip tackles a broader class • Identifying new skillsets that need to be — Autonomous reading and of AI workloads: Loihi offers a higher nurtured for successful development of “understanding” of data supports degree of connectivity than competing neuromorphic initiatives. integrating and operationalization of data implementations. Qualcomm, an early for different enterprise applications. exponent of neuromorphic processors, has Business Impact: Neuromorphic hardware shifted its focus to conventional processors. faces the largest barriers in advancing DNN, • Data reuse/cross-industry collaboration but also may unlock the most powerful — Being linked conceptually chunks data There are three major barriers to the results. There are likely to be major leaps and metadata, which can be shared deployment of neuromorphic hardware: forward in hardware in the next decade, if more easily and hence foster reuse. not from neuromorphic hardware, then from • GPUs are more accessible and easier to other radically new hardware designs. However, it is too early to tell whether program than neuromorphic silicon. knowledge graphs will deliver on the Neuromorphic systems promise lower broader promise. • Knowledge gaps: Programming power, but will likely operate across smaller neuromorphic hardware will require new input sets. As such, they will likely first Benefit Rating: High tools and training methodologies. appear in edge devices, where they will process images and sound. These devices Market Penetration: 1% to 5% of target • Scalability: The complexity of may also execute lower levels of a DNN at audience interconnection challenges the ability of the edge, reducing bandwidth and central semiconductor manufacturers to create processing constraints. Maturity: Emerging viable neuromorphic devices. We are in the midst of an extremely rapid Sample Vendors: Facebook; Google At the moment, these projects are not evolution cycle, enabled by radically new (Cloud Platform); Intelligent Views; Maana; on the mainstream path for deep neural hardware designs, suddenly practical DNN Microsoft (Microsoft Graph); Mindbreeze; networks (DNNs), but that could change with algorithms and huge amounts of big data Neo4j; Semantic Web Company (PoolParty); a surprise breakthrough in programming used to train these systems. Neuromorphic TopQuadrant techniques. devices have the potential to drive the reach of DNNs further to the edge of the Neuromorphic Hardware User Advice: Neuromorphic computing network, and potentially accelerate key Analysis By: Chirag Dekate; Martin Reynolds architectures can deliver extreme tasks such as image and sound recognition performance for use cases such as deep inside the network. They will require Definition: Neuromorphic hardware neural networks because they operate at significant advances in architecture and comprises semiconductor devices very low power and are potentially capable implementation to compete with other DNN conceptually inspired by neurobiological of faster training than the GPU-based DNN architectures. architectures. Neuromorphic processors systems deployed today. Furthermore, feature non-von-Neumann architectures neuromorphic architectures can enable Benefit Rating:Transformational and implement execution models that native support for graph analytics. Most of are dramatically different from traditional the neuromorphic architectures today are not Market Penetration: Less than 1% of target processors. They are characterized by ready for mainstream adoption. However, audience simple processing elements, but very high these architectures will become viable over interconnectivity. the next five years, and will deliver new Maturity: Embryonic opportunities. I&O leaders can prepare for Position and Adoption Speed Justification: neuromorphic computing architectures by: Sample Vendors: BrainChip; Hewlett Neuromorphic systems are at the very Packard Enterprise; IBM; Intel; Micron early prototype stage. IBM has delivered • Creating a roadmap plan by identifying a TrueNorth-based system to Lawrence key applications that could benefit from AI-Related C&SI Services Livermore National Laboratory. BrainChip’s neuromorphic computing. Analysis By: Susan Tan Spiking Neuron Adaptive Processor technology and Hewlett Packard Enterprise’s • Partnering with key industry leaders in Definition: Artificial intelligence (AI)- Labs Dot Product are other early entries, neuromorphic computing to develop related consulting and system integration proof of concept projects.

47 (C&SI) services are a subset of intelligent • Lack of internal skills and competencies to • Engage service providers to help you automation services to help clients ideate initiate an AI program and roadmap. understand the impact of AI on your use cases, design business or IT processes, organization’s processes and workforce, select technologies, curate data, build and • Availability of ready-to-use data for and take steps to mitigate risks, institute train models, deploy solutions, assess training AI, long lead time for training AI change management and apply and mitigate risks, and adapt talent mix to and lack of process standardization and responsible AI ethics such as avoiding successfully incorporate intelligent solutions. documentation. bias and unintended consequences, and Intelligent solutions must involve one or more developing AI systems that are secure, advanced technologies, such as machine • Limited understanding of how to scale transparent and explainable. learning, deep learning and natural- and integrate AI into existing systems and language processing. workflows. • Favor service providers that have invested in building AI-leveraged or pretrained Position and Adoption Speed Justification: • Fear of the impact of intelligent solutions solutions, such as AI-infused predictive Organizational buyers are engaging on jobs and tasks. maintenance for your industry. service providers to explore the inclusion of AI in solutions. A large majority of these Due to limited internal capabilities, when • Decide what data and IP you need to engagements (68% according to a survey organizations are ready to apply AI, a protect explicitly in contracts to avoid of 24 service providers) are in ideation, high-proportion turn to service providers for service providers using them in their exploration and proof-of-concept. To consulting and implementation. solutions for other clients. accelerate time-to-value, service providers are using rapid, phased approaches, User Advice: Clients looking to engage AI- • Ensure service providers bring the right platforms and prebuilt assets and/or related C&SI service providers should: mix of interdisciplinary consultants with pretrained models to deliver intelligent relevant experience, including technical, solutions. • Use a “start small, achieve benefits, domain and industry/process knowledge, then scale up” approach by focusing while understanding that the newness of While the market is emerging, many leading on a narrow domain use case and the the technologies means few have direct AI SIs are already working with their clients associated business outcomes where solutioning experience. on intelligent solutions, often including AI AI approaches can add value beyond with other more proven technologies. Their traditional techniques. • Get references and discuss with them track record has proven success using AI how their implementation went and what to achieve targeted business outcomes • Contract a time-boxed engagement for were areas they did not anticipate to avoid such as increased productivity, increased service providers to help build a minimum repeating the same mistakes. consistency, reduction in error rates viable product or automate defined tasks and improvement in customer retention in a single knowledge domain to make Business Impact: AI-related C&SI services and revenue, which should improve the training the AI faster and reaping the can be applied to any business process confidence of other clients using such benefits quicker. or model. A recent Gartner study found services and lead to higher adoption in the organizations are deploying the following next two to three years. • Avoid “moonshot” projects that take use cases: years of training and validation, unless However, obstacles to organizations it results in the potential to disrupt an • Predictive analytics: Providing insights, adopting AI-related C&SI services at this entire industry or bring disproportionate detecting anomalies, providing stage include: competitive advantage as a personalization, predicting likely events consequence. In this case, ensure you by using learning systems that use data • The technology is new and some aspects have contract terms that prevent service mining and pattern recognition across — for example, security, privacy, risks, providers from divulging such intellectual large amounts of data. liabilities — are still unknown. property (IP) to competitors and ensure you have executive commitment at the • Automating tasks and replacing human • Limited understanding of the capability, highest level. judgment using a combination of AI and limitations and implications of AI. RPA. An example of a business process

48 49

optimization use case is the use of Human-in-the loop crowdsourcing has three solving, model training, classification and intelligent search which can summarize key characteristics: validation capabilities compared to internal contract data — intelligent technology or traditional outsourced capabilities. ingests structured and unstructured data, 1 The ability to reach prequalified people at extracts relevant key clauses in contracts scale. Data and analytics leaders should use and policies that requires attention, human-in-the-loop crowdsourcing when: shrinking the amount of text to be read 2 The ability to aggregate human (crowd) and enables employees to concentrate contributions into meaningful results. • Rules are hard to describe for automation their time on relevant clauses. (mostly for data collection, verification 3 Engaging contributors for a specific, mostly and enhancement), such as labelling • Chatbots or virtual agents that use text information-centric, task (not as full-time images or data enrichment with data or voice to communicate with users in employees). from unspecified sources. Humans can natural language to scale call centers find the right information and their input quickly and have more compelling Position and Adoption Speed Justification: can serve as a training dataset for further personalized conversations, often in Human-in-the-loop crowdsourcing is a big improvement of the algorithm. multiple languages. shift from entrusting problem solving to the known personnel (in-house or outsourced). • The problem cannot be solved efficiently • Products with embedded sensors and AI Currently, academia and market leaders by machines, for example, when a technologies to make them smarter so — including Google, Facebook, Amazon, machine learning algorithm reaches the they learn about their environments. Microsoft, IBM, eBay, Baidu and many limit of its accuracy, humans can further others — routinely incorporate this approach. improve the output (such as content • Mechanical equipment such as drones Over the past year, adoption has been moderation, detecting subtleties in the and robots with AI that can perform tasks substantially accelerated, mostly driven by text or validation of information retrieval in dangerous or remote environments, the machine learning requirements of data and search results). and/or learn from its environment and its labeling and the quality of training data. For experience to perform better. example, the growing machine learning • Tasks or projects require rare skills (for business made CrowdFlower rebrand itself to example, data science competency). Benefit Rating: Transformational Figure Eight. Data science problem solving is Such cases usually involve competitions also on the rise: Kaggle reported accelerated and access to validated experts (via Market Penetration: 5% to 20% of target member growth in 2018. Kaggle, Topcoder, Experfy, Aigency or audience Gigster, for example). The beauty of this While the market potential is high, human-in- approach is in the self-selection of the Maturity: Emerging the-loop crowdsourcing faces many barriers participants. Another example of rare to adoption — including low awareness skills is knowledge of a narrow market or Sample Vendors: Accenture; Atos; Deloitte; of its benefits, and “perceived” (rather than a specific niche, like validating usability of EPAM; Fujitsu; IBM; Infosys; Luxoft Holding; real) concerns about quality, security and a solution in different geographies with Mindtree; PwC confidentiality. Adoption will grow alongside their cultural differences. the maturity of the overall AI market, as Human-in-the-Loop Crowdsourcing organizations will realize that human-in-the- • Traditional solutions require too much Analysis By: Svetlana Sicular; Gilbert van der loop crowdsourcing is a viable (and probably setup, for example, a one-time job for Heiden the most reliable) solution to improving coding a user interface for all kinds of accuracy of machine learning models. mobile devices. Definition: Human-in-the-loop crowdsourcing is the complementary use of User Advice: Companies working on AI and Data and analytics leaders should allow humans and algorithm-based automation machine learning should employ human- time to adjust to crowdsourcing capabilities. to solve a problem or perform a task, where in-the-loop crowdsourcing as an enabler They should also consider the potential risks, the human input further improves the of AI solutions. This approach yields more- including labor-related legal implications, IP automated AI or data management solution. fluid costs and a wider access to problem protection and inconsistent quality.

49 Business Impact: The business impact of changes dynamically as the user interacts natural-language-query-based platforms human-in-the-loop crowdsourcing ranges with data to explain key findings or the such as ThoughtSpot and AnswerRocket from modest to high: meaning of a chart or dashboard. NLG — can automatically generate a written or combines natural-language processing with spoken context-based narrative of findings • Modest, because in some instances, machine learning and artificial intelligence in the data. This accompanies visualization, organizations have been able to achieve to dynamically identify the most relevant storyboard and batch reports to fully inform savings by procuring niche solutions via insights and context in data (trends, the user about what is most important and crowdsourcing. relationships, correlations). actionable. BI teams can now also integrate stand-alone NLG engines (such as those of • High, because it enables AI, machine Position and Adoption Speed Justification: Automated Insights, Narrative Science and learning and information quality that Whereas text analytics focuses on deriving Yseop) with modern analytics and BI or data would not have been accomplished analytic insights from textual data, NLG science platforms to explain findings from without this innovative approach. is used to synthesize textual content by analytics to information consumers and combining analytic output with dynamic citizen data scientists. Narrative Science, A wide range of industries will find human- selection-driven descriptions. Automated Insights, and Yseop all now offer in-the-loop crowdsourcing indispensable. APIs for their platforms. Qlik, Microsoft (with The approach will greatly benefit analytics Although still in the early stages of adoption, Power BI), Tableau, Sisense, Information teams interested in applying human NLG is being used effectively to reduce the Builders, and MicroStrategy have integrated intelligence to unstructured text, image, time and cost of conducting repeatable with one or more of these NLG vendors. audio and video data for AI, machine analysis, such as for operational and Other partnerships are emerging. Integration learning and information quality, as well as regulatory reporting, earnings reports in the of NLG with analytics and BI platforms those who are looking for one-time solutions financial services sector, benefits statements and virtual personal assistants — such or rare skills, such data science. The tasks and weather forecasts in the government as Amazon Alexa, Apple Siri, or Google could include conditioning of training sector, and personalized messages in Assistant for conversational analytics — will data for algorithms, metadata extraction, the advertising sector. It is also used for further drive adoption. proofreading, image recognition, content data products such as sports analytics creation and classification, data collection, (personalized “fantasy football” analysis and Easy configuration and multilanguage product categorization, refining product reports), customer communications, and support will be necessary for broad descriptions, text translation, creating marketing and research information services. adoption. We expect that, due to NLG’s photos of real estate properties, and audio potentially beneficial impact on the transcriptions. The combination of NLG with modern expansion of analytics to a broader analytics and business intelligence (BI) — audience, NLG will be a feature of most Benefit Rating: High used to create analytics content including modern analytics and BI platforms by 2019. analytics applications — is one of the most We already see this happening, with most Market Penetration: 5% to 20% of target promising applications improving insights modern analytics and BI vendors offering audience derived from analytics for all users. Modern or planning to offer NLG through integration analytics and BI platforms have made with third-party NLG vendors or organic Maturity: Adolescent significant advances in visualizing data in development. interactive dashboards and storyboards, and Sample Vendors: Alegion; Amazon Web have collaboration capabilities for sharing User Advice: Data and analytics leaders Services; Experfy; Figure Eight; Kaggle; and socializing findings. However, many should: Playment; Topcoder; TwentyBN users have varying degrees of analytics skill to correctly interpret and act on statistically • Integrate NLG with existing modern Natural-Language Generation significant relationships in visualizations. analytics and BI and data science Analysis By: Rita L. Sallam; Cindi Howson Moreover, without NLG, the annotation and initiatives, or explore emerging presentation of findings is manual and static. augmented data discovery tools that Definition: Natural-language generation embed NLG with automated data (NLG) automatically generates a natural- With the addition of NLG, augmented preparation and pattern detection. language description of insights in data. analytics platforms — for example, those of Within the analytics context, the narrative Salesforce (Einstein Discovery) and search/

50 51

• Assess their organization’s readiness Context-based narration will reinforce Chatbots in social media, service desk, HR for business-user-accessible advanced mobile BI use cases, where a lack of screen or commerce, as enterprise software front analytics in terms of alignment with space is a major impediment to information ends, and for self-service, are all growing business outcomes. consumption. It will also expand the use rapidly. Still, the vast majority of chatbots of conversational analytics that combine are simple, relying on scripted responses • Monitor the NLG capabilities and NLQ, chatbots and NLG via virtual personal in a decision tree and relatively few intents. roadmaps of their analytics and BI and assistants. Moreover, it will reduce the Related to chatbots are virtual agents, which data science platforms, as well as of time and cost involved in creating regular are broader in scope and sophistication, startups. operational and regulatory batch reports. require more infrastructure and staffing to maintain, and are designed for a longer • Be aware of a solution’s maturity, Benefit Rating:High relationship with its users outside of particularly in terms of data integration single interactions. Users will interact with and preparation requirements, the Market Penetration: 1% to 5% of target hundreds of chatbots, but few virtual agents. platform’s self-learning capabilities, audience upfront set-up and configuration required, Enterprises with successful chatbot the range of languages supported, the Maturity: Emerging installations are already looking at the extent of narration for a single chart or challenge of managing multiple chatbots across a dashboard, and the accuracy of Sample Vendors: AnswerRocket; Arria from different vendors performing different the findings and narration. NLG; Automated Insights; Marlabs; use cases. It is likely that more enterprises Narrative Science; Salesforce (BeyondCore); will seek out platform offerings and • Understand potential drawbacks relating ThoughtSpot; Yseop middleware offerings as the space matures. to multilingual user scenarios, as NLG The space is currently oversaturated with requires specific libraries for each Chatbots companies and offerings, the vast majority language in use. Additionally, industry- Analysis By: Magnus Revang; Anthony of which will not manage to keep up with the specific use cases need to be considered Mullen; Brian Manusama pace of innovation as alternatives to decision carefully with respect to jargon, tone and trees, such as fact extraction and process specialized ontologies. Definition: A chatbot is a stand-alone mapping, become more common — and conversational interface that uses an app, voice and multimodality become more • Recognize that NLG could be attractive messaging platform, social network or chat viable. Looking at the investments, attention to government organizations that are solution for its conversations. Chatbots vary and research by big software companies required to have their analytics and BI in sophistication, from simple, decision-tree- in this space, we are looking at a rapid solutions comply with the Americans with based marketing stunts, to implementations evolution until we reach productivity in about Disabilities Act (in the U.S.), and similar built on feature-rich platforms. They are four years. mandates in other countries. always narrow in scope. A chatbot can be text- or voice-based, or a combination of User Advice: Business Impact: NLG supports a number both. of productivity-enhancing use cases that • Start proofs of concept for chatbots reduce the need for writers (such as of Position and Adoption Speed Justification: today — the window of opportunity for financial reports, sports analysis or product Chatbots have really increased in hype over experimentation is still here, but will likely recommendations) outside analytics. the last couple of years. But still, only 4% of close by the end of 2018. The lessons enterprises have deployed conversational from those experimental projects will be The combination of NLG with automated interfaces, which includes chatbots. However, invaluable as the technology evolves. pattern/insight detection and self-service 38% of enterprises are planning or actively data preparation has the potential to drive experimenting, according to the Gartner • Treat vendors as tactical, not strategic — the user experience of next-generation 2018 CIO Survey. This sets chatbots up for acknowledge that you’ll most likely want augmented analytics platforms. It could also tremendous growth over the next few years, to switch vendors two to three years from expand the benefits of advanced analytics but also sets it up for a large backlash once it now. to a wider audience of business users and reaches the top of the Hype Cycle. citizen data scientists.

51 • Focus on vendors offering platforms that Position and Adoption Speed Justification: using the exact same dataset, and then can support multiple chatbots The AI PaaS hype is rapidly increasing, selecting one that best addresses your with the leading cloud service providers, requirements. Business Impact: Chatbots are the face including Amazon Web Services (AWS), of artificial intelligence and will impact Google, IBM and Microsoft, clamoring to • Increase your organization’s AI project all areas where there is communication become the platform of choice. Over the success by selecting AI cloud services that between humans today. Customer service is last several years, AI applications utilizing balance your data science, developer and a huge area in which chatbots are already cloud services have continued to gain infrastructure expertise. impacting. Indeed, it will have a great impact traction and acceptance in the market both on the number of service agents employed by data scientists and developers alike. AI Business Impact: AI PaaS offerings are by an enterprise, and how customer PaaS offerings are primarily focused on focused on the three key AI portfolio services service itself is conducted. For chatbots as the three key areas of machine learning, of machine learning, natural-language application interfaces, the change from “the natural-language processing and computer processing and computer vision: user having to learn the interface” to “the vision. The AI cloud approach is beginning chatbot is learning what the user wants” to disrupt the more established on-premises • Machine learning: Packaged ML services has great implications for onboarding, data science and machine learning offered by the AI cloud service providers training, productivity and efficiency inside platform market, especially as organizations unify the end-to-end ML workflow. They the workplace. To summarize, chatbots will experiment and build AI prototypes. extend the capabilities of an isolated have a transformational impact on how we The availability of specialized hardware ML engine by providing integrated interact with technology. instances with AI-optimized chips and large access to all phases of the project — amounts of data storage makes the cloud from data preparation to deployment Benefit Rating: Transformational an ideal environment for organizations to in a managed training and execution build and deploy AI applications without environment accessible through APIs. For Market Penetration: 5% to 20% of target the risks, costs and delays of conventional technical professional teams with little audience on-premises procurement. Cloud service to no data science expertise, features providers are also offering packaged APIs like automated algorithm selection and Maturity: Adolescent and tools that make it easier for developers training-set creation will offload some of to integrate AI capabilities into existing the complexity of the project and leverage Sample Vendors: Amazon; Facebook; applications. The promise of using cloud existing expertise on operating cloud Google; Gupshup; iFLYTEK; IBM; Microsoft; services to more quickly and easily build and services. OneReach; Oracle; Rulai deploy AI solutions will push AI PaaS to the Peak of Inflated Expectations. This will be • Natural-language processing: At the Peak followed by some level of disillusionment as Organizations can use pretrained NLP AI PaaS organizations experience and understand systems to create cloud-based chatbots the limitations of AI PaaS offerings. for a variety of use cases. Major AI PaaS Jim Hare; Bern Elliot Analysis By: vendors provide a language-processing User Advice: Enterprise architecture and catalog as part of their conversational Cloud artificial intelligence and Definition: technology innovation leaders responsible platform that can be used to deliver machine learning platform services are for AI-enabled applications should take applications through a natural-language known collectively as AI platform as a service these steps: interface. (PaaS). They provide AI model building tools, APIs and associated middleware that • Consider AI PaaS over on-premises • Computer vision: This enables enable the building/training, deployment options to reduce the overhead of organizations to apply facial detection, and consumption of machine learning packaging and for easier deployment and recognition and analysis to unlock new models running on prebuilt infrastructure as elastic scalability. sources of image-based data. Pretrained cloud services. These cover vision, voice and systems require no data science expertise general data classification and prediction • Improve chances of success of your AI and allow developers to gain unique and models of any type. strategy by experimenting with different new insight by invoking an API. AI techniques and PaaS providers,

52 53

The combination of the above as cloud not make the TPU2 available other than Benefit Rating:High services will accelerate digital business through a cloud-based service. Other cloud technology platform viability in the short vendors are following suit. Market Penetration: Less than 1% of target term. audience Other dedicated silicon is coming. Graphcore Benefit Rating: High has developed a custom processor to deliver Maturity: Adolescent extreme performance for DNN-based Market Penetration: 1% to 5% of target applications and plans to launch the next- Sample Vendors: Google; Graphcore; Intel audience generation “Colossus” processor in 2018. Their marketing materials suggest close to Smart Robots Maturity: Adolescent an order of magnitude improvement over Analysis By: Annette Jump; Kanae Maita GPUs, although performance improvements Sample Vendors: Amazon Web Services; move faster than presentations. Intel is Definition: Smart robots are Google (Cloud AI); IBM (IBM Cloud); also developing an ASIC code named electromechanical form factors that work Microsoft (Azure AI Platform) “Lake Crest,” optimized for DNN, based on autonomously in the physical world, learning the technology it acquired from Nervana in short-term intervals from human- Deep Neural Network ASICs Systems in 2016. supervised training and demonstrations Analysis By: Chirag Dekate; Martin or by their supervised experiences on the Reynolds; Alan Priestley User Advice: The benefits of DNN ASICs in job. They sense environmental conditions performance and energy consumption are and recognize and solve problems. Some Definition: A deep neural network (DNN) significant. However, widespread use of can interact with humans using voice application-specific integrated circuit DNN ASICs will require the standardization language, while some have a specialized (ASIC) is a purpose-specific processor that of neural network architectures and support functional form, like warehouse robots. accelerates DNN computations. across diverse DNN frameworks. Plan an Others have general forms and/or effective long-term DNN strategy comprising humanoid appearances. Due to advanced Position and Adoption Speed Justification: DNN ASICs by choosing DNN ASICs that offer sensory capabilities, smart robots may work Deep neural networks (DNNs) are statistical or support broadest set of DNN frameworks alongside humans. models that detect and classify patterns in to deliver business value faster. Compare input data such as sound and images, or the return on investment of a GPU-based Position and Adoption Speed Justification: text patterns such as sentences. There are solution against an ASIC solution, and plan Smart robots have had significantly less two phases in DNN systems. In the training to retire the GPU solution if your business adoption to date compared with their phase, the DNN iterates across a large will perform better with a dedicated neural industrial counterparts (predefined, dataset and distills it down to a small DNN network processor. unchanged task) — but they received great parameter set. In the inferencing phase, hype in the marketplace, which is why smart the DNN uses this parameter set to classify Business Impact: Hardware acceleration robots are positioned climbing the Peak of an input such as an image, speech or will enable neural-network-based systems Inflated Expectations. In the last 12 months, text. Today, a vast majority of training and to address more opportunities in a business, we have seen some of the established inferencing tasks use GPUs. DNN ASICs can through improved cost and performance. robot providers expanding their product deliver significantly higher performance and Use cases that can benefit from DNNs line and new companies entering the lower power consumption than CPUs or include speech-to-text, image recognition market (particularly from China). Therefore, GPUs when accelerating neural networks. and natural-language processing. the market is becoming more dynamic, opening to new technology providers and Google has deployed DNN ASICs (known as IT leaders deploying deep neural network technologies, and the barriers for entry are Tensor Processing Units [TPU, TPU2, TPU3]), applications should include DNN ASICs in the slightly dropping. at scale, providing inferencing across its planning portfolio. We expect this market to businesses for, for example, speech and mature quickly, possibly within the three-year Hype and expectations will continue to build image recognition. The TPU2 and TPU3 depreciation horizon of new systems. around smart robots during the next few also accelerate the training process, a task years, as providers execute on their plans to formerly delegated to GPUs. Google does

53 expand their offerings and deliver solutions to do physical work, with greater reliability, Conversational User Interfaces across the wider spectrum of industry- lower costs, increased safety and higher Analysis By: Magnus Revang; Van L. Baker specific use cases and enterprise sizes. productivity, is common across these Hype is quickly building for smart robots as a industries. The ability for organizations to Definition: Conversational UI (CUI) is a result of several key vendors’ actions during assist, replace or redeploy their human high-level design model in which user and the past few years: workers in more value-adding activities machine interactions primarily occur in the creates potentially high — and occasionally user’s spoken or written natural language. • Amazon Robotics (formerly Kiva Systems) transformational — business benefits. Typically informal and bidirectional, these deployed robots across Amazon Typical and potential use cases include: interactions range from simple utterances warehouses. through to highly complex interactions, • Medical materials handling with subsequent highly complex results. • Google has acquired multiple robotics As design models, CUI depends on technology companies. • Disposal of hazardous wastes implementation via applications or related services or on a conversational platform. • Rethink Robotics launched Baxter and • Prescription filling and delivery Sawyer, which can work alongside Position and Adoption Speed Justification: human employees. • Patient care CUIs have seen an explosive growth in interest over the last couple of years, with • SoftBank Robotics introduced the • Direct materials handling chatbots, messaging platforms and virtual humanoid Pepper and created the assistants, especially home speakers such Pepper for Business Edition. • Stock replenishment as Amazon Echo and Google Home, all contributing to the increased hype. Still, only • In early 2018, LG introduced CLOi, a series • Product assembly 4% of enterprises have a conversational of robots developed for commercial use interface solution in production, while a in hotels, airports and supermarkets. • Finished goods movements further 38% is experimenting or planning, Also, various hotels in the U.S. and two according to Gartner CIO Survey 2018. Shangri-La hotels in Singapore now use • Product pick and pack Expected growth will be greatly fueled by robots for delivering room service. enterprises entering production from those • E-commerce order fulfillment planning and experimentation phases. User Advice: Users in light manufacturing, distribution, retail, hospitality and healthcare • Package delivery The promise of CUIs is a dramatic shift in facilities should consider smart robots as responsibility between the user and the both substitutes and complements to their • Shopping assistance interface — where the responsibility shifts human workforce. Begin pilots designed from the user having to learn the software, to to assess product capability, and quantify • Customer care the interface learning what the user wants. benefits. Examine current business- and This promise warrants a transformational material-handling processes into which • Concierge impact — even if current CUIs are far from smart robots can be deployed; also, living up to this promise. consider redesigning processes to take Benefit Rating: High advantage of the benefits of smart robots Since 2017, there has been an explosion in with three- to five-year roadmaps for large- Market Penetration: 1% to 5% of target the availability of conversational platforms scale deployment. Smart robots could also audience used to implement CUI. These tools have be a quality control (QC) check at the end of made it a lot easier for developers to the process, rejecting product with faults and Maturity: Emerging build CUIs. We have, as a consequence, collecting data for analysis. also seen CUIs being implemented inside Sample Vendors: Aethon; Amazon Robotics; popular applications as an alternative to Business Impact: Smart robots will make ARxIUM; Google; iRobot; Panasonic; Rethink GUI, and even in application suites. We their initial business impact across a wide Robotics; Savioke; SoftBank Robotics; expect application suite vendors to bring spectrum of asset-centric, product-centric Symbotic to market CUIs in front of their business and service-centric industries. Their ability

54 55

applications — which can quickly lead to Avoid retrofitting CUI front ends to existing Sample Vendors: Amazon; Baidu; hundreds of different chat interfaces being applications unless this improves usability Facebook; Google; IBM; IPsoft; Microsoft; available to employees of a large enterprise and user delight. Oracle; Salesforce; SAP — on multiple messaging platforms. The emerging pattern of chatbots acting Prepare for new roles in the enterprise. Intelligent Applications as a guide or concierge in front of these Dialogue designer, AI trainer, digital coach, Analysis By: Jim Hare; Helen Poitevin conversational interfaces will likely gain a lot humanizer and AI interaction designer are of traction over the next year. all titles Gartner is seeing in the market Definition: Intelligent applications are to support the creation of conversational enterprise applications with embedded or Most CUI implementations are still primitive, experiences. integrated AI technologies to support or and thus are not able to respond to complex replace manual human-based activities queries. Increases in capabilities will, at Business Impact: CUIs are the interaction with intelligent automation and improved first, largely come from improvements in pattern of many chatbots and virtual decision making. natural-language understanding (NLU) and assistants — both will be significant speech recognition, which will bring CUIs contributors to the impact of CUIs, especially Position and Adoption Speed Justification: closer to the promise and hype. Additional in high-touch communicative fields of AI has become the next major battleground. capabilities around context handling, user customer service and Q&A-type interactions Every application and service will incorporate identification and intent handling will likely with significant volume. AI at some level over the next few years. arrive within the next year, but will still not Enterprise application vendors are beginning be good enough to avoid a disillusionment Outside of this, CUIs will appear primarily to embed AI technologies within their phase in two-to-three years’ time. Only at in new applications. Enterprise IT leaders offerings as well as introducing AI platform that point will we see a standardization should be on the lookout for (and biased capabilities — from ERP to CRM to HCM around design methodologies for creating toward) CUIs to improve employee (and to workforce productivity applications. AI flows and personality in interactive customer) effectiveness, as well as to cut has the potential to be organizationally conversations. operating expenses and time spent learning transformational and is at the core of digital arcane computer semantics. business. Customer-facing and back- User Advice: CUIs shift the responsibility office enterprise applications are a vital for learning from the user to the software, There will also be some retrofitting. Over component of that transformation effort so the software learns what the user wants. the next five years, we do not expect large because they provide the digital foundation The impact on training, onboarding and enterprises to invest heavily in retrofitting upon which most of the endeavors rest. AI expansion of use cases is profound. The existing systems of record where the will run unobtrusively in the background of need for literacy-related training and tools employee base is experienced and stable, many familiar application categories while will thus significantly diminish during the and the feature set well-known to the giving rise to entirely new ones. next decade. Plan on CUIs becoming the user base. However, where there is high dominant model. employee turnover or significant rapid There is an AI “land grab” from both large changes in feature sets, or where enterprises vendors making “big bets” and from startups Be wary, however, of committing to CUIs too face a continuing burden of providing seeking to gain an edge. They all aim to deeply. Conversational interfaces can make computer literacy training, IT leaders need to support or replace manual human-based machines smarter and improve the ability consider creating people-literate front ends activities with intelligent automation. For of people to handle novel situations (people to make it easier for employees to adapt and example, the main enterprise software and machines collaborating will be better excel. vendors are emphasizing sales, service, than either working alone), but they also marketing, human resources and ERP as carry an extra burden. For well-developed, Benefit Rating:Transformational particularly valuable areas for applying AI repetitive skills that can be performed techniques. almost effortlessly, injecting conversation Market Penetration: 1% to 5% of target can degrade performance — unless the audience Intelligent applications will use AI in the technology is able to recognize the repetitive following ways: patterns and invoke many steps of a routine Maturity: Emerging process with a single, user-generated command.

55 • Analytics: AI can be used to create more underlying insights. Build upon these recommending which jobs to apply for and predictive and prescriptive analytics that early successes and apply the lessons answering questions or conducting initial can then be presented to users for further learned from any failures. candidate screening. evaluation, or plugged into a process to drive autonomous action. AI is also being • Be aware of “AI washing” as more and Benefit Rating: Transformational used for augmented analytics. more startups, and even aging solutions, claim AI as part of their solution. Ask Market Penetration: 1% to 5% of target • Process: AI can drive more intelligent them how they use AI to deliver advanced audience actions by an application. For example, analytics, intelligent processes and new you can use AI for intelligent invoice user experiences. Maturity: Emerging matching or analysis of email documents to improve service flow. In the future, • Prioritize investments in highly specialized Sample Vendors: Google Docs; Microsoft this can be extended further to identify and domain-specific intelligent Office 365; Oracle Applications; Salesforce patterns of work, from which process applications delivered as individual point Einstein; SAP Leonardo; ServiceNow; models can be built and executed. solutions. They can help solve problems Workday in domain areas such as customer • User Experience: Natural-language service, talent acquisition, collaboration, Digital Ethics processing used to create VPAs is one engagement and more. Analysis By: Jim Hare; Frank Buytendijk; application of AI to the user experience. Lydia Clougherty Jones Further examples include facial Business Impact: Intelligent enterprise recognition and other AI applications for applications that leverage AI can offer the Definition: Digital ethics comprises the understanding user emotions, context or following benefits: systems of values and moral principles for intent, and predicting user needs. the conduct of electronic interactions, and • Reshape how tasks are performed the use and sharing of data between people, User Advice: Enterprise application leaders by allowing workers to focus on more businesses, governments and things. should: value-adding activities through the use The scope of digital ethics is broad and of automation — via bots, sensors and includes security, cybercrime, privacy, social • Explore how AI can alter your machine learning. interaction, governance, free will, and society organization’s processes and operations and economy at large. by adding more intelligent automation, • Deliver business efficiencies at scale via dynamic workflows, and guided decision packaged AI technologies available in Position and Adoption Speed Justification: making. enterprise applications. Digital ethics jumped several positions toward the Peak of Inflated Expectations • Challenge your packaged software • Enable organizations to derive better due to the recent wake of well-publicized providers to outline how they’ll be using performance outcomes from their assets. negative press, rising public discourse, and AI to add business value in new versions new regulatory compliance including data in the form of advanced analytics, For example, in the area of Human Capital privacy considerations. Current themes intelligent processes and advanced user Management (HCM), AI is increasingly being such as “artificial intelligence,” “fake news” experiences. added to HCM applications to match talent and “digital society” are triggers driving the supply and demand, predict recruitment increased need for digital ethics. Innovations • Acclimatize employees to the idea of success, or optimize recruitment marketing. such as the Internet of Things, 3D printing, automation and “bots” by deploying Predictors include the fit of a particular cloud, mobile, social and AI are moving robotic process automation (RPA). candidate for a job, the likelihood that a faster than business, governments and candidate would be open to exploring society can organize around it or even • Develop a deeper understanding of a new job opportunity, and behavioral comprehend. The probability that unintended and expertise in AI by piloting initiatives profiles through analysis of voice or video consequences will occur is high as the use during the next two years in areas where interviews. Candidate-facing chatbots are of technology creates distance between significant opportunity exists to mine becoming increasingly common in enabling morals and actions. For business and the data quickly and efficiently to uncover further automation of this process, such as technologies used in business, a morally

56 57

agnostic stance is a position that simply • Promote trust — trust is needed to make Graph analytics are typically portrayed via a cannot and should not be sustained. the other three calls to action work. It is visualization for business user consumption. Digital ethics require societal, economic, great to take responsibility, but if your Graph analytics consists of models that political and strategic debate, new types stakeholders do not trust you to do so, determine the “connectedness” across of governance, and new processes and your offer will not be accepted. data points to create data nodes/clusters technologies to control new technologies. and their demarcation points. Nodes are Business Impact: The four areas of business connected explicitly or implicitly, indicating User Advice: Privacy rules and data impact, listed in increasing order of “moral the levels of influence, frequency of protection provide a legal minimum in development” are: interaction, or probability. handling data that is insufficient. Instead, take a “care ethics” approach to the • Submitting to compliance — staying Position and Adoption Speed Justification: application of digital technologies in the within the boundaries of the law. Graph analytics are steadily climbing to business world to reconcile principles and the Peak of Inflated Expectations, primarily consequences. The core question of care • Mitigating risk — being mindful of not due to a lack of broad awareness. The ethics is, “How do we take responsibility for using technology in ways that can upset growing adoption of graph analytics is the consequences of our actions, even if they stakeholders, or cause reputational or driven largely by the need to find insights are unintended?” (see “ethics of care” on financial risk in other ways. across an exponentially larger amount of Wikipedia). In the digital world, the concept heterogeneous data, and the demand to of care ethics is not only about people, • Making a difference — making ethical analyze it. Once highly complex models are but also about how businesses and even use of technology as a proposition that developed and trained, the output is more technologies act. Care ethics teaches that sets you apart in the market. For example, easily stored because of the expanded ethics is about taking responsibility when this could be in terms of data for good capabilities, computational power and confronted with situations you feel are not initiatives or social purpose. adoption of graph databases, which present OK. Apply “care” ethics by following these an ideal framework for storing, manipulating call to actions: • Follow your values — there is a direct and analyzing graphs. correlation between the use of technology • Be empathetic — put yourself in the and delivering value to customers, other Analytics experts are beginning to claim other person’s shoes; develop a sense stakeholders and yourself. specialization in graph analytics, and some of right and wrong that goes past just traditional analytics vendors are offering being afraid of punishment or hoping to Actively engage and participate in online capabilities that enable users to build generate a product sale whether legally data ethics and data for good initiatives such interactive network graphs — as additional or in terms of customer loyalty. as DataEthics and DataKind. features of their products. Importantly, the utilization of graph analytics is necessary • Take responsibility — taking responsibility Benefit Rating: High in order to develop knowledge graphs — a is essential for taking the lead within your highly useful output of graph analytics. ecosystem, and being the interface to the Market Penetration: 5% to 20% of target Commercialization of graph analytics is still customer or citizen. In emerging digital audience at quite an early stage, with a small number environments, taking responsibility over of emerging players. However, the unique the use of digital technologies, even if Maturity: Adolescent method of storing and processing data legally not required, builds and improves within many graph databases, combined trust. Graph Analytics with the need for new skills related to Analysis By: Mark A. Beyer; Rita L. Sallam; graph-specific knowledge, may limit growth • Display competence — build the capacity Alexander Linden in adoption. For example, knowledge and and expertise to be able to quickly and experience with the Resource Description adequately address problems. Don’t Definition: Graph analytics is a set of analytic Framework (RDF), SPARQL Protocol and RDF simply acknowledge the need to care and techniques that allows for the exploration of Query Language (SPARQL), and emerging accept the responsibility; you also need to relationships between entities of interest such languages such as Apache TinkerPop or the be able to follow through. as organizations, people and transactions. recently open-sourced Cypher.

57 User Advice: Data and analytics leaders • Epidemiology Maturity: Emerging should evaluate opportunities to incorporate graph analytics into their analytics portfolio • Genome research Sample Vendors: Ayasdi; Cambridge and strategy. This will enable them to Semantics; Centrifuge Systems; Databricks; address the high-value use cases less- • Detection of money laundering Digital Reasoning; Emcien; Intel (Saffron); suited to traditional SQL-based queries Maana; Palantir; SynerScope and visualizations (such as computing Business Impact: Graph analytics is and visualizing the shortest path, or the highly effective at both assessing risk and Prescriptive Analytics relationship between and influence of two responding to it to analyze fraud, route Analysis By: Peter Krensky; Carlie J. Idoine nodes or entities of interest in a network). optimization, clustering, outlier detection, They should also consider using graph Markov chains, discrete-event simulation Definition: The term “prescriptive analytics” analytics to enhance pattern analysis. and more. The engines used to expose describes a set of analytical capabilities fraud and corruption can also be used to that specify a preferred course of action The user can interact directly with the graph identify it within the organization and answer to meet a predefined objective. The most elements to find insights, and the analytic issues of liability in a proactive manner. A common examples of prescriptive analytics results and output can also be stored for most recent example of identifying networks are optimization methods (such as linear repeated use in a graph database. of relationships was the International programming), a combination of predictive Consortium of Investigative Journalists (ICIJ) analytics and rules, heuristics, and decision Business situations in which graph analytics research revealing the Panama Papers. analysis methods (such as influence constitute an ideal framework for analysis Relational analytics is typically ideal for diagrams). Prescriptive analytics differs and presentation include: structured, static data in columns and rows from descriptive, diagnostic and predictive in tables. Graph analytics, by contrast, is a analytics in that the output is a recommended • Route optimization new kind of “lens” for exploring fluid and (and sometimes automated) action. indirect relationships between entities across • Market basket analysis multistructured data. It can deliver the kind Position and Adoption Speed Justification: of insight that is difficult to reach with SQL- Although the concepts of optimization and • Fraud detection based relational analytics. decision analysis have existed for decades, they are now re-emerging along with a • Social network analysis Graph analytics processing is a core greater awareness of and experience with technology underlying many other advanced data science, better algorithms, cost-effective • CRM optimization technologies, such as virtual personal cloud-based computing power and available assistants, smart advisors and other smart data. In addition, the focus on the business • Location intelligence machines. Platforms such as those of prioritization of providing actionable, Cambridge Semantics, Digital Reasoning, proactive insight — as opposed to the more • Supply chain monitoring Ayasdi and Maana also use graph analytics traditional reactive reporting — has further to identify important findings. fueled the resurgence. Prescriptive analytics • Load balancing augments a user’s decision making by Graph analytics can extend the potential recommending a course of action to achieve • Special forms of workforce analytics, such value of the data discovery capabilities a defined objective. Some use cases are very as enterprise social graphs and digital in modern business intelligence and mature. These include optimization in supply workplace graphs analytics platforms. Once a graph process is chain and logistics, or combining predictive completed, it can be visualized — using size, scores with business rules for credit and • Recency, frequency, monetary (RFM) color, shape and direction — to represent lending decisions, database marketing and analysis of related networks of objects, relationship and node attributes. churn management. Many new use cases assets and conditions continue to emerge. It is, therefore, still early Benefit Rating:High days for broad adoption and awareness, There are also more specialized but many more organizations today are applications, such as: Market Penetration: 1% to 5% of target expressing an interest in prescriptive audience techniques. • Law enforcement investigation

58 59

Because prescriptive analytics often can be deployed to improve performance Position and Adoption Speed Justification: leverages and extends predictive methods, because it recommends a course of The internet giants deploy systems its adoption tends to be higher among action that best manages the trade-offs based on DNNs across their businesses. companies that have already built predictive among conflicting constraints and goals. Examples of well-developed DNN systems capabilities. Although it is a necessary Significant business benefits are common underpin Amazon Alexa’s speech-to-text competence, prescriptive analytics does and are obtained by improving the quality capability, Google’s search capability, not automatically result in better decision of decisions, reducing costs or risks, and image recognition and self-driving cars, and making. With improvement in analytics increasing efficiency or profits. Facebook’s face-tagging technology. solutions, data quality, skills and broader use of predictive analytics, prescriptive Common use cases include customer DNNs are, however, tricky to build and train. analytics will continue to advance, reaching treatments, loan approvals, claims triage, To achieve consistently good results, you the Plateau of Productivity in five to 10 years. and many kinds of optimization problems need large quantities of labeled data, data such as supply chain or network optimization science expertise and special-purpose User Advice: Data and analytics leaders and scheduling. Prescriptive analytics hardware. Most enterprises struggle to should: can also be a business differentiator for obtain enough labeled data to support their planning processes, whether it be financial or DNN initiatives. Furthermore, data science • Start with a business problem or decision production or distribution planning, allowing experts are scarce, as the IT and internet where there are complicated trade-offs users to explore multiple scenarios and giants have hired aggressively. Additionally, to be made, multiple considerations and compare recommended courses of action. optimized computational resources for DNNs multiple objectives. require a great deal of capital expenditure. Although prescriptive analytics has been • Look for packaged applications that traditionally relegated to strategic and The most widely implemented DNNs are provide specific vertical or functional tactical time horizons, more advanced convolutional neural networks (CNNs) and solutions, and service providers with the capabilities can support real-time or near- recurrent neural network (RNNs). CNNs are necessary skills. real-time decision making. This can support used, for example, for image classification automation of some operational decisions. and speech to text. RNNs are good for, • Understand the breadth of prescriptive among other things, extracting meaning analytics approaches and decision Benefit Rating:High from snippets of speech. Additionally, models available, and which best cater hyperscalers are developing generative to the nature of your specific business Market Penetration: 5% to 20% of target adversarial networks (GANs), a technology problems and skills. audience that is most useful in gameplay situations, but that will no doubt be pressed into service • Gain buy-in and willingness from Maturity: Emerging for business applications. stakeholders — ranging from senior executives to front-line workers carrying Sample Vendors: AIMMS; Decision Lens; The level of hype about DNNs is not very out the recommended actions — to rely FICO; Gurobi Optimization; IBM; River Logic; different from last year. on analytic recommendations. SAS; Sparkling Logic; Veriluma User Advice: Data and analytics leaders of • Ensure that your organizational structure Deep Neural Nets (Deep Learning) modernization initiatives should: and governance will enable the company Analysis By: Alexander Linden; Chirag to implement and maintain functional, Dekate • Explore DNNs: These technologies could as well as cross-functional, prescriptive help them solve previously intractable analytics recommendations. Definition: Deep neural nets (DNNs) are classification problems, especially relating large-scale neural networks, often with many to images, video and speech. Business Impact: Prescriptive analytics processing layers. They underpin most recent can be applied to strategic, tactical and advances in artificial intelligence (AI) by • Start with tools from cloud providers: operational decisions to reduce risk, enabling computers to process much more Wherever possible, begin by using tools maximize profits, minimize costs, or more complex data than before, such as video, available from the major cloud providers. efficiently allocate scarce or competing image, speech and textual data. resources. Importantly, prescriptive analytics

59 They have enormous resources invested photographs; and recognize and understand entertainment, new use cases in business in image, speech and facial classification speech (which, in time, may make most environments, such as hospitality services systems, and in training and data. Their devices conversational devices). have been developed. The evolution of VPAs systems will likely outperform almost into domain experts (such as digital doctors, anything you build and deploy yourself. Completely new product categories are likely lawyers and insurance agents) will open in fields such as personal assistance and additional use cases and digital business • Develop and acquire skills: Improve surveillance. models. At the end of 2017, the Alexa for your machine learning experts’ skills Business announcement launched VPA through training. Engage with academic Benefit Rating:Transformational speakers’ entry into the managed enterprise teams. Use crowdsourcing providers like environment. Algorithmia, Experfy, Kaggle and TunedIT. Market Penetration: 5% to 20% of target Although it’s currently difficult to compete audience User Advice: Technology service providers with the big cloud companies, there is a should capitalize on the current hype good stream of graduates skilled in this Maturity: Adolescent around VPA speakers by voice-enabling area, and talent will become easier to their devices and services, and join one acquire. Sample Vendors: Amazon; Baidu; or more of the ecosystems. New recurring deepsense.ai; Google; H2O.ai; Intel; revenue streams (for example, from licensing • Focus on data for deep learning as a Microsoft; NVIDIA; Skymind digital content or in-skill purchases) can be long-term investment: DNNs are within monetized, and consumer relationships and your field of competency, and the value VPA-Enabled Wireless Speakers affinities are there to be exploited. Develop of the right data will grow over time. Analysis By: Fernando Elizalde; Werner new interactive, 24-hour user experiences, Don’t assume that DNNs will derive Goertz; Annette Jump piggybacking on the further proliferation insights from any type of data through (and international expansion) of hands- unsupervised learning. So far, results Definition: Gartner defines VPA-enabled free, far-field voice and video. Recognize have mostly been achieved using wireless speakers as cloud-enabled, far- the communal and ambient nature of this supervised learning. field voice capturing devices connecting the device category, and deliver use value that user to a virtual personal assistant (VPA) differentiates between individual users and Business Impact: DNNs have service such as Alexa, Google Assistant, respects privacy and confidentiality, paying transformational, and therefore disruptive, Siri, Cortana, WeChat and others. With the particular attention to confidentiality in potential for all industries. The challenge advent of screen-enabled VPA speakers in the enterprise context. Improve consumer for those wanting to realize this potential is 2017, multimodal interactions were brought experience in relevant use cases such as to identify the business problems to solve, to the VPA experience. Our definition refers to home shopping by adding a camera and and to secure availability of enough experts stand-alone products and does not include display to aid visualization. Continue to and reasonably good datasets. DNNs implementations of VPA into products such improve the user experience beyond early demonstrate superior accuracy to past as lighting fixtures, home appliances or cars. adopters, for example, language support state-of-the art algorithms in detecting fraud, and accent recognition, relevance, setup, determining quality, predicting demand and Position and Adoption Speed Justification: troubleshooting and address privacy other classification problems that involve Even though the conversational experience concerns. During the past year, many skills sequences (using, for example, video, audio delivered by VPAs is still far from perfect, the and use cases have emerged that leverage or time series analysis). higher-than-expected consumer adoption emotion detection and emotive analysis. of VPA speakers justifies upgrading the Understanding your clients’ emotional states The basis of a DNN’s potential is its ability to position at the pinnacle of the Hype Cycle can be leveraged for improved marketing. produce granular representations of highly this year. Microsoft’s Cortana is now enabled dimensional and complex data. A DNN in Harman Kardon’s device and Apple Business Impact: The business impact of can, for example, give promising results began shipping its HomePod product, VPA speakers and the ecosystems built when interpreting medical images in order adding to forerunners Amazon Echo and around them are truly transformational. to diagnose cancer early; help improve the Google Home. Originally marketed as Through cloud integration with connected sight of visually impaired people; enable consumer and connected home products, home solutions, VPA speakers have the self-driving vehicles; colorize black-and- as a source of general information, manage potential to become the focal center of white photographs; add missing elements to connected home devices and play digital control for all smart devices in the home.

60 61

VPA speakers will have an impact on learning (where labels are omitted); and to continuous education, and set explicit e-commerce experiences, especially with reinforcement learning (where evaluations expectations that this is a learning repetitive purchases. The ability to cast are given of how good or bad a situation is). process and mistakes will be made. shopping request images to a connected screen or the addition of screens onto Position and Adoption Speed Justification: • Track what initiatives you already have the device itself facilitates the shopping Machine learning is still one of the hottest underway that have a strong machine experience and reduces transactional concepts in technology, given its extensive learning component — for example, friction. The quality of the built-in speakers range of effects on business. The drivers of customer scoring, database marketing, of these devices, and their ability to stream its continued massive growth and adoption churn management, quality control and music from online services and Bluetooth are the growing volume of data and the predictive maintenance — to accelerate players can easily displace the appeal for complexities that conventional engineering machine learning maturation through wireless music systems. approaches are unable to handle. An cross-pollination of best practices. increasing number of organizations are Monitor what other machine learning The advent of VPA speakers affects and exploring use cases for machine learning initiatives you could be a part of and accelerates digital business opportunities in and many are already in the initial phases what your peers are doing. The choice enterprise segments, such as the following: of pilot/POC. Tech providers are adding of machine learning algorithms is also embedded machine learning capabilities influenced by the ability to explain • Customer portals and interactive voice into their software. Despite the heightened how the algorithm arrived at a certain response (IVR) system replacements interest in the technology, most organizations outcome. are still dabbling in their approaches to • Home healthcare and, especially, elder machine learning. Finding relevant roles and • Assemble a (virtual) team that prioritizes care skills needed to implement machine learning machine learning use cases, and projects is a challenge for such organizations. establish a governance process to • Multifactor authentication (MFA) for As the volume and sources of data increase, progress the most valuable use cases building access, asset management and the complexity of systems will also grow through to production. homeland security applications and, in such scenarios, traditional software engineering approaches would produce • Focus on data as the fuel for machine Benefit Rating: Transformational inferior results. In the future, advances in learning by adjusting your data many industries will be impossible without management and information Market Penetration: 1% to 5% of target machine learning. governance for machine learning. Data is audience your unique competitive differentiator and User Advice: For data and analytics leaders: high data quality is critical for success of Maturity: Adolescent machine learning initiatives. Although the • Start with simple business problems choice of fundamental machine learning Sample Vendors: Ainemo; Amazon; Apple; for which there is consensus about the algorithms is fairly limited, the number of Centralite Systems; Google; Lenovo; LG; Sony expected outcomes, and gradually move algorithm variations and available data toward complex business scenarios. sources are vast. Machine Learning Analysis By: Shubhangi Vashisth; Alexander • Utilize packaged applications, if you find Business Impact: Machine learning Linden; Carlie J. Idoine one that suits your use case requirements. drives improvements and new solutions to These often provide superb cost-time-risk business problems across a vast array of Definition: Machine learning is a technical trade-offs and significantly lower the skills business, consumer and social scenarios: discipline that aims to solve business barrier. problems utilizing mathematical models that • Automation can extract knowledge and pattern from • Nurture the required talent for machine data. There are three major subdisciplines learning, and partner with universities • Drug research that relate to the types of observation and thought leaders to keep up to date provided: supervised learning, where with the rapid pace of advances in data • Customer engagement observations contain input/output pairs (also science. Create an environment conducive known as “labeled data”); unsupervised • Supply chain optimization

61 • Predictive maintenance cases based on conversational agents and Initial projects should start with modest automatic machine translation, among goals in order to demonstrate success. As • Operational effectiveness others. Existing syntactic- and semantic- experience is obtained, projects should based methods are increasingly augmented iterate and scope can increase. • Workforce effectiveness and displaced with deep neural networks (DNNs) approaches. Additional current NLP opportunities exist • Fraud detection for enterprises but are not as mature, Visible accomplishments include or will require effort before they provide • Resource optimization technologies that: consistent returns on investment. Translation or transcription services, for instance for Machine learning impacts can be explicit or • Improve natural-language parsing (via meetings or documents, offer opportunities implicit. Explicit impacts result from machine Google’s SyntaxNet, an open-source, to improve operations and lower costs. learning initiatives. Implicit impacts result DNN-based, natural-language parsing However, these NLP-based solutions are less from products and solutions that you use framework for TensorFlow). accurate than similar human-based options without realizing they contain machine and may benefit in some cases from human learning. • Translate in real time from one spoken involvement. language to another (as in Microsoft’s Benefit Rating: Transformational Skype Translator). As enterprises enhance their NLP implementations, new skills should be Market Penetration: 5% to 20% of target • Build large-scale knowledge graphs explored. Computational linguists, for audience (illustrated by the work of Google, IBM example, are versed in the manipulation and Microsoft). of various linguistic techniques and the Maturity: Adolescent impact of natural communications on users. • Offer answers instead of a list of page Upskilling of data scientist talents might also Sample Vendors: Alteryx; Amazon Web links (as in Google’s information cards). be necessary given the increasing use of Services; Domino Data Lab; Google Cloud data science techniques in NLP applications. Platform; H2O.ai; IBM (SPSS); KNIME; However, human language is complex; and Microsoft (Azure Machine Learning); while NLP solutions have made progress, Finally, the quality of NLP solutions offering RapidMiner; SAS there are many subtleties and nuances that knowledge-based consolidation, content require human intervention to enable proper mapping, search enhancements and NLP interpretation. These limitations are slowing text summarization will vary. As a result, Analysis By: Bern Elliot; Erick Brethenoux adoption. For instance, dialogue capabilities enterprise planners should test and are weak, DNNs are experimental and verify the effectiveness of these solutions Definition: Natural-language processing fragile, and understanding, inferences, before making significant commitments. If (NLP) provides an intuitive form of context and synthesis pose significant enterprises invest in specialized grammars, communication between humans and challenges. Additionally, many NLP solutions care should be taken that these be systems, i.e., NLP includes computational require specialists in order to ensure compatible across vendor solutions. linguistic techniques aimed at parsing, continued accuracy of the grammars and interpreting (and sometimes generating) models. Business Impact: To obtain clear near-term human languages. NLP techniques deal ROI and to build enterprise knowledge and with the pragmatics (contextual), semantics User Advice: NLP offers enterprises skills in the area of NLP, planners should (meanings), grammatical (syntax) and significant opportunities to improve leverage NLP applications such as: lexical (words) aspects of natural languages. operations and services. For many The phonetic part is often left to speech- enterprises, the strongest and most • Virtual assistants and chatbots to improve processing technologies that are essentially immediate use cases for NLP are related to interactions, including employee and signal-processing systems. improved customer service (impacting cost, customer services in select environments. service levels, customer satisfaction and Position and Adoption Speed Justification: upselling) and employee support (including Enterprise NLP usage is increasing as making them smarter and more effective in capabilities improve, along with new use their work).

62 63

• Text analytics to extract and summarize Definition: Robotic process automation (RPA) Users contracting for new labor-intensive the focus of textual reports and preview is a combination of user interface recognition BPO deals should consider RPA, but what questions are most common before technologies and workflow execution. It can remember that it alone is not a complete building chatbots. mimic the mouse clicks and keystrokes of a solution. Automation of existing processes human using screen and keyboard to drive is akin to the old “mess for less,” achieved • Basic transcription and translation applications and execute system-based by simply offshoring inefficient processes. services. work. It can sometimes also be designed to Therefore, users should look for solutions automate application to application. While that incorporate both process transformation • Language-generation applications that called robotic process automation, this tool and automation — reducing or eliminating produce natural-language descriptions of is not a physical object. It can be paired with tasks and activities where possible, and tabular data, making it easier for many to other tools, BPMS or AI, for example. It is a perhaps automating when not. understand. type of automation that needs structured data to work. One criticism of RPA is that it is a temporary • Keyword tagging in documents, making fix in lieu of implementing straight through it easier to determine relevant sections or Position and Adoption Speed Justification: systems. Users should consider other to extract other information such as intent The RPA concept is experiencing much hype technology solutions that may be more and entities. in consulting, business process outsourcing preferable as longer-term solutions. (BPO) and shared-service centers as it can • Content moderation services that have an impact on replacing humans in tasks Pilots and proofs of concept to date have examine user-generated content (text that are rule-based and repetitive rekeying proved that it is possible to rapidly develop or images), to flag potentially offensive or data collation. This work has often been automation capabilities and migrate work content or to identify fake news in social relocated to less expensive nearshore or from humans to software. There are many media. offshore delivery locations. advantages to finding relatively quick ways to remove people from the process, including • Sentiment analysis to identify the feeling, User Advice: RPA is a “glue” type of the following: opinion expressed in statements — from technology, which will allow you to stick negative to neutral, to positive. systems together. In order to do more • RPA tools do not make keying errors. sophisticated activities than just automate • Search improvements by better work, you need to be able to read • Automated work is time-stamped, understanding the intent of a search handwriting or structure unstructured data trackable and auditable. query as well as by summarizing content or process activities performed by chatbots that is retrieved. or machine learning activities. Whether • RPA software can cost less than you may be charged more money or not employees for comparable workloads Benefit Rating:Transformational will depend on the RPA vendor or machine — as long as the RPA tool is utilized learning vendor offerings. Users with manual optimally. Market Penetration: 1% to 5% of target rule-based processes, where data entry is audience not automated, should consider finding out • Automation potentially reduces the need where RPA makes sense, and whether their for multilingual capabilities. Maturity: Emerging BPO and IT suppliers have built RPA tools or are piloting RPA in other accounts. The • RPA tools can be designed to operate Sample Vendors: Bitext; Clarabridge; process automation technology can then be 24/7. CognitiveScale; Digital Reasoning; Google; evaluated to see if it can be programmed IBM Watson; Microsoft; Narrative Science; and used to replace employees. If the users’ • User interface recognition is less able to SAS; Yseop current BPO or consulting provider is not work in Citrix environments. piloting such technologies, it is in the users’ Robotic Process Automation Software interest to encourage their provider to do so • RPA can help address the problematic Analysis By: Cathy Tornbohm or to explore running pilots internally. issue of staff attrition.

63 • RPA can help mitigate the impact of wage recommend actions. VAs can be deployed in • Anticipate that VAs will proliferate inflation seen in most offshore locations. several use cases, including virtual personal as people and businesses move to assistants, virtual customer assistants and conversational user interfaces. Individuals Use automated business process virtual employee assistants. may use several different VAs, while discovery tools to accelerate the discovery businesses migrate from one deployment of automation opportunities for RPA and Position and Adoption Speed Justification: to multiple VAs that are composed to inform changes/enhancements to The VA space is increasingly dominated of groups of specialist chatbots, with existing scripts to cover more exceptions. by conversational interfaces such as narrowly scoped intents, working together Complement RPA with content analytics Apple’s Siri, Google Assistant, Microsoft’s with a master chatbot to coordinate the services (which may be machine-learning- Cortana, IPsoft’s Amelia, Nuance’s Nina, classification of requests. powered) to expand the types of content/ Amazon Alexa, and IBM’s Watson Assistant. channels that the RPA tools can interact with. Increasingly, behavior and event triggers • Businesses that haven’t begun the will enhance VAs. Devices such as Amazon’s process of deploying VAs to interact with Business Impact: If organizations have Echo and Google Home, together with the customers and employees should start been outsourcing or offshoring heavily broad deployment in cellular phones, have now, because customers and employees labor-based, data entry or consolidation put VAs in a position of importance in the are increasingly expecting conversational of data work, then RPA could decrease consumer’s mind. We also continue to see interfaces to be available to address help the need for as many people, increase more business-oriented VAs being created, desk and customer service desk issues. quality and lower overall process costs. with tools such as Dailogflow Enterprise RPA allows organizations to automate work Edition, Alexa for Business and Watson • Adopt the VAs that are emerging in from being manual and to look at new Assistant. Adoption grows as users get more cloud office suites first, followed by ways to automate work to deliver business comfortable with them, technologies improve SaaS offerings such as those from outcomes. Potential savings will depend on and the variety of implementations multiplies: SAP, ServiceNow and Salesforce, and your organization’s IT legacy applications consumer application environments such regarding what hasn’t already been • Unobtrusive, VA-like features — such as Facebook. automated. as Gmail’s Smart Compose with recommended sentence completion, • Look for opportunities to leverage VAs to Benefit Rating: High and the discovery features in Microsoft’s make users more productive with their Graph that find unknown resources — are business apps and mobile platforms in Market Penetration: 5% to 20% of target embedded in existing products. targeted, well-defined use cases. audience • Use-case-specific VAs have also • Incorporate analytics to measure Maturity: Early mainstream emerged — such as personal financial the impact of VAs on behavior and advisors, health and wellness coaches, performance. Closely monitor the use Sample Vendors: Automation Anywhere; and calendaring agents. of VAs, especially in virtual customer Blue Prism; Infosys; Kofax; Kryon Systems; assistant (VCA) use cases, and implement NICE; Pegasystems; Softomotive; UiPath • Chatbots that are subsets of VAs are an architecture where handoff to human increasingly used to answer customer agents is automated to ensure customer Virtual Assistants questions about products and services. satisfaction. Analysis By: Van L. Baker • VAs can act on behalf of consumers, • Utilize VAs in different use cases: including Definition:Virtual assistants (VAs) help users employees and businesses, but the customer support and engagement, and or enterprises with a set of tasks previously use cases are all based on the same, employee support and enablement, as only made possible by humans. VAs use constantly improving, language-centric well as employees’ use of personal virtual AI and machine learning (such as natural- artificial intelligence (AI) technologies. assistants for services such as HR. language processing, prediction models, recommendations and personalization) User Advice: App development leaders Business Impact: to assist people or automate tasks. VAs should develop a VA strategy that includes listen to and observe behaviors, build and voice and text enablement, because VAs will • VAs have the potential to transform the maintain data models, and predict and deliver significant benefits to the enterprise’s nature of user behavior, and customer workforce and its customers.

64 65

and employee service, as well as the of cognitive tasks. These systems are • Virtual assistants (VAs) that will assist with way work is done and how workplace interactive, iterative and stateful in dialogue, email and administrative issues activities are structured. able to recall previous interactions; they are also contextual and able to adapt to changes • Cognitive expert advisors (CEAs) that pair • There are many providers of VAs and in information and/or goals. We recognize with specialist knowledge workers to the quality varies dramatically, so expect “cognitive computing” as a promotional term solve very narrow problems and make rapid changes to the provider landscape. overused by vendors in the marketplace profound discoveries today; we do not believe these systems • VAs can be built using tools and hosted are truly capable of cognition; they merely • Computer vision (CV), augmented reality AI services licensed from providers or mimic/extend the cognitive abilities of (AR) systems that enhance humans’ created using professional services. humans. sensory abilities Performance of the VA is dependent on the quality of the dataset used to add Position and Adoption Speed Justification: Develop a mission statement with clear domain-specific information, and the Cognitive computing rapidly climbed to the objectives and planning for performance quality of the hosted-language-oriented Peak of Inflated Expectations due to the improvement via cognitive computing as AI services. pervasive promotion of the term by major part of a five-year technology adoption plan. vendors seeking differentiation in the latest Make sure to include overall employment • Security and the collection of personal generation of the AI marketplace. While goals, new policy considerations, impacts on information are still concerns, but some classes of AI such as autonomous workers, and sufficient time and resources users are growing more comfortable vehicles and virtual customer assistants may to implement communication and change in their interactions with VAs. VAs that replace human workers, cognitive computing management programs. are embedded will be the first to gain enhances them. Usability still suffers from traction; but as enterprises deploy the difficulty in assembling the right bundles of Resist the temptation to select “winners” at technology, so VAs will be broadly used technology matched to rich bodies of data, this stage, and make experimental trials by employees and customers. lack of skills to train rather than code systems, involving many vendors. and organizational and cultural acceptance. • As they mature, VAs may act for the Thus, while the hype and expectations will Employ Mode 2 development, cognitive user, forming a relationship with the user continue to build, there is considerable ergonomics and design thinking to cognitive over time. VAs shift the responsibility for disillusionment with cognitive computing computing adoption plans. understanding the process from the user still ahead. We expect these obstacles will to the system, by corresponding with the be resolved for the mainstream adopters Business Impact: Cognitive computing can user. during the next five years. This will occur as impact the business in broad and deep users demand solutions for making sense of dimensions. VAs, for example, will impact Benefit Rating:Transformational patterns in the Internet of Things (IoT), digital productivity horizontally and across many job business development and big data, coupled categories, including performers of routine Market Penetration: 5% to 20% of target with significant investment and innovation by work. Meanwhile, CEAs will impact primarily audience large and startup vendors. vertical-specific use cases in the banking, insurance, healthcare and retail sectors, Maturity: Adolescent User Advice: Beware of the hype and in the narrow fields of nonroutine, surrounding cognitive computing at this stage knowledge work. CV/AR will enhance Sample Vendors: 1-Page; Amazon; Apple; of its development, with vastly overstated human perception, decision making and Google; IBM; IPsoft; Microsoft; Nuance; expectations along the lines of an artificial productivity in utilities, mining, construction, Oracle; [24]7.ai general intelligence. manufacturing, and maintenance repair and overhaul functions. Cognitive Computing Realize that cognitive computing is not a Analysis By: Kenneth F. Brant single technology; it is a broad class of Some of the business benefits you should technologies that share an approach to seek to verify and quantify in cognitive- Definition: Cognitive computing is a class augment human cognition, ranging from: computing-based business models and of technology enabling the improved trials include: performance of a human in a wide range

65 • Higher output per dollar of selling, programmed with a procedural language. User Advice: FPGA accelerators can enable general and administrative (SG&A) Instead, they are configured with a circuit dramatic performance improvements within expenses design language that is alien to the typical significantly smaller energy consumption programmer. “VHDL” is a language which is footprints than comparable commodity • Faster cycle times difficult for most software engineers to learn, technologies. I&O leaders need to evaluate which makes FPGA programming difficult. applicability of FPGA accelerators by: • Improved productivity of field maintenance workers In the data center, FPGAs can be used in a • Identifying subset of applications that can range of use cases that require applying be meaningfully impacted using FPGAs. • Reduced risk and opportunity costs due to consistent processing operations to large poor/late decisions volumes of data, such as high-frequency • Outlining costs associated with skill set trading (HFT), hyperscale search and DNA and programming challenges. • Greater return on R&D investments sequencing. For example, Microsoft is leveraging FPGAs for search analytics and • Evaluating the availability of FPGA-based • Improved employee safety and networks, and Edico Genome’s FPGA-based hardware for use in data center server satisfaction DRAGEN Bio-IT Platform enables high- deployments — either FPGA-based PCIe performance genome sequencing workflows. add-in cards or servers with processors Benefit Rating:Transformational that integrate FPGAs. FPGAs are typically configured using Market Penetration: 1% to 5% of target hardware programming languages such • Leveraging cloud-based FPGA services to audience as RTL and VHDL that are very complex accelerate development. to use, this has held back widespread Maturity: Emerging adoption. However, the major FPGA I&O leaders should use FPGA accelerators vendors (Intel and Xilinx) are working to when: Sample Vendors: Accenture; CognitiveScale; address this with libraries and toolsets Deloitte; Digital Reasoning; Google; that enable FPGAs to be configured using • Preconfigured solutions exist that can help IBM; Intel Saffron; IPsoft; Microsoft; software-centric programming models. dramatically transform key workloads SparkCognition Adoption is also becoming easier, helped (e.g., financial trading analytics, genome by new frameworks such as OpenCL that sequencing, etc.). Sliding Into the Trough lower the time and skills required to use FPGA Accelerators FPGAs. Emerging workloads like deep • Algorithms will evolve requiring frequent learning (inference) are driving interest in updates at the silicon level that can be Chirag Dekate; Alan Priestley Analysis By: FPGAs. Intel’s integration of FPGAs with utilized by broader applications (example, mainstream server CPUs and easier access Microsoft Project Catapult). Field-programmable gate Definition: to development platforms exemplified array (FPGA) accelerator is a server-based by Amazon Web Services’ (AWS’s) FPGA- Business Impact: FPGAs can deliver extreme reconfigurable computing accelerator that enabled instance types are also driving performance and power efficiency for a delivers extremely high performance by adoption of FPGAs with the data center. growing number of workloads. FPGAs are enabling programmable hardware-level well-suited for AI workloads as they excel application acceleration. Today, the biggest growth opportunity for in low precision (8 bit and 16 bit) processing FPGAs in the data center is in the inference capabilities in exceptionally energy-efficient Position and Adoption Speed Justification: portion of deep learning workloads. Given footprints. While programmability continues FPGA accelerators feature a large array of the evolving nature of this new use case and to be a major challenge, limiting broader programmable logic blocks, reconfigurable the surrounding software ecosystem, the adoption of FPGAs, I&O leaders should interconnects and memory subsystems FPGA accelerator position has been moved evaluate FPGA-based solutions for genome that can be configured to accelerate to postpeak — 25%. sequencing, real-time trading, video specific algorithmic functions. This allows processing and deep learning (inference). FPGA processors to offload tasks from the I&O leaders can further insulate against main system processor. FPGAs are not risks by utilizing cloud-based infrastructures for provisioning FPGAs (example, Amazon

66 67

Elastic Compute Cloud F1, Microsoft Azure, identification. In turn, this has led to the rise Benefit Rating:High Baidu Cloud). of “new players” in CV (outside of academia) such as Amazon, Baidu, IBM, Microsoft and Market Penetration: 1% to 5% of target Benefit Rating:Moderate Google. Adoption is still limited, but interest is audience ramping up quickly for a number of reasons: Market Penetration: Less than 1% of target (1) the interest and hype in using DNNs Maturity: Adolescent audience (although CV can and will continue to use geometric and rule-based systems in many Sample Vendors: Amazon; Apple; Baidu; Maturity: Early mainstream circumstance) and the associated artificial Clarifai; Facebook; Google; IBM; Microsoft; intelligence hype; (2) CVs broad applicability Tencent Sample Vendors: Amazon Web Services; across numerous domains such as robotics, Baidu; Bigstream; Intel; Microsoft Azure; autonomous vehicles, drones, augmented Predictive Analytics Xilinx (DeePhi Tech) reality and virtual reality; (3) most enterprises Analysis By: Peter Krensky; Alexander are challenged over what to do with all the Linden; Carlie J. Idoine Computer Vision image/video data they are collecting and Analysis By: Tuong Huy Nguyen; Brian Blau how to automate the processing of that Definition: Predictive analytics is a form of image data; (4) computer vision is a special advanced analytics that examines data or Definition: Computer vision (CV) is a use case and natural extension for the content to answer the question, “What is process that involves capturing, processing Internet of Things (IoT). It’s the external sensor going to happen?” or more precisely, “What and analyzing real-world images and that extends and expands the reach of IoT. is likely to happen?” It is characterized by videos to allow machines to extract techniques such as regression analysis, meaningful, contextual information from User Advice: Technology innovation leaders: multivariate statistics, pattern matching, the physical world. There are numerous use computer vision to augment your predictive modeling and forecasting. different and important CV technology workforce and automate the processing of areas, including machine vision, optical image and video data. For example, it can Position and Adoption Speed Justification: character recognition, image recognition, be used in automation, such as bin picking, The excitement surrounding predictive pattern recognition, facial recognition, edge in assistive technologies for people with analytics continues to drive more interest detection and motion detection. disabilities and act as expert advisors in and adoption at all maturity levels. fields that require analysis of images and However, recent disillusionments connected Position and Adoption Speed Justification: video. In light of ongoing security and privacy to the wave of open-source adoption Building algorithms and models to solve concerns as well as legislation such GDPR, are beginning to surface in the form of vision problems have existed for more than you will also need to evaluate exposure to overinflated vendor promises and failed half a century. The convergence of deep legal liabilities associated with the collection early notebook projects that were excessively neural networks, availability of large swaths and processing of image/video data. ambitious. Still, this technology is unlikely of data and massively parallel processors to spend significant time in the Trough of has breathed new life by significantly Business Impact: Visually enabling devices Disillusionment as the rate of evolution and advancing the field of computer vision — will transform how they can interact with underlying value of predictive analytics enabling supervised and unsupervised the environment. Vision makes an excellent drives the technology rapidly toward the learning, identification, classification, complement to other sensor data, such as Plateau of Productivity in the near future. prediction and operation of CV models. For geolocation, inertia and audio. As such, it example, 30 years ago, object classification also enhances humans’ abilities to interact From those just getting started with was a difficult, manual task. Results over with the digital and physical world. This predictive analytics, to enterprises with the past eight years from the ImageNet has fueled broad interest in this discipline mature data science labs, organizations are Challenge best demonstrates current for applications such as autonomous evangelizing the value and potential impact progress in this field. Miscalculation rates vehicles, robotics, drones, augmented of predictive models. Interest is also driven have decreased by 30% — meeting and — mixed and virtual reality — security, by improved availability of data, lower-cost occasionally exceeding human levels of biometrics and more. compute processing (especially in the cloud)

67 and proven real-world use cases. Predictive Maturity: Early mainstream societal considerations, such as permits for models are no longer just produced by operation, liability, insurance and the effects data science platforms; predictive analytics Sample Vendors: Alteryx; H2O.ai; IBM; of human interaction. is being embedded in more business KNIME; MathWorks; Microsoft; RapidMiner; applications than ever before. Client SAS Autonomous vehicle technology has searches on gartner.com for “predictive disruptive potential not only for applications analytics” continue to trend steadily upward. Autonomous Vehicles in smart mobility and logistics, but also for Analysis By: Carsten Isert shipping, mining, agricultural, industrial, User Advice: Predictive analytics can be security and military operations. quite easy to use if delivered via a packaged Definition: Autonomous or self-driving application. However, packaged applications vehicles can navigate and drive certain parts Continued advancements in sensing, do not exist for every analytics use case. or the whole distance from a starting point positioning, imaging, guidance, mapping Packaged applications may also often to a specified destination without human and communications technologies, combined not provide enough agility, customization intervention by using various onboard with AI algorithms and high-performance or competitive differentiation. In these sensing and localization technologies, computing capabilities, are converging situations, organizations are advised to such as lidar, radar, cameras, GPS and to bring the autonomous vehicle closer to build solutions either through an external map data, in combination with AI-based reality. However, in 2018, complexity and cost service provider, or with typically highly decision-making capability. While self- challenges remain high, which is impacting skilled in-house staff using a combination driving cars are getting most of the reliability and affordability requirements. of open-source technologies and a data attention at present, the technology can science platform. Many organizations also be applied to nonpassenger vehicles In terms of investments, 2018 has not seen increasingly use a combination of these for transportation of goods. the major numbers as those from 2017. tactics (build, buy, outsource) and some vendors have hybrid offerings. Finally, to Position and Adoption Speed Justification: Initially, the pace of technology innovations secure the success of predictive analytics A number of signs of moving into the Trough in individual country, state and global projects, it is important to focus on an of Disillusionment have occurred in the past legislation will likely result in specific, limited- operationalization methodology to deploy year. In early 2018 several accidents with use deployments of self-driving vehicles these predictive assets. automated vehicles occurred, including the in the short term (for example, low-speed death of a pedestrian. In addition, some operations in a campus environment or Business Impact: By understanding likely announced milestones have passed without designated area within a city, and high- future outcomes, organizations are able the promised launches. speed operations on certified highways) as to make better decisions and anticipate demonstrated by the planned commercial threats and opportunities, being proactive But there has also been progress. The first launch of Waymo in Arizona. rather than reactive (for example, predictive commercial services are expected to be maintenance of equipment, demand launched in 2018. Overall, the reduced funding, the accidents prediction, fraud detection, and dynamic and evidence gathered in discussion with pricing). Interest and investment continues The efforts of automobile manufacturers industry experts leads to the prediction of to grow in both new use cases, and more and technology companies to develop autonomous driving entering the trough. traditional applications of predictive analytics self-driving vehicles have been prominently (for example, churn management, cross- featured by mainstream media, leading User Advice: The adoption of autonomous selling, propensity to purchase, database to unrealistic and inflated expectations for vehicle technology will develop in three marketing, and sales and financial the technology. Artificial Intelligence (AI) is a distinct phases — automated driver forecasting). critical technology for enabling autonomous assistance, semiautonomous and fully vehicles, and development of machine driverless vehicles. Each phase will require Benefit Rating:High learning algorithms for autonomous vehicles increasing levels of technical sophistication has accelerated. Key challenges for the and reliability that rely less and less on Market Penetration: 20% to 50% of target realization of autonomous vehicles continue human driving intervention. Automotive audience to be centered on cost reductions for the companies, service providers, governments technology and industrialization, but they and technology vendors (for example, increasingly include regulatory, legal and software, hardware, sensor, map data and

68 69

network providers) should collaborate on reduced accidents, injuries and fatalities, as are types of operations that are heavily joint research and investments to advance well as improved traffic management, which regulated in most countries. Additionally, the required technologies, as well as work on could impact other socioeconomic trends. the high cost of vertically specialized end- legislative frameworks for self-driving cars. For example, if people can use travel time to-end drone solutions — which cover the for work or entertainment while being driven devices, the supporting software and the Furthermore, educate all constituencies in an autonomous vehicle, living near a city flight operations — holds back large-scale of the benefits of self-driving vehicles. center to be close to work won’t be as critical, adoption among end users. Gartner expects Consumer education is critical to ensure which could slow the process of urbanization. commercial UAVs to approach the Plateau that demand meets expectations once of Productivity in two years, assuming that autonomous vehicle technology is ready When autonomous driving enters the Trough regulatory conditions and certain technology for broad deployment. Specific focus of Disillusionment, it might be a good elements continue to improve as expected. must be applied to the transitional phase, opportunity for new market entrants. Especially, autonomous flights will represent where autonomous or semiautonomous a boost to the market, but their enablement vehicles will coexist with an older fleet of Benefit Rating:Transformational requires both regulatory changes and nonautonomous vehicles. technological advancements. Market Penetration: Less than 1% of target Autonomous vehicles will have a disruptive audience User Advice: Overall, a corporate drone impact on some jobs, such as bus, taxi and program should have both short-term truck drivers. Develop policies and programs Maturity: Emerging and long-term objectives. This is because to train and migrate employees that will be commercial UAVs can deliver major affected by automation to other roles. Sample Vendors: Audi; BMW; Daimler operational benefits already today, but once Group; Ford Motor Co.; General Motors; the relevant aviation regulations become Business Impact: The main implications Nissan; Tesla; Uber; Volvo Cars; Waymo more permissive, their potential can shoot of self-driving vehicles will be in economic, up practically overnight. Organizations business and societal dimensions. Commercial UAVs (Drones) considering drones, therefore, should Automotive and technology companies will Analysis By: Aapo Markkanen not solely plan on the basis of available be able to market autonomous vehicles as technology, but also factor in the local having innovative driver assistance, safety Definition: Commercial unmanned aerial regulatory outlook. It makes sense to and convenience features, as well as an vehicles (UAVs, also known as drones) are proactively identify relevant regulatory option to reduce vehicle fuel consumption small helicopters, fixed-wing airplanes, changes and take advantage of them and to improve traffic management. The multirotors and hybrid aircrafts that have as early as possible. The Low Altitude interest of nonmobility companies (such no human pilot on board. They are either Authorization and Notification Capability as Waymo and Baidu) highlights the remotely controlled by human pilots on (LAANC) initiative in the U.S., aiming to opportunity to turn self-driving cars into the ground or outfitted for autonomous accelerate the waiver approvals for flying in mobile computing systems that offer an navigation. Unlike their consumer or restricted airspace, is one such example. ideal platform for the consumption and military counterparts, they are used for Today, the leading use cases include aerial creation of digital content, including location- commercial purposes. photography, mapping and surveying, based services, vehicle-centric information volume measurement, and remote and communications technologies. Position and Adoption Speed Justification: inspection. All of these can be considerably In 2018, commercial UAVs have entered the enhanced by the right of use analytics, so Autonomous vehicles are also a part of Trough of Disillusionment. In the technological as part of their UAV planning, the adopters mobility innovations and new transportation sense, such drones are relatively mature and should also take into account how they can services that have the potential to disrupt capable of increasingly sophisticated tasks. exploit the captured data in the best possible established business models. For example, However, their adoption is often held back way. Use cases involving physical tasks — autonomous vehicles will eventually lead by regulations that restrict many use cases. such as delivering objects or repairing assets to new offerings that highlight mobility-on- In particular, flying drones beyond visual — are currently largely in their nascency, but demand access over vehicle ownership by line of sight (BVLOS), above people or in they can be expected to become gradually having driverless vehicles pick up occupants restricted airspace, such as close to airports, more viable over the next two to three years. when needed. Autonomous vehicles will deliver significant societal benefits, including

69 However, benefits of commercial UAVs overlays. It is this “real world” element that User Advice: Decide on the audience for in applications that rely on completion of differentiates AR from virtual reality. AR your AR solution. Internal- and external- physical tasks will take longer to materialize aims to enhance users’ interaction with the facing solutions are not transposable. than in the ones that focus on data capture environment, rather than separating them Restrict initial trials to a specific task or goal. and analysis. from it. Set benchmarks against unaugmented solutions to understand risks and benefits. Business Impact: Most of all, commercial Position and Adoption Speed Justification: Set the business goals, requirements and UAVs are a technology to enhance the Current technology is best-suited for purpose- measurements for your AR implementation capabilities of the roles such as surveyors, built, specialized solutions. As such, position before choosing a provider. Rich and robust inspectors, drivers and cameramen who and adoption speed will vary by vertical and offerings can bring value only if you have traditionally perform labor-intensive tasks industry. Current horizontal tasks seeing a clear intention for the deployment. For in potentially unsafe conditions. As such, the most traction are task itemization, visual external-facing implementations, use AR as drones offer productivity improvements by design and video guidance. This profile an extension of your brand and experience. reducing and/or redeploying head count, represents a homogeneous view of AR For internal-facing implementations, use while enabling real-time data capture and implementations across market segments. AR as a tool that will enhance employee job improving employee safety. Examples function. This could include, for example, of industry verticals where commercial Market interest remains fairly steady delivering context-specific information at the UAVs can particularly add value include according to Google Trends, even as point of need for mobile workers, reduction agriculture, construction, emergency high-profile developments in the AR space of head count in plant and maintenance services, extractive industries, media and continue to fuel interest and hype in this operations, or enhancing business entertainment, as well as utilities. In most of area. These developments include Apple’s processes via AR-based training and the verticals, the value of commercial UAVs launch of ARKit, Google’s launch of ARCore instruction. is in reducing operational expenditure and and Magic Leap’s long-rumored HMD. improving safety, but there are also revenue- Business Impact: By leveraging device generating opportunities in industries such AR is currently struggling with mismatched sensors, AR acts as a digital extension as cinematography, surveying and logistics. expectations (vendors promising solutions of users’ senses, and it serves as an beyond current capabilities) and poor interface for humans to the physical world. It Benefit Rating:High implementations (for example, solutions provides a digital filter to enhance the user’s delivered without immersive development surroundings with relevant, interesting and/ Market Penetration: 5% to 20% of target knowledge or workflow integration, or or actionable information. audience not mapped to business value or need). B2C implementations are still struggling to AR bridges the digital and physical world. Maturity: Adolescent show consumers value. Better and more This has an impact on both internal- and transparent hardware, coupled with more external-facing solutions. For example, Sample Vendors: Boeing; Cyberhawk compelling use cases, is needed before internally, AR can provide value by enhancing Innovations; DJI; Lockheed Martin; further progress can be made. training, maintenance and collaboration PrecisionHawk efforts. Externally, it offers brands, retailers Based on Gartner inquiry and industry and marketers the ability to seamlessly Augmented Reality news, B2B AR continues to gain traction combine physical campaigns with their Analysis By: Tuong Huy Nguyen; Brian Blau as more enterprises are discovering digital assets. and seeing the value of using AR in their Definition: Augmented reality (AR) is the workflow. HMD sales reflect the burgeoning As such, AR is broadly applicable across real-time use of information in the form pilot deployments. Advancements in HMD many markets, including gaming, industrial of text, graphics, audio and other virtual hardware will provide more compelling design, digital commerce, marketing, mining, enhancements integrated with real-world hands-free use cases for AR, as well. engineering, construction, energy and objects and presented using a head- utilities, automotive, logistics, manufacturing, mounted-type display or projected graphics healthcare, education, customer support, and field service.

70 71

Benefit Rating:High KM tools are becoming more commonplace Sample Vendors: ComAround; Upland in IT organizations, and Gartner estimates Software Market Penetration: 1% to 5% of target that market penetration is between 20% audience and 50%. Many organizations struggle to Climbing the Slope realize the ROI and true value, due to cultural Virtual Reality Maturity: Adolescent issues, behavioral challenges and a lack Analysis By: Brian Blau of understanding regarding the successful Sample Vendors: Blippar; Catchoom; implementation of the underlying KM Definition: Virtual reality (VR) provides a DAQRI; Google; Microsoft; Ubimax; Upskill; practices. computer-generated 3D environment Wikitude that surrounds a user and responds to an User Advice: Knowledge management tools individual’s actions in a natural way, usually Knowledge Management Tools should be an integral part of an I&O strategy, through immersive head-mounted displays whether through stand-alone options or as Analysis By: Rich Doheny; Kenneth Gonzalez (HMDs). Gesture recognition or handheld part of an ITSM suite. Integration is necessary controllers provide hand and body tracking, to reap the benefits of a knowledge base, Definition: Infrastructure and operations and haptic (or touch-sensitive) feedback and buyers should assess which platform (I&O) leaders use knowledge management may be incorporated. Room-based systems best suits their needs. Don’t overemphasize (KM) tools to create, modify and access IT provide a 3D experience while moving the tools’ potential for success. Tools enable knowledge bases. KM tools are often linked around large areas or can be used with processes, but are only as good as the to portals that support self-service, so that multiple participants. end users can access relevant intellectual processes, procedures and policies you have in place. Formal knowledge management assets themselves. The products are defined Position and Adoption Speed Justification: governance mechanisms are crucial to by their ability to federate, store and access Immersive VR applications are more ensure that the content is reviewed, updated information about IT and non-IT services. KM advanced than other graphical simulations and corrected on an ongoing basis. tools are available as stand-alone options or and the Time to Plateau of five to 10 years is integrated components of broader IT service consistent with awareness, exposure to the Business Impact: Used optimally, a good management (ITSM) tools. technology and overall adoption is mainly knowledge base can create significant in the consumer market, but growing in efficiencies across I&O, although are often Position and Adoption Speed Justification: business. KM provides significant untapped potential targeted for incident handling, request fulfillment, training, impact assessments and for many IT organizations to optimize, VR is usually experienced using HMDs. self-service implementation. Knowledge tools drive efficiencies and realize economies of The well-known devices on the market in can drive down support costs as well as free scale in ITSM. Done correctly, it can greatly 2018 are the Oculus Rift and Oculus Go, Sony up IT service desks and other resources to improve I&O effectiveness and business PlayStation VR, HTC VIVE, Samsung Gear VR be deployed elsewhere. The effective use of user self-sufficiency. In addition, KM can and Google Daydream. VR is mature enough KM can also pay off in terms of the qualitative provide a vital component in enabling future for enterprise use, but caution is required as perspective, driving customer satisfaction and automation as a repository of information while the devices are capable, the success of VR overall customer perception. to teach emerging technologies, including usage depends on the quality of the device and chatbots and virtual support agents. Many user experience. Most VR user engagement Benefit Rating: Moderate intermediate and advanced ITSM vendors come from video games or watching video, are enhancing their products’ capabilities which can be 360-degree surround or TV and Market Penetration: 20% to 50% of target in the area of KM. As a result, the market movie content. VR HMD deployments are slow audience for stand-alone tools targeting the ITSM use but growing. New areas for VR include retail case has seen consolidation. and e-commerce, and improved HMD quality Maturity: Adolescent and system ease of use.

71 User Advice: Virtual reality can be used in a allow application developers to integrate VR Entering the Plateau variety of business scenarios: more easily into their products. GPU Accelerators Analysis By: Chirag Dekate; Martin • Complex simulation and training VR developers should consider targeting Reynolds; Alan Priestley applications immersive video game development, interactive movies and new storytelling Definition: GPU-accelerated computing • Military simulation and training, such as experiences, live events and business- is the use of a graphics processing unit flight simulators focused scenarios where using advanced (GPU) to accelerate highly parallel compute- visualization and HMDs can benefit the task intensive portions of the workloads in • Telepresence in scenarios such as remote or customer interaction point due to their conjunction with a CPU. medicine ability to offer higher degrees of visual fidelity and personalization over what flat-screen- Position and Adoption Speed Justification: • Equipment operator training based systems can provide. GPUs are highly parallel floating-point processors designed for graphics and • Entertainment and social experiences, Business Impact: VR can support a visualization workloads. Over the last such as video games or 360 surround wide variety of simulation and training decade, NVIDIA and others have added video or interactive movies applications, including rehearsals and programmable capability to GPUs, enabling response to events. VR can also shorten applications to access deep, fast-floating- • Product marketing to extend in the brand design cycles through immersive point resources. GPUs also have very high- interaction or in product design collaboration and enhance the user interface bandwidth memory subsystems. For many experience for scientific visualization, highly parallel, repetitive, compute-intensive • Architectural walkthroughs and scientific education and entertainment. Businesses applications, these capabilities deliver visualization, such as genome mapping will benefit due to VR’s immersive interfaces, dramatic performance improvements. helping create task efficiencies or reducing • Modeling, such as geomodeling in the oil costs associated with new product design, Compute-intensive applications including industry or can enhance the understanding of molecular dynamics, computational information through advanced graphical fluid dynamics, financial modeling While VR can be amazingly sophisticated visualization and simulation technologies. and geospatial applications can utilize and beneficial, the level of customization GPUs today. Programming GPUs can be can come at a high cost. Recent advances Benefit Rating:Moderate challenging, as execution order and code in HMD technologies may help ease these optimization are critical. However, toolkits obstacles, so developers should focus on Market Penetration: 1% to 5% of target like NVIDIA’s CUDA can dramatically building effective and quality experiences. audience lower the programming challenges. GPU Standards for artificial intelligence scripting, computing has moved forward on the Slope object metadata and social identity data are Maturity: Adolescent of Enlightenment to account for new use becoming more popular, due to increased cases and the evolutionary nature of deep use of personal and social networking Sample Vendors: Google; HTC; NextVR; neural networks (DNNs). technologies, which will help developers Oculus VR; Samsung Electronics; Sony; Valve; make VR more personalized and intelligent. WorldViz We anticipate that DNN technologies Technologies such as cloud graphics will mature quickly, supported by open processing and mobile video games, as well frameworks from the large cloud providers. as the proliferation of broadband access, will These frameworks include TensorFlow, Torch, Caffe, Apache MXNet and Microsoft Cognitive Toolkit.

72 73

User Advice: GPU-accelerated computing Benefit Rating: High scenarios involving high precision. They can deliver extreme performance for highly are often able to achieve a 5% to 30% parallel compute-intensive workloads Market Penetration: 1% to 5% of target reduction in error rates, which may result in in high-performance computing (HPC) audience a substantial impact on modeled metrics. and DNN training. GPU computing is also Ensemble learning is especially valuable available as a cloud service, and may Maturity: Mature mainstream for novel projects where it is difficult to be economical for applications where identify a model of best fit. However, utilization is low but urgency of completion Sample Vendors: AMD; Cray; Dell; Hewlett deployment of ensemble techniques can is high. Cloud GPUs shift the balance of Packard Enterprise; IBM; Lenovo; NVIDIA; be a computational burden to current supercomputing from on-premises toward Supermicro infrastructures. the cloud. Ensemble Learning Ensemble techniques may not be an option I&O leaders can accelerate compatible Analysis By: Peter Krensky; Alexander Linden in regulated industries, where predictive applications using GPU solutions by: models must be entirely explainable and Definition:Ensemble learning techniques are transparent. • Selecting GPU compute platforms that machine learning algorithms wherein a set of offer the most mature software stack. predictive models is created and its outputs Data and analytics leaders should combined to become the single output of understand the advantages and • Optimizing infrastructure costs the entire ensemble. This methodology disadvantages of ensemble learning: by evaluating cloud-hosted GPU draws heavily on the “wisdom of the crowd” environments for proof of concept (POC) principle, where the diversification of opinions Pros: and prototype phases. or model outputs is key. The most well-known patterns of bagging, boosting and stacking • A proven method for improving the I&O leaders should use GPU accelerators techniques include random forests and accuracy of a model that works in most when applications require extreme gradient boosting. Such techniques regularly cases. performance and have high degrees of achieve very high accuracy for supervised compute parallelism (example, many high- learning problems. • Makes models more robust and stable. performance computing and deep learning applications, etc.). Position and Adoption Speed Justification: • Can be used to capture linear as well as Adoption of ensemble techniques continues nonlinear relationships in the data. Business Impact: High-performance to steadily grow. All major data science computing and deep learning are essential vendors offer this technology as part of their Cons: to many digital business strategies. For this portfolios. Ensemble learning has become fast-growing workload, traditional enterprise a widely accessible approach for both data • Can be time-consuming in terms of ecosystems based on CPU-only approaches scientists and citizen data scientists. As the performance, and generally not well- will not suffice. Leverage mature GPU use of ensemble techniques becomes even suited for real-time applications. technologies for select HPC applications more commonplace within data science and deep learning infrastructures. teams, the technology will reach the Plateau • Selection of models for creating an Programmability challenges have been of Productivity in the next 12 months. ensemble is a skill that can take time to largely solved in GPU-accelerated computing master. by frameworks like CUDA. I&O leaders can User Advice: For even a moderately skilled minimize risk by using cloud-hosted GPU data science professional, ensemble environments for testing and evaluation. techniques are relatively easy to apply to

73 Business Impact: Almost every predictive Market Penetration: 20% to 50% of target In tandem with algorithmic advances, analytics use case and machine learning audience speech-to-text applications have been task can benefit significantly from the propelled by hardware progress, the application of ensemble techniques. Success Maturity: Early mainstream adoption of conversational agents such stories of applied techniques continue to as chatbots and virtual assistants by bolster ensemble learning’s reputation for Sample Vendors: Alteryx; DataRobot; H2O. enterprises and consumer adoption of increasing predictive accuracy. Ensemble ai; IBM; KNIME; Microsoft; RapidMiner; SAP; speech interactions on smartphones, game methods are frequently deployed in SAS consoles and, in particular, virtual personal analytics competitions such as the KDD assistant speakers such as Amazon Echo Cup and Kaggle competitions, and acquit Speech Recognition and Google Home. The use of speech- themselves finely. Analysis By: Anthony Mullen to-text technologies is also growing for connected home and automotive domains Data and analytics leaders should inquire Definition: Speech recognition technology and embedded solutions running on edge with their data science teams as to how translates human speech into text for further devices without the need for cloud to create and when they are employing ensemble processing. new usage scenarios. techniques. These techniques are robust against outliers and overfitting. Predictions Position and Adoption Speed Justification: User Advice: From a human computer based on ensemble methods can be used Speech recognition performance has interface standpoint, speech recognition is as rank-ordering scores or interpreted rapidly accelerated in the last three applicable and useful where users have: as regression functions, making them years. Heavyweights such as IBM, especially useful for a wide range of tasks Microsoft, Google, Amazon and Baidu all • An interest or motivation, e.g., injuries or across financial services and marketing demonstrated rapid scientific improvements disabilities. applications, where customer behavior in 2016-2017, claiming equal or better needs to be predicted. performance than human transcription. • Their “hands busy, eyes busy” and need data entry or system control performed Ensemble techniques can offer an invaluable Services continue to improve especially via voice alongside other tasks, i.e., in new perspective on a model and provide around developer and process support, warehouses, factories, hospitals, shop validation for existing models already in for example, in 2018, Google overhauled floors, cars or homes. production. Ensemble learning is also a well- their Cloud Speech to Text API to improve established idea creation tool for data science performance by providing multiple machine • A need for sustained, voluminous or teams working in the business exploration learning models to suit different use-case repeated input such as office, medical and and advanced prototyping use cases. contexts (e.g., phone call, voice commands, legal dictation, particularly in applications video) as well as improving punctuation to where speech shortcuts can be used to Benefit Rating: High make transcriptions more readable. insert commonly repeated text segments.

74 75

• Domain knowledge but not system support and transcription, sales support, After a series of breakthroughs with the knowledge. i.e., interactions are voice access to analytics and reports technology and while the rapid pace has expressed through natural speech, rather to hands-free warehousing and virtual eased, there is still a regular cadence of than proprietary system commands and employee assistant (VEA) use cases. innovation and improvement in areas interfaces. Further, there are legal imperatives, such such as edge-based speech to text, as GDPR, compliance and redaction, that hybrid models using semantic and DNN Typical use-case scenarios of adoption require businesses to be able to obtain techniques and GPU/TPU hardware. These include: transcripts of voice calls. gains were largely driven by deep learning. Using techniques like convolutional neural • Supporting Users. Consumer electronics Vendors in this space can generally be split networks (CNNs), long short term memory providers should consider the use broadly into two camps — general purpose (LSTM), recurrent neural networks (RNNs). of speech recognition services for platforms and specialists that provide a Also, end-to-end neural architectures using applications, smartphones, smart homes managed service. Generalist platforms connectionist temporal classification (CTC) and cars, either licensing technology to tend to cover many languages and target loss (championed by Baidu) are improving work online/offline for their own devices general purpose speech. Specialists offer time to train models. or using cloud services to enrich the tailored solutions designed to perform well experience and presence of their devices for a specific business context and lexicon Tech heavyweights like Google, Apple and and services. using custom dictionaries and semantic Microsoft also collect large troves of training tools to work with DNN models to improve data from opt-on programs with consumers • Supporting Customers. Speech disambiguation. and this ongoing cycle of training data and recognition for telephony and contact improved algorithms will see the issue of center applications enable enterprises Making speech to text work for most speech to text as a largely solved problem to automate call center functions such as organizations entails more than simply within the next two years. Specialist vendors travel reservations, order status checking, activating an off-the-shelf solution. with custom language models designed ticketing, stock trading, call routing, Organizations should plan for an extended for verticals will continue to be essential directory services, auto attendants and period of human involvement to monitor, to organizations looking to embed this name dialing. Further applications include train and improve performance — especially technology deep into their business. the use of speech to text for marketing around modelling proprietary business and commerce interactions. terms, dialects and noisy/complex Benefit Rating: Transformational environments. • Supporting Employees. Existing Market Penetration: More than 50% of enterprise application developers should Business Impact: Unlike other elements target audience consider the use of speech recognition of the natural-language processing chain, and natural-language entry as a speech to text (and text to speech) can be Maturity: Mature mainstream method of simplifying UIs and increasing considered to be a stand-alone commodity productivity. There are an increasing where its modules can be plugged into a Sample Vendors: Amazon; Baidu; Google; amount of use cases for speech to text variety of natural-language workflows. iFLYTEK; IBM; Intelligent Voice; Microsoft; in the workplace from meeting room NICE; Nuance

75 Appendixes

Figure 3. Hype Cycle for Smart Machines, 2017

Source: Gartner (July 2017)

76 77

Hype Cycle Phases, Benefit Ratings and Maturity Levels

Table 1. Hype Cycle Phases

Phase Definition Innovation Trigger A breakthrough, public demonstration, product launch or other event generates significant press and industry interest. Peak of Inflated Expectations During this phase of overenthusiasm and unrealistic projections, a flurry of well-publicized activity by technology leaders results in some successes, but more failures, as the technology is pushed to its limits. The only enterprises making money are conference organizers and magazine publishers. Trough of Disillusionment Because the technology does not live up to its overinflated expectations, it rapidly becomes unfashionable. Media interest wanes, except for a few cautionary tales. Slope of Enlightenment Focused experimentation and solid hard work by an increasingly diverse range of organizations lead to a true understanding of the technology’s applicability, risks and benefits. Commercial off-the-shelf methodologies and tools ease the development process. Plateau of Productivity The real-world benefits of the technology are demonstrated and accepted. Tools and methodologies are increasingly stable as they enter their second and third generations. Growing numbers of organizations feel comfortable with the reduced level of risk; the rapid growth phase of adoption begins. Approximately 20% of the technology’s target audience has adopted or is adopting the technology as it enters this phase. Years to Mainstream Adoption The time required for the technology to reach the Plateau of Productivity. Source: Gartner (July 2018)

77 Table 2. Benefit Ratings

Benefit Rating Definition Transformational Enables new ways of doing business across industries that will result in major shifts in industry dynamics High Enables new ways of performing horizontal or vertical processes that will result in significantly increased revenue or cost savings for an enterprise Moderate Provides incremental improvements to established processes that will result in increased revenue or cost savings for an enterprise Low Slightly improves processes (for example, improved user experience) that will be difficult to translate into increased revenue or cost savings Source: Gartner (July 2018)

Table 3. Maturity Levels

Maturity Level Status Products/Vendors Embryonic • In labs • None Emerging • Commercialization by vendors • First generation • Pilots and deployments by industry • High price leaders • Much customization Adolescent • Maturing technology capabilities and • Second generation process understanding • Less customization • Uptake beyond early adopters Early mainstream • Proven technology • Third generation • Vendors, technology and adoption rapidly • More out-of-the-box methodologies evolving

Mature mainstream • Robust technology • Several dominant vendors • Not much evolution in vendors or technology Legacy • Not appropriate for new developments • Maintenance revenue focus • Cost of migration constrains replacement Obsolete • Rarely used • Used/resale market only Source: Gartner (July 2018)

78 Evidence Evidence for this note was garnered from:

• “Forecast: The Business Value of Artificial Intelligence, Worldwide, 2017-2025”

• Gartner 2018 CIO Survey conducted among 3,138 respondents

• Gartner 2018 Artificial Intelligence Consumer Perceptions Survey conducted among 4,019 respondents.

• Gartner search analytics.

• Gartner client inquiry analytics.

Source: Gartner Research Note G00357478, S. Sicular, K. Brant, 24 July 2018

79 A New Era Empowered By Artificial Intelligence

80