An Investors' Guide to Artificial Intelligence - Flamingo AI · PDF fileGlobal Equity Research 27 November 2017 An Investors' Guide to Artificial Intelligence AI adoption at an inflection

Global Equity Research27 November 2017

An Investors' Guide to Artificial IntelligenceAI adoption at an inflection point

Global Technology

Stacy Pollard AC

(44-20) 7134-5420

[email protected]

Bloomberg JPMA POLLARD <GO>

J.P. Morgan Securities plc

Varun Rajwanshi

(44-20) 7742-2814

[email protected]


Mark R Murphy

(1-415) 315-6736

[email protected]

J.P. Morgan Securities LLC

Sterling Auty, CFA

(1-212) 622-6389

[email protected]


Tien-tsin Huang, CFA

(1-212) 622-6632

[email protected]


Doug Anmuth

(1-212) 622-6571

[email protected]


Alex Yao

(852) 2800-8535

[email protected]

J.P. Morgan Securities (Asia Pacific) Limited

Gokul Hariharan

(852) 2800-8564

[email protected]


Andreas Willi

(44-20) 7134-4569

[email protected]


Paul Coster, CFA

(1-212) 622-6425

[email protected]


See page 109 for analyst certification and important disclosures, including non-US analyst disclosures.J.P. Morgan does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of this report. Investors should consider this report as only a single factor in making their investment decision.

www.jpmorganmarkets.com

While the concept of AI has been around for over 50 years, we are at a pivotal point for its adoption today due to the availability of big data, high-powered computing and advances in algorithms – which all make AI cheaper and faster to implement. This report is intended to serve as both an educational primer on AI as well as an investment framework highlighting JPM-covered companies currently offering AI solutions (43 companies profiled).

What is AI? In simple terms, artificial intelligence is the simulation of human intelligence by machines. AI is different from traditional software programs in that it extracts knowledge from data and can alter its behavior (or learns) without being specifically programmed. Traditional software pre-defines the logic, whereas AI discovers the patterns and logic. These ‘self-learning’ systems are impacting nearly every industry vertical from manufacturing to financial services, giving rise to new business models while making some legacy models obsolete.

The scope of AI applications is huge and growing. AI can be used by enterprises to: 1) Drive sales and customer engagement – AI can improve the overall customer experience in a multi-channel world with the use of recommendation systems, virtual assistants, chatbots and AI-managed marketing platforms. 2) Improve operational efficiencies – AI functions are enhancing quality control, predictive maintenance and prescriptive responses, 3) Enhancing products with embedded AI, and 4) generate new insights and enable new business models. AI tools will be intimately linked with the overall digital transformation going on now within businesses, and AI is likely to be embedded in numerous technology applications within a few years.

AI will continue to take share within the IT budget. IDC forecasts spending on AI-focused hardware, software, and services to reach $58bn by 2021, up from ~$12bn in 2017, making this one of the fastest-growing technology segments (growing at nearly 50% 2017-2021 CAGR). We believe AI functionality (direct or embedded) will take market share within the IT budget, and therefore vendors that are ahead of the curve in embedding AI will benefit. The AI platform vendors have taken an early competitive lead; however, we believe there is ample opportunity both for established traditional IT enterprise vendors to embed AI into their core offerings and for a variety of new, emerging players to offer creative solutions (from niche and vertical-specific to even broad-scope industry-disrupting ideas).

Within our global software/internet stock coverage, we highlight the following companies as well positioned to benefit from AI adoption.

J.P. Morgan coverage companies which stand to benefit from the move to AI

Sector CompaniesInternet Google, Amazon, Facebook, Baidu, Tencent, AlibabaAI PaaS Amazon, Microsoft, Google, IBM, Alibaba, BaiduEnterprise Software Vendors SAP, Salesforce, Microsoft, Oracle, Adobe and othersSpecialist AI functions Dassault Systèmes, Hexagon, NICE, Siemens, Software AGIT Services Accenture, Capgemini, Atos, CognizantSource: J.P. Morgan.

Further contributing analysts:

Alexander Mees

Sandeep Deshpande

Toby Ogg


Completed 27 Nov 2017 01:09 PM GMTDisseminated 27 Nov 2017 04:35 PM GMT

2


Stacy Pollard(44-20) [email protected]

The AI Software Investment Framework

Artificial Intelligence (AI) is not a winner-takes-all market, in our view. On the one hand, we could argue that digital-only companies (Alphabet, Facebook, Microsoft, etc.) will always dominate in the new information age. On the other hand, many traditional companies also own and regularly create valuable vertical-specific information (on their customers, products, processes, supply chain, etc.). At the very least, these companies need to make sure their information assets are digitized, and then utilize AI and other analytics or operational efficiency tools to maximize the value of the data.

We see digital transformation as a prerequisite for AI adoption and believe thatdigitally mature enterprises will emerge as the lead adopters of AI and thus be able to reap the benefits that AI brings to the enterprise (be that improvements in operational efficiency or product offerings with embedded AI capabilities). AI continues to grab a growing portion of overall IT spending. Accordingly, we believe that the strategic shift toward AI adoption will continue to benefit digital transformation enablers.

Given that data (large amounts of reliable and diverse data sets) is the fuel powering current mainstream AI technologies (such as deep learning), we see platform vendors with well-established ecosystems as better positioned against the competition to take advantage of AI-driven innovation. Furthermore, large incumbent enterprise software vendors should have a role to play, given that they generate the current system of record and intimately understand the IT systems and needs of the enterprise.

Consultants and IT services vendors, which play a key role in designing, implementing, and integrating digitization, automation, and AI applications for enterprises, should continue to benefit as enterprise AI adoption grows. Finally, specialist and industry-specific AI vendors should fare well with productized or semi-packaged AI solutions for specific functionality.

Within our global technology/internet stock coverage, we believe the following companies are best positioned to benefit from Artificial Intelligence.

Figure 1: J.P. Morgan coverage companies which stand to benefit from the shift to AI

Region Technology SectorCompanies best prepared for the shift to AI

AI function / use Analyst

Europe

European Software SAP, Dassault Systèmes, HexagonAI platforms, AI-based analytics, object/pattern recognition, predictive maintenance, smart cities.

Stacy Pollard

European IT Services Capgemini, AtosAdvisory & education, creation and solution design, deployment and integration.

Stacy Pollard

European Industrial Tech SiemensMindSphere cloud IoT, digital factory, robotics, industrial optimization.

Andreas Willi

America

US Enterprise Software Salesforce, Microsoft, AdobeAI platform, predictive analytics, automation bots, speech recognition and NLP, image recognition.

Mark Murphy / Sterling Auty

US IT Services IBM, Accenture, CognizantAI platform, analytics/data discovery, chatbots, natural language, computer vision, advisory & implementation.

Tien-Tsin Huang

US Applied & Emerging Tech NICE, CognexRobots as digital assistants, utilizing real-time customer data in chatbots, machine vision.

Paul Coster

US Internet Google, Amazon, FacebookSearch, cloud, home assistants, autonomous vehicles, photos, news feed…many others.

Doug Anmuth

Asia Asian Internet / software Baidu, Tencent, AlibabaAI Platforms, autonomous driving, conversational (NLG, voice), computer vision, customer analytics, etc.

Alex Yao / Gokul Hariharan

Source: J.P. Morgan.

3



The internet companies (like Google, Amazon, Facebook, Baidu, Tencent, Alibaba, etc), as pure digital entities, have taken the lead in both generating big data and utilizing AI to analyze big data.

However, many companies are also offering AI PaaS (AI Platform-as-a-Service – a combined package of big data, cloud computing power and AI/ML toolkits) to companies that may not have their own existing capabilities. Key AI PaaS vendors include Amazon, Microsoft, Google, IBM, Alibaba, Oracle, and Baidu.

Since digital transformation is necessary to take full advantage of AI, we believe there is an opportunity for the many technology vendors that are facilitating the digital transformation – including enterprise software vendors like SAP, Salesforce, Microsoft, Adobe and others.

Many companies have focused on one of several aspects of AI – such as computer vision, natural language (text and speech), chatbots, geolocation, predictive design and pattern recognition, etc. Several coverage companies offering these AI functions include: Dassault Systèmes, Hexagon, and NICE. Other companies are specializing in IoT, industrial optimization, predictive maintenance or robotics (like Siemens and Software AG).

Consultants and IT services vendors play a key role in designing, implementing, and integrating digitization, automation, and AI applications for enterprises. Top picks amongst our coverage include: Accenture, Capgemini, Atos and Cognizant.

Internet

AI PaaS

Enterprise software vendors

Specialist AI functions

IT Services

4



Table of ContentsThe AI Software Investment Framework ................................2

10 FAQs on AI...........................................................................7

1. What is AI? .........................................................................................................7

2. Is the current hype around AI justifiable?.............................................................7

3. Why now?...........................................................................................................7

4. What are some of the top AI use-cases (by industry)?...........................................8

5. What is the market size for AI?............................................................................9

6. Is AI a competitive advantage? ..........................................................................10

7. What does AI bring to the enterprise? ................................................................11

8. Who are some key AI vendors? .........................................................................11

9. Jobs and social disruption? ................................................................................11

10. Will AI make humanity obsolete? ....................................................................12

AI will be a disruptive force ...................................................13

Why Now?............................................................................................................15

AI in Enterprise Software.......................................................16

Layers of the AI ecosystem ...................................................18

Data is the New Oil .................................................................19

AI Brings Life to IoT ...............................................................20

The Five Senses of AI ............................................................22

3 – ‘See’: Computer Vision enabling a host of next-gen applications......................24

4 - ‘Analyze and Act’: Ushering in a golden age for data science ...........................25

5 - Remember (Knowledge): Data Discovery and Integration.................................26

The AI application ecosystem ...............................................28

What does AI bring to the Enterprise? .................................30

Is AI a competitive advantage?..............................................................................30

How are enterprises thinking about AI? Build, buy or outsource?...........................31

Enterprise AI adoption still in its infancy; growing digitization and awareness to spur growth ..................................................................................................................31

Leading Enterprise AI Use Cases by Industry ........................................................33

The AI-spending opportunity.................................................34

AI-focused spending to witness sharp growth in the coming years; Software / Services to take the biggest slice............................................................................34

AI-focused spending still a small portion of overall enterprise IT spending.............36

Will AI boost overall IT spending?........................................................................36

Mapping the Vendor Landscape ...........................................38

AI: Is this a winner-takes-all market?.....................................................................38

Categorizing AI vendors........................................................................................40

AI Platform-as-a-Service.......................................................................................41

AI Platforms and deep learning frameworks/libraries .............................................41

Traditional enterprise software vendors .................................................................42

5



Data Analytics and BI Vendors..............................................................................43

Data and its role in shaping the competition landscape...........................................44

Data integration tools play a role ...........................................................................45

Specialist AI / Vertical-focused / Niche players .....................................................45

Industrial applications, IoT+AI, IT/OT integration and robots ................................46

AI-enhanced customer-facing channels..................................................................48

IT Services............................................................................................................48

Intelligent things/devices .......................................................................................49

Specialized Infrastructure for AI............................................................................51

Cybersecurity—immediate benefactor to machine learning ....................................52

Appendix 1: Artificial Intelligence 101 ..................................53

Historical foundations (philosophy and mechanics)................................................53

Untangling the jargon............................................................................................55

Man vs. machine – the intelligence perspective......................................................56

AI vs. machine learning vs. deep learning..............................................................56

Why do we want machines to learn?......................................................................57

Machine learning – An alternative to rules-based programs....................................57

‘Hey Siri, what is machine learning?’ ....................................................................57

Automating the learning process............................................................................58

Types of learning and application areas .................................................................59

Neural networks and deep learning ........................................................................60

Artificial Neural Networks – A simplified model of the human brain .....................61

Deep learning process flow....................................................................................62

Can machines think and reason?............................................................................63

Roadmap to AGI...................................................................................................64

The Singularity .....................................................................................................64

Employment in the face of AI: Disruption or Evolution? ....65

Predictable / Routine tasks most susceptible to automation.....................................65

Will AI be a net creator or destroyer of jobs? .........................................................67

Appendix 2: What companies are doing in AI......................68

Accenture (ACN US, OW) ....................................................................................68

Adobe (ADBE US, OW) .......................................................................................71

Alibaba .................................................................................................................71

Alphabet (GOOGL US, OW) ................................................................................72

Alteryx (AYX US, N) ...........................................................................................73

Amadeus (AMS SM, N) ........................................................................................74

Amazon (AMZN US, OW)....................................................................................75

Atos (ATO FP, OW) .............................................................................................76

Aveva (AVV LN, OW) .........................................................................................77

Baidu (BIDU US, N).............................................................................................77

Barracuda Networks (CUDA US, OW)..................................................................79

Box (BOX US, N).................................................................................................79

Capgemini (CAP FP, OW) ....................................................................................79

Cloudera (CLDR US, OW)....................................................................................80

Cognex (CGNX US, UW) .....................................................................................80

6



Cognizant (CTSH US, OW) ..................................................................................81

Cornerstone OnDemand (CSOD US, N) ................................................................81

Coupa Software (COUP US, N).............................................................................81

Dassault Systèmes (DSY FP, OW) ........................................................................82

Face++/Megvii (Private company).........................................................................84

Facebook (FB US, OW) ........................................................................................84

FireEye (FEYE US, N)..........................................................................................86

Fortinet (FTNT US, OW) ......................................................................................86

Hexagon (HEXAB SS, N) .....................................................................................86

HubSpot (HUBS US, OW) ....................................................................................86

IBM (IBM US, N).................................................................................................87

iFlytek (002230 CH, Not Covered) ........................................................................88

Imperva (IMPV US, N) .........................................................................................89

Indra (IDR SM, N) ................................................................................................89

Micro Focus (MCRO LN, N).................................................................................91

Microsoft (MSFT US, OW)...................................................................................92

NICE Ltd. (NICE US, N) ......................................................................................92

Oracle (ORCL US, OW) .......................................................................................93

Palo Alto Networks (PANW US, OW) ..................................................................93

Sage (SGE LN, OW).............................................................................................94

salesforce.com (CRM US, OW).............................................................................94

SAP (SAP GR, OW) .............................................................................................94

SenseTime (Private Company) ..............................................................................98

Siemens AG (SIE GY, N)......................................................................................99

Software AG (SOW GY, N) ................................................................................100

Sophos (SOPH LN, OW).....................................................................................101

Splunk (SPLK US, N) .........................................................................................102

Tencent (700 HK, OW) .......................................................................................102

Twilio (TWLO US, OW).....................................................................................103

Ultimate Software (ULTI US, OW) .....................................................................104

Workday (WDAY US, OW)................................................................................104

Yitu (Private Company) ......................................................................................104

Appendix 3: AI acquisitions 2012-2017 ..............................105

Appendix 4: Other J.P. Morgan reports covering the topic of AI............................................................................................106

7



10 FAQs on AI

1. What is AI?

In simple terms, artificial intelligence is the simulation of human intelligence by machines. For example, the development of computer systems with the ability to learn, reason, discover meaning, perceive environment, learn from experience, and interact.

In practice, AI is a group of technologies that help facilitate the discovery and analysis of information for the purpose of making predictions and recommendations, support decision making, facilitate interactions and automate certain responses. AI tools will be intimately linked with the overall digital transformation going on now within businesses, and AI is likely to be embedded in numerous technology applications within a few years.

AI is different from traditional software programs in that it extracts knowledge from data and can alter its behavior (or learn) without being specifically programmed. Traditional software pre-defines the logic, whereas AI discovers the patterns and logic.

2. Is the current hype around AI justifiable?

Yes; while it is relatively early days, we fully believe AI will be a disruptive force in the coming years. With advances in computing power, increasing sophistication of learning algorithms, and vast swathes of data streams available for training, computers are now able to perform tasks that were once considered the exclusive domain of the human mind. These ‘self-learning’ systems are impacting nearly every industry vertical from manufacturing to financial services, giving rise to new business models while making some legacy models obsolete.

Nearly every major technology player is employing machine learning (ML)-driven intelligence in some form or another as the relevance of this technology across different consumer and enterprise applications grows. Further, thanks to the availability of open-source software platforms / libraries and availability of cheap computing power (via cloud service providers such as Amazon, Google, Microsoft, etc.) several start-ups have emerged that are extending the scope of available applications of ML. Businesses are now deploying AI as a differentiator in their offerings – be that software platforms such as chatbots for customer service, AI-enabled systems for fraud detection, customer analytics, etc. or hardware devices such as smartphones, wireless speakers, and intelligent robots (incl. autonomous cars). We firmly believe that AI will be at the center of the revolution that will continue to shape the broader industry landscape in the coming decades.

3. Why now?

If Artificial Intelligence has been around since the 1950s, then why is interest gaining momentum now? In short, it’s the availability of big data, computing power, the cloud and advances in algorithms – which make AI easier, cheaper and faster to implement. 1) Massive sets of data are now available to train and validate the AI algorithms, and IoT will provide an even greater explosion of data. 2) Computing power has improved exponentially, and the use of GPUs has greatly accelerated computations required for AI workloads. 3) The cloud makes aggregated large data sets and cheap computing power available to data scientists who want to build and

8



train AI algorithms. 4) And finally the virtuous feedback loop enabled by massive computing power and big data gives data scientists the opportunity to hone, refine and perfect their algorithms.

Why not now? Barriers to adoption

Some of the key challenges impeding the adoption of AI across enterprises include a shortage of available ‘AI-talent’ (data scientists, machine learning experts, etc.), difficulty in porting existing infrastructure to AI-ready infrastructure, challenges in quantifying value derived from deploying AI solutions, and lack of a strategic directive from top management (C-level management, business unit heads, etc.). In addition, there are concerns surrounding the ‘black-box’ nature of AI algorithms (the way ML algorithms interpret data to produce an outcome or a set of outcomes is often not known), which hinder its adoption (especially in highly regulated industries, such as financial services).

Having said this, we believe that growing enterprise digitization, increasing awareness of AI solutions and their direct/indirect benefits, and rising sophistication of AI algorithms will help drive enterprise AI adoption. Further, as the marketmatures, we expect consolidation across the AI software vendor landscape and this should help enterprises select vendors more effectively for AI-related initiatives, aiding adoption of AI. While adoption may be a point of discussion at this stage, we believe that AI will be ubiquitous across all the industry verticals and all digitally mature enterprises in the future.

4. What are some of the top AI use-cases (by industry)?

Current mainstream AI models (such as machine learning) learn from training examples by identifying and stitching together different features that are representative of a particular segment within the input data. These algorithms work well in situations for which there is a large amount of data available for training and the relationships between the different features in the data are more or less steady. Some of the prominent use-cases for AI currently include:

Digital virtual agents, chatbots, NLG (natural language generation), robo-advisors, automated customer service agents, human emotion perception,

Customer analytics, sensor data analysis from IoT, legal/contract analysis, healthcare diagnosis and personalized treatment/drugs,

Quality control, predictive machine maintenance, yield improvement, warehouse automated retail stock checks, automation, logistics & fleet management, industrial and consumer robots,

Computer vision, image analysis and tagging, autonomous cars, smart surveillance cameras; automated geophysical feature detection, object detection (avoidance and navigation), geospatial awareness, video analysis,

Automated threat intelligence; prevention against cybersecurity threats.

The table below presents a snapshot of different use cases by industry verticals. An important point to note here is that machine learning algorithms are only as good as the data that is used for its training. Hence, biases / errors inherent in the datasets will more often than not lead to broken machine learning algorithms that yield incorrect results.

9



Table 1: Sample AI Use Cases Across Different Industry Verticals

Industry vertical Sample AI Use Cases

Banking & SecuritiesAutomated trading & investment discovery, trading strategies, robo-advisors, voice-based commerce, customer behavior analysis, chatbots for customer services, identity verification, and fraud detection.

GovernmentSmart surveillance, threat detection, Smart Cities and Utilities, AI-enhanced and personalized education and training, chatbots for info distribution and citizen engagement.

Manufacturing & Natural Res.Predictive maintenance, machine learning driven insights for yield improvement, optimization.

Comms, Media & ServicesCustomer analytics, forecasting and customer demand trends, video analytics, computer vision interactivity (e.g. in video games and other immersive media).

RetailCustomer analytics, forecasting, anticipating demand trends, reducing revenue churn, supply chain management, warehouse automation, chatbots for customer services, conversational commerce.

InsuranceClaims management and fraud detection, analyzing customer behavior and reducing revenue churn, automated underwriting, pricing, conversational platforms for customer services, complying with regulations, trading strategies.

UtilitiesEnhanced supply-demand management based on AI-driven analytics, predictive maintenance, dynamic pricing based on consumption analytics (provided by smart meters, for example), chatbots for customer service.

Healthcare ProvidersDiagnostics, image analytics for early disease detection, drug discovery, patient monitoring (pre-emptive warning systems), personalized medicine and treatment, telehealth (like inpatient robots), VR for surgical training and simulation.

TransportationSelf-driving vehicles, Advanced driver assistance systems, personalized content delivery / productivity enhancement tools used by providers of transportation services.

EducationCustomized / adaptive learning programs, skill upgrade programs based on real-time insights gleaned from job market trends.

Wholesale TradeWarehouse automation, inventory management based on insights gleaned from demand analytics, autonomous delivery.

Sources: J.P. Morgan, Gartner, Capgemini.

5. What is the market size for AI?

It is important to note that AI is not a product offering in itself – it is essentially a model (a set of software tools / algorithms) that helps identify patterns and associations in large amounts of data. Further, AI can either be embedded in hardware products or software platforms or be deployed as a component in a larger software process implementation. Hence, rather than focusing on AI, it is helpful to talk about the markets / applications enabled by AI, which can then be assigned a dollar value.

IDC forecasts spending on AI-focused hardware, software, and services to reach $58bn by 2021, up from ~$12bn in 2017, making this one of the fastest-growing technology segments (growing at nearly 50% ’17-’21 CAGR). The components of this AI-focused spend include:

AI applications – applications that learn, discover, and make recommendations / predictions or core AI components.

AI software platforms – tools built on core AI components that enable AI-driven use cases.

AI-related IT & business services – for example: consulting / implementation services provided to an enterprise for deploying AI-related technologies.

AI-dedicated server and storage spending (hardware).

10



Figure 2: Global AI-focused spending* ($, bn)

Source: AI-spending estimates from IDC. *Includes AI-focused spending on hardware, software (applications + software platforms),

and services (IT consulting & system implementation).

To put this spending size into perspective, AI growth projections of ~50% CAGR through 2020/21 are more than twice the growth rate of other high-growth tech sub-sectors (such as Big Data 23% and Cloud 20%). By 2020, AI-focused spending couldbe about the same size as the security software market.

Figure 3: AI market growth compared to other high-growth technology segments

Source: Bloomberg, IDC, J.P. Morgan.

6. Is AI a competitive advantage?

Yes, we believe that early adopters of AI functionality will see significant benefits, which also speaks to the current urgency in the market for enterprises and governments to consider their approach and strategy around AI.

Whereas last year most organizations were testing proof of concepts in AI, 2017 has seen more AI applications moving into production. It is still early stage, though, and we expect to see continued enhancements and fine tuning for these applications to find their full value. Today many AI applications are automating existing repetitive tasks, but we see the future as very much about whole new processes or even business models being created because of the capabilities of AI.

Leaders in digital transformation will also be leaders in adoption of AIAcross industries, we believe those corporates addressing their own digital transformation now will also be the leaders in the adoption of AI. Certainly, the puredigital companies have already shown a lead in using AI (Google, Amazon, Uber,

5.3 7.5 12.0

57.6

-

10.0

20.0

30.0

40.0

50.0

60.0

70.0

2015 2016 2017E 2021E

AI-focused spending*

AI-focused spending, 50%Big data, 23%

Total cloud, 20%

Security, 8%

0

50

100

150

200

250

0% 10% 20% 30% 40% 50% 60%

Approximate 2016-2020 CAGR

App

roxi

mat

e 20

20 m

arke

t siz

e ($

, bn)

48% CAGR

11



Netflix, etc.), and the faster other industries digitize, the faster they can take advantage of the insights and efficiencies of AI.

7. What does AI bring to the enterprise?

Artificial Intelligence is a buzzword in every boardroom, and enterprises want to use AI for multiple purposes. 1) Drive sales and customer engagement – AI can improve the overall customer experience in a multi-channel world with the use of recommendation systems, virtual assistants, chatbots and AI-managed marketing platforms. 2) Operational efficiencies – AI functions are enhancing quality control, predictive maintenance and prescriptive responses. Efficiencies can range from reducing operational costs and churn to increasing legal/regulatory compliance and improving machine and process performance. 3) Enhancing products with embedded AI. Incorporating AI into a product or service can provide end-customer benefits. 4) Generating new insights, and enabling new business models. Better data analysis is allowing companies to think differently and often more creatively. Employees are spending less time in routines, and more time thinking of new products, new go-to-market strategies and engaging with and understanding customers at a higher level.

8. Who are some key AI vendors?

As is characteristic of any nascent market, the AI vendor landscape is in a state of flux, spanning across large technology firms to fledgling start-ups. Although vendors such as Google, Facebook, Baidu, Apple, Microsoft, and Amazon may have taken early advantage of their large data sets and established platforms, the large enterprise software vendors (like SAP, IBM, Oracle, salesforce.com, Sage, etc.) are gradually embedding AI into multiple aspects of their offerings. In addition, we expect certain revolutionary aspects of AI to bring to market new and/or specialist vendors. We have categorized vendors into several functional areas of AI (while noting that many vendors cross functions may fit into multiple categories). Key vendor categories (and sample vendors) include:

AI Platform-as-a-Service: Amazon, Microsoft, Google, Alibaba, Baidu, IBM

Enterprise software vendors: SAP, Oracle, Microsoft, Salesforce.com, Adobe

Data analytics vendors: SAS, SAP, Oracle, Tableau, Mathworks, Qlik, Palantir, IBM, Informatica

Specialist AI / Vertical-focused / Niche players: Clarifai, Kore.ai, Aspect, IPSoft, Iflytek, Sensetime

Industrial IoT + AI: Siemens, GE, PTC, SAP, Amazon, Software AG

IT Services: Accenture, IBM, Capgemini, Atos, Cognizant

9. Jobs and social disruption?

Today, AI-enabled systems can technically replace not only low-skilled jobs but also ones that require a high degree of expertise (such as reading CT scans). While adoption of self-learning AI systems will result in meaningful productivity improvements, its impact on the job market in the near to medium term cannot be ignored. As per a McKinsey study, ~50% of time spent by the US workforce is highly susceptible to automation using currently available technologies. This could potentially lead to job polarization – concentration of jobs either in the high-paying, non-routine cognitive and low-paying, non-routine manual buckets (jobs least

12



susceptible to automation by currently available technologies) – and job displacement, especially in certain verticals such as manufacturing, customer service, financial services, etc. that have a higher proportion of automatable jobs, and associated social unrest.

However, overall we believe AI is more likely to supplement humans and make them more productive. In addition, while many of today's job functions will be done by machines and computers, we think many new roles as yet unknown will be created for humans. No doubt we will need to adapt our skills, education and perhaps social systems to the evolving AI-embedded world.

10. Will AI make humanity obsolete?

The political and philosophical debate we mainly leave to others, but in our opinion there is no need for immediate panic. We are actively implementing narrow artificial intelligence (task focused), but so far nothing close to (an all-powerful and all-knowing) artificial general intelligence. That said, there are risks in not understanding the logic behind even some of the simple tasks already automated by AI (such as resumé filtering, criminal sentencing recommendations, etc.). But as for a dystopian future where humans are no longer needed or somehow made extinct? Well, all things are possible (and AI is not our only threat here), but it seems unlikely machines would have the sense of purpose, motivation, emotional engagement with the future, and sense of limited time that drives humanity’s intelligence. There is also a strong argument to suggest that humans and AI merge…but that is a debate for a different time.

13



AI will be a disruptive force

With advances in computing power, the increasing sophistication of learning algorithms, and vast swathes of data streams available for training, computers are now able to perform tasks that were once considered the exclusive domain of the human mind. These "self-learning" systems are impacting nearly every industry vertical from manufacturing to financial services, giving rise to new business models while making some legacy models obsolete. Nearly every major technology player is employing ML-driven intelligence in some form or another as the relevance of this technology across different consumer and enterprise applications grows. Further, thanks to the availability of open-source software platforms / libraries and availability of cheap computing power (via cloud service providers such as Amazon, Google, Microsoft, etc.) several start-ups have also emerged that are extending the scope of available applications of ML. Businesses are now deploying AI as a differentiator in their offerings – be that software platforms such as chatbots for customer service, AI-enabled systems for fraud detection, customer analytics, etc. or hardware devices such as smartphones, wireless speakers, and intelligent robots (incl. autonomous cars).

McKinsey notes that companies invested between $26bn and $39bn in AI-related technologies in 2016, with large technology and digitally mature manufacturing firms comprising roughly $20-30bn of these investments (private financing comprised the remaining $6-9bn). The US government invested around $2bn in AI technologies last year (Source: White House paper: Artificial Intelligence, Automation and the Economy) and the Chinese government has made AI a national priority – estimating that by 2020 the scale of core industries of AI will exceed 150bn yuan ($23bn).

Despite the scale of AI investments (which continues to grow), AI is far from being pervasive across the consumer and enterprise landscape. AI adoption has just begun and its disruptive potential can already be seen by instances where organizations have deployed AI – we highlight three sample examples below:

Netflix recommender system: A billion dollars per year in value

According to a paper1 published by Netflix, 80% of hours streamed on Netflix are based on recommendations (driven by its recommender system), while the remaining 20% is search-driven. A typical Netflix user spends roughly 60-90 seconds searching / reviewing titles before losing interest. Thus, if the user does not find a title of his/her liking within 90 seconds, Netflix runs the risk of the user abandoning its services. This is where the company's recommender system, driven by machine learning, comes in. Netflix estimates the combined impact of personalization and targeted recommendations (enabled by its recommender system) saves the company more than $1bn per year. In fact, broadly speaking, ‘personalization’ ranks among the top categories of AI application areas that are used by customer-facing organizations to increase engagement, loyalty and mind-share. A recent BCG report (Profiting from Personalization, May 8 2017) mentions that personalization will drive a revenue shift of ~$800bn over the next five years to the 15% of companies (in the retail, health care, and financial services verticals) that employ it effectively.

Firms often use different nomenclature such as artificial

intelligence, cognitive

computing, deep learning, machine intelligence when

referring to applications driven

by self-learning systems. While keeping the broader distinctions

in mind, we will refer to ML-

driven intelligence as AI in the subsequent sections.

1The Netflix recommender

system: Algorithms, Business

value, and Innovation by Carlos A. Gomez-Uribe and Neil Hunt.

14



AI in Journalism: 12x increase in earnings coverage2

Associated Press (a news service agency) could only cover roughly 6% of all the US corporate earnings (5,000+ companies) in 2013 given its limited staff of business writers. The agency began working with an AI firm, Automated Insights (which specialized in Natural Language Generation), in 2014 and estimates that employing AI solutions freed up roughly 20% of journalists' time. The freed up time (which was previously consumed by the manual work of poring through earnings data) could now be used for more complex / quality work. By 2015, the AI system used by Associated Press was writing reports on 3,700 corporate earnings, more than 12x increase in coverage vs. 2013. No jobs were lost – rather, journalists could now focus more on qualitative, high-value-added work.

AI enabling hundreds of millions of dollars in datacenter-related power savings

Power consumption and cooling costs comprise a meaningful chunk of overall datacenter operating costs and thus minimizing these costs is one of the key priorities of any datacenter operator. Google, which has an installed base of several million servers, decided to take on this problem using AI (technology acquired by its DeepMind acquisition in 2014) to manage power usage across its datacenters. This resulted in nearly 15% savings in overall power consumption (after accounting for electrical losses and other non-cooling inefficiencies), according to DeepMind. Considering Google’s overall datacenter power consumption, this could translate to meaningful savings running into hundreds of millions of dollars.

Figure 4: AI-driven reduction in power consumption in Google's datacenters

Source: Adapted from DeepMind blog titled: DeepMind AI reduces Google datacenter cooling bill. ML = Machine Learning. PUE =

Power Usage Effectiveness overhead.

The broader enterprise complex is just starting to realize the potential benefits available from deploying AI solutions. Indeed, we believe that enterprise AI adoption is set to gain rapid pace over the coming years as the benefits derived from AI become more tangible, awareness improves, and enterprises become more digital. This is likely to drive a shift in enterprise spending, away from traditional ‘run-the-business’ applications toward more ‘transform-the-business’solutions. This creates both opportunities and risks for enterprise software vendors. We will discuss these dynamics in the subsequent sections.

2 The future of augmented

journalism: A guide for newsrooms in the age of smart

machines, by Francesco

Marconi, Alex Siegman, and multiple AI systems.

https://deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/

15



Figure 5: Survey: technologies having the greatest impact on the firm in the next decade

Source: NewVantage Partners; survey respondent sample had a heavy financial services / insurance skew (~75%).

Why Now?

If Artificial Intelligence has been around since the 1950s, then why is interest peaking now? In short, it is the availability of big data, computing power, the cloud, and advances in algorithms – which make AI easier, cheaper and faster to obtain.

1. Big data and Internet of Things (IoT). Massive sets of data are now available to train and validate the AI algorithms. Image analysis is being perfected through the large and growing numbers of digital images being stored on the Internet and various clouds (viewed as testing grounds for AI models). The wide variety of text, documentation, and audio is useful for training natural language generators as well as speech recognition. IoT will also provide an explosion of data with multiple purposes, from predictive maintenance to Smart Cities.

2. Computing power: Improvements in computing power have more or less followed Moore's Law (an observation which states that the transistor count of an integrated circuit will double every 18 -24 months). This has enabled tremendous innovation in the electronics industry, from the room-sized computers in the 1950s to the point that we can now carry a much more powerful computer in our pocket. However, analyzing troves of datasets to make computers learn still proved to be a daunting task for many cutting-edge microprocessors until researchers found that the use of Graphics Processing Units (GPUs), traditionally used for computer game graphics, greatly accelerated the computations required for AI workloads. Advancements in AI research and increased sophistication of algorithms have paved the way for new alternative computing architectures (such as Google’s TPU) specially designed for AI-related computations. The diversity and power of computing elements available today are enabling rapid advances in the field of AI. Indeed, as Shane Legg of Google’s DeepMind notes, training an AI algorithm that would take one day on a Google TPU would have taken a quarter of a million years using a cutting-edge 1990s microprocessor.

3. The Cloud makes big data more widely accessible. Aggregating more and more data into the Cloud makes big data more accessible to a wider set of users, which boosts opportunities for AI as well. The same goes for cheap compute and data engineering in the cloud.

4. Algorithmic sophistication. The virtuous feedback loop enabled by massive computing power and big data gives data scientists the opportunity to hone, refine and perfect their algorithms. It also widens the opportunity for many more potential applications.

5%

8%

12%

26%

44%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50%

Blockchain

Cloud computing

FinTech solutions

Digital tech | Internet of Things (IoT)

Artificial Intelligence | Machine Learning

16



AI in Enterprise Software

In time, AI is likely to exert a profound impact across the enterprise software landscape. AI is expected to drive the next leg of automation in Enterprise software, spanning across Intelligent Bots acting as the first point contact for Customer Service use cases, to cataloguing huge volumes of data (text/images) to Robotic Process Automation (Robotic Process Automation is the use of software robots, created using Machine Learning and AI, to imitate humans to handle high volume, repeatable tasks). However, the true impact of AI in software stretches beyond automation, into revenue generating/cost saving functions by predicting actions based on historical patterns, thus enabling customers to do things like increasing customer retention (by predicting churn) or preventing downtime (via predictive maintenance on machines/infrastructure).

Discussion of AI is spreading like wildfire among enterprise software companies. The incorporation of AI can be seen across the entire stack, from the application layer (SAP, Salesforce, HubSpot) to the platform layer (Microsoft, Cloudera) to the Infrastructure layer (Nutanix). Among enterprise software companies, the cloud computing companies are in a unique position to take advantage of AI due to the massive volume of data that these companies store and process, which is a crucial component for delivering effective AI solutions.

Incorporating AI directly into the application layer delivers value to customers much faster, as they do not have to invest in infrastructure/resources to build Machine Learning models. For example, SAP’s Leonardo Machine Learning tools embed AI into existing enterprise applications – like Service Ticket Intelligence which processes inbound social media posts and emails, or the Customer Retention application which can anticipate customer behavior. Salesforce also provides a great example: it has incorporated AI (called Einstein) into the application layer to drive higher value to customers across its stack of clouds in Sales, Service, Marketing, Commerce and Platform. Based on the data that salesforce already stores in its platform, Salesforce Einstein can predict the likelihood of a deal closure, can allow its customers to build a marketing campaign to engage with the customers at the right time to have the maximum impact, or provide product recommendations to end-customers using AI built into the salesforce Commerce Cloud. Other SaaS companies such as HubSpot are also using AI to offer such capabilities as Predictive Lead Scoring or allowing customers to create intelligent bots, while companies like Workday or Cornerstone are using AI to improve employee retention and recommend learning content.

While incorporating AI into the application layer drives a faster time to value, it will likely provide a narrower set of use cases, relatively restricted to pre-defined models. Allowing customers to use AI/Machine learning as a platform broadens the opportunity set and enables customers to apply AI to use cases of their choosing based on their own Machine Learning/Deep Learning models. However, one important point to consider in using an AI platform is the infrastructure investment, since Machine Learning or Deep Learning algorithms are very iterative and compute-heavy processes thus requiring a very powerful infrastructure environment. As such, an AI platform offered as a PaaS layer on a pay-as-you-go licensing model could have its advantages, as customers wouldn’t have to invest in the infrastructure environment and could focus on the algorithms/models they want to build. Microsoft is one of the very few companies that is offering an AI platform as a service as part of Microsoft Azure. Microsoft not only offers a set of Machine Learning services including an ML studio, workbench (free), data prep and model management

17



capabilities, it also offers a set of cognitive services such as Image recognition and Speech recognition, which can be incorporated by developers into their applications. Other companies such as Cloudera are also making strides in this area by offering Spark as a platform for AI along with a new Data Science Workbench.

The application of AI can also be seen on the infrastructure side, led by companies such as Nutanix which recently introduced Nutanix X-fit (Cross-fit) which can not only predict problems with hardware before they occur but can also offer suggestions, for instance suggest if a workload should be hosted on-premises or in the cloud based on SLA requirements and cost.

18



Layers of the AI ecosystem

AI in its purest form is just self-learning software, so it is important to examine the broader AI ecosystem to better understand the interdependency of the entire IT stack. See Figure 6 below (source: Cognizant, and Frank, Roehrig, & Pring).

The Experience layer includes the interface, device or application - which is essentially how humans interact with the underlying AI software.

The Intelligence layer encompasses the AI algorithms, as well as the broader software ecosystem which provides task-specific functionality, and of course the middleware which bridges the gap between applications, networks, etc.

The Data layer is critical for AI, and the sources of data can be very wide. Systems of record inside the organization create data, as does the Internet of Things (sensors, cameras, etc.). Data beyond the enterprise is also highly valuable to bring context (particularly to customer analysis or competitive environment), and this can include social media, web search, geo-location, etc.

The Infrastructure layer, which includes the cloud, data centers, networks, etc., supports the collection, storage and transference of data.

Figure 6: Layers of the AI ecosystem

Source: Adopted from Cognizant and "What to Do When Machines Do Everything", Frank, Roehrig, and Pring.

INTELLIGENCE

EXPERIENCE

DATA

INFRASTRUCTURE

Interface - listen/talk/voice, see, typing, motion, etc.

Device - PC, mobile phone, virtual reality glasses/headsets, car, home, etc.

Application - The logic and workflow of a process - e.g. booking a car, processing a claim, making a trade, or analysing an MRI screen etc.

Software Ecosystem

AI - algorithms, automationprocesses, machine learning, neural networks.

Process Middleware

AI

Process Middleware - software that acts as a bridge between applications, networks, etc.

Software Ecosystem -linkages to other software tools via APIsthat provide task specific functionality.

Data & Metadata from other sources -social media, web search, geo-location, etc

Cloud, data centers, networks etc.

IoT Infrastructure - sensors, cameras, and connectivity.

Systems of Record - ERP, SCM, etc.

19



Data is the New Oil

Data is the fuel of the digital economy

The same way as oil powered an industrial revolution, data will drive the Fourth Industrial Revolution. As a raw material, the uses of oil are limited, but once refined, the uses are vast. Data in its raw state also has fairly limited uses, but once aggregated and analyzed the uses of data appear almost unlimited. And with data, the more the better – such that quantity creates its own quality.

We are in the early days of the Fourth Industrial Revolution where big data, analytics and AI are becoming increasingly relevant and useful. But we see the potential as every bit as large as the transformation that refined oil has brought to the global economy over the last 100 years (impacts on energy, transport, and creating whole new industries based on plastics, petrochemicals, etc.).

AI in our view will also transform nearly every industry and create new ones. Already, Netflix, Amazon and Uber are using AI and automation to enhance customer experience (e.g. recommendations, mapping/routing/scheduling) in ways that have allowed them to take market share from incumbents.

Data is also superior to oil in several ways:

The value of data is exponential. The value of oil is linear (one barrel is worth $60, 10 barrels are worth $600), but data increases in value at an exponential rate. For example, combining one set of data with another is more than 1+1 because the interaction, correlation and relationships between the data can be worth far more than 2.

We are creating data exponentially. Oil is ultimately a finite resource, and it is expensive and time consuming to obtain. Data on the other hand is relatively cheap to create, store and manipulate – with the technologies getting better, faster, cheaper all the time. We frequently hear statistics about how much data is crossing the internet, or how much is being collected by the Internet giants. On top of this, the Internet of Things (IoT) is about to create another exponential step-up in the volume of data collected around the globe. Data-generating sensors will be embedded into just about everything, from cars and assembly lines to toothbrushes, light bulbs and other IoT end-points.

Digital twins. In the not-too-distant future, we expect there will be a digital twin for just about every physical object in the world. Digital twins will exist for vehicles, airplanes, ships, trains, buildings, plants, geospatially-tuned landscapes, people, genetic codes, drugs, and networks themselves (such as the electricity grid or the Uber network of cars).

Data can multiply itself and is easily distributed. Not only is data easily replicable, but lessons learned by one AI machine can be handed directly to another machine (assuming they aren't already networked). Compare this to the slow rate at which humans gather and transfer knowledge (basically, humanity has benefitted enormously from its collective intelligence, but there are certain things at a local level which each human must learn for himself).

Digitize everything, refine your own oil

On the one hand, we could argue that digital-only companies (Alphabet, Facebook,Netflix, etc.) will always dominate in the new world. On the other hand, many

How big is big data?

IDC predicts that global data volumes will reach 180 zettabytes by 2025 (a zettabyte is 10

21, or

more precisely 270

bytes). Compare this Cisco’s estimate that global internet traffic will total 1.3 zettabytes this year.

20



traditional companies also own and regularly create valuable information (on their customers, products, processes, supply chain, etc.). At the very least, these companies need to make sure their information assets are digitized, and then utilize AI and other analytics or operational efficiency tools to maximize the value of the data. To repeat a theme we’ve mentioned several times in this report: digital transformation and AI will go hand-in-hand.

There are several data problems still facing AI applications today:

Weak, biased, incomplete (including small sample sizes), poorly tagged, non-relevant or outright incorrect data can lead to invalid analysis or undesirable outcomes.

Data security will become increasingly important, including regulatory requirements around the protection and physical storage of personal data in particular (GDPR – Global Data Protection Regulation in Europe, for example).

Concerns surrounding the ‘black-box’ nature of AI algorithms (the way ML algorithms interpret data to produce an outcome or a set of outcomes is often not known).

Narrow applications of AI may fail to understand external impacts or omit future relevant data.

Lack of an ethical framework, or a machine’s inability to truly understand ethical implications and comprehend context.

Data haves and have-nots (potential monopoly power of an entity controlling vast amounts of data for its semi-exclusive use).

AI Brings Life to IoT

AI + IoT: A Partnership to Drive Scale

Internet-of-Things (IoT), and the associated interconnectivity (via the internet or local networks) of computing devices and sensors embedded into everyday objects, will enable the receipt and distribution of vast sums of data. The relationship between AI and IoT is symbiotic and complementary in nature; AI requires huge volumes of data from which it can learn, and IoT benefits from self-learning, trained algorithms in order to process, interpret and render useful insights from the volumes of data generated. IoT devices (e.g. temperature sensors, noise sensors, airborne drones equipped with cameras, energy meters) will serve as a portal for information and data capture. AI + IoT relationships can be structured in three overarching categories:

IoT generated data as an input to the AI. Here the IoT end point serves as the data gatherer, and feeds the Machine Learning model in the AI system with captured data. Ongoing training through data processing will make the model more 'intelligent'. An example of this type of relationship can be reflected on a construction site, an area where emphasis is placed on site safety, and resultantly site managers are constantly assessing potential safety hazards. An AI system connected to IoT endpoints (e.g. surveying and monitoring tools, on-site cameras, drones) will be embedded with a trained model that can identify potentially hazardous objects, and its accuracy will continuously improve over time as it is fed more data from the IoT end points.

21



AI as a component of the IoT System. In this structure, AI will be a feature or component in the IoT system. The AI will help to interpret and make sense of the data collected by the IoT enabled device, for the specific application the system has been designed for. Visual impairment support devices, for example, coupled with an AI inference engine can help to interpret the visual feed data, and convert the data into useful audio to help the user.

AI and IoT in a system of feedback. In this setup, the IoT system continuously feeds data to the AI system, resulting in an AI that is continuously evolving and updating. Over time and with improved ‘intelligence’, the AI produces and deploys a new production system, which then aids and enhances the IoT system’s inferencing capability. A key use case for this particular structure is in autonomous vehicles; in which an up to date inferencing engine and AI system are imperative. Changing variables such as weather conditions, roadworks and pedestrian flow among others require the inferencing engine embedded into the sensors to be continuously updated and be alert to the changing environment.

Figure 7: AI + IoT Data Cycle

Source: Gartner.

IoT AI

AppsDecisionmaking

Analytics

VideoAudio

Sensors

22



The Five Senses of AI

AI-driven systems and solutions are enabling a fundamental shift in the way we interact with machines and broadening the scope of tasks amenable to automation. Just as we humans use the five primary sensory functions (vision, hearing, taste, smell, and touch) to interact with our environment, an AI-driven system can be thought of as having five primary senses, enabling the system to listen/talk, see, analyze/act and remember (knowledge). We refer to the technologies associated with these five AI senses as core AI components, as these are central building blocks on which further applications or use-cases can be developed. Any AI-driven platform or system can be thought of as a solution built on one or more of these core AI components.

Table 2: Core AI components, sample application areas and vendors

Core AI component DescriptionCorresponding AI ‘sensory’ function Application areas Sample vendors

Natural Language Processing / Speech recognition / text analytics / Natural Language Generation

The ability of computers to understand (and converse in) human languages. Although NLP techniques have existed in the past, recent advances in machine learning have dramatically improved the accuracy of NLP algorithms.

1. Talk2. Listen

Digital virtual agents, chatbots, speech biometrics, automated customer service agents

Amazon, Google, IBM, Microsoft, Nuance, Lexalytics, Basis Tech, Expert System, iFlytek

Computer Vision / Smart Vision Technologies

Computer Vision (or Smart Vision) technologies enable extraction of meaningful, actionable information by analyzing digital images/videos. Examples include machine vision in autonomous cars, image recognition, facial recognition, motion tracking, etc.

3. See

Autonomous cars, smart surveillance cameras, automated retail stock checks, industrial and consumer robots

Amazon, Baidu, Google, IBM, Microsoft, Clarifai, Allied Vision, Bosch, Nice Systems, IntelliVision, Cognex

Augmented / Prescriptive Analytics

Analyzing data to provide actionable insights to enhance decision making. Traditional analytic tools are often descriptive or diagnostic in nature, while AI-enabled analytic tools have predictive / prescriptive (recommending action in response to an event)capabilities. Broadly speaking, AI-enabled analytic tools are capable of operating on unstructured data (big data).

4. Analyze/Act

Automated threat intelligence, quality control, predictive machine maintenance, etc.

IBM, Oracle, SAP, MicroStrategy, Tableau, Qlik, Tibco

Data stores, connectors, search engines, smart data discovery

Data integration as the ability to extract, integrate, structure, warehouse and migrate data. Smart data discovery enables the user to automatically locate, project and curate findings which are relevant to them (e.g. anomalies, correlations, pattern, clusters and forecasts), without the need to write algorithms. Encapsulates interactive data visualization, predictive analytics, pattern matching and machine learning to provide autonomous decision making support in a visual way.

5. Remember

Extracting, preparing, migrating, accessing, inferencing, representing data.

IBM, SAP, Talen, Oracle, Cambridge Semantics, Dassault (EXALEAD), Micro Focus (IDOL)

Source: J.P. Morgan, Gartner, Capgemini.

1/2 - ‘Talk/listen’: Conversational Platforms on the rise

One can now order a cab, book a table at a restaurant, and make payments using messaging platforms such as Line, WeChat, and Facebook's Messenger. Similarly, Virtual Personal Assistants (VPAs) such as Apple's Siri, Google Assistant, Microsoft’s Cortana, Amazon’s Alexa, or Baidu’s Duer allow users to look for information and execute commands at the user's request – using voice-based inputs. On the enterprise front, chatbots are not only being deployed to service customer queries but are also being used to enhance productivity and improve efficiency. These examples represent a paradigm shift in the way we interact with devices / platforms, driving new business models centered on conversations (chats/texts, or

23



voice). Advances in speech recognition and Natural Language Processing (both enabled by machine learning) are driving this shift.

Any conversational agent (be it a chatbot or a VPA) has the following key components: a user interface to accept inputs (voice commands or text), NLP (Natural Language Processing)/speech recognition element (to understand user-provided inputs), dialogue management (which helps provide context for the conversation), and back-end infrastructure (which connects the bots/VPAs to different applications/services).

Figure 8: Framework of a typical conversational commerce application

Source: Adapted from Gartner, J.P. Morgan. API = Application Programming Interface.

Messaging services and VPA providers are opening their platforms to third-party developers to expand the diversity of services / applications that can be provided. As an example, Messenger is now host to more than 50,000 chatbots since Facebook opened the platform to bot developers. Similarly, Alexa boasts of over 25,000 skills, thanks to Alexa Skills Kit which enables developers to build new skills (or applications) for Alexa. For product/service providers, conversational platforms provide an opportunity to maintain a deeper level of engagement with customers (via personalized conversations), helping drive loyalty and mind-share. This point is corroborated by a recent survey from Nuance (VPA provider), which shows that a majority of customers (89%) prefer a conversational interaction with a virtual agent when it comes to customer service / interacting with a business (vs. searching through Web pages or a mobile app on their own).

Apart from B2C (Business to Consumer) applications, enterprises are boosting productivity by deploying conversational bots across several functions such as vendor payments/invoicing, internal HR functions (training/onboarding/query-handling), search, etc. Traditional enterprise software vendors (such as SAP and Oracle) are also developing chatbot strategies to address this paradigm shift.

As per Gartner, by 2021, >50% of enterprises will spend more per annum on bots and chatbot creation than traditional mobile app development, signaling the high growth nature of conversational platforms.

Other notable enterprise chatbot / conversational AI platform vendors include Microsoft, IPSoft, Nuance, Slack, VoiceBox, MindMeld, Next IT, Aspect, Kore, etc.

http://www.nuance.com/company/news-room/press-releases/Opus-Intelligent-Assistants-and-Authentication-Conference-2016.docx

24



3 – ‘See’: Computer Vision enabling a host of next-gen applications

Computer Vision (or Smart Vision) is the ability of machines to extract meaningful and actionable information from images or video streams. Video and image traffic comprise the biggest portion of the growth in ‘unstructured’ data (driven by the proliferation of cameras). This introduces challenges in both manual monitoring of video/image feeds as well as tagging/indexing data for further analysis. Driven by advances in machine learning techniques (and improvements in camera technologies and computing power), computer vision is seeing brisk adoption across enterprises for tasks such as identifying objects, people, facial expressions, monitoring activity, and surveillance.

Figure 9: Identification and classification of static/dynamic objects using machine learning

Source: Nvidia.

Figure 10: Tally – Retail shelf auditing and analytics robot

Source: Simbe Robotics.

Figure 11: Urban mapping solutions - for Smart Cities

Source: Hexagon Leica RealCity Urban Mapping Solution

Figure 12: LiDAR read for terrain, construction, flood analysis, etc.

Source: Hexagon Leica Geosystems; ALS80 Airborne LiDAR Sensor.

Computer Vision technologies are also driving a wave of innovation in the field of robotics, extending its use-cases across several industry verticals. Examples of some industry use-cases of Computer Vision include: healthcare (image analytics for early disease detection & classification), retail (automated stock checks, facial recognition to enhance shopping experience/personalized advertising), automotive (advanced driver assistance systems), banking and insurance (ATM fraud control, identity

25



verification), manufacturing (production line monitoring), and security (smart surveillance).

Computer Vision APIs (Application Programming Interface – essentially a set of tools on which Computer Vision applications can be built) are offered by all the leading cloud service providers (Amazon Rekognition, Google Cloud Vision API, Microsoft Azure Cognitive Services, Watson Visual Recognition) and certain specialized vendors such as Clarifai and Imagga. In addition, there are several vertical-specific vendors that offer tailored computer vision technologies – we summarize some representative vendors in the table below.

Table 3: Representation computer vision software vendors across different industry verticals

Industry vertical Representative Computer Vision software vendors

RetailCatchoom, Clarifai, Syte.ai, ViSenze, VisionLabs, FLIR, ShopperTrak, Johnson Controls, Amazon Go, Trax technologies, Vispera

Manufacturing Cognex, Siemens, SmartFactory, Isra VisionMedia Clarifai, IBM, Metaliquid, VuDigitalHealthcare Arterys, Google DeepMind, IBMSecurity / Surveillance Bosch, Qognify, CrowdVision

Source: Gartner, J.P. Morgan.

4 - ‘Analyze and Act’: Ushering in a golden age for data science

Put simply, data science entails translating raw data into actionable insights. This sensory function can be considered as the central element of any AI platform. Enterprises are deploying AI-driven tools to augment decision-making (corroborated by findings from Gartner survey as highlighted in Figure 14) – these tools go beyond the descriptive and diagnostic aspects of traditional analytic tools and address predictive and prescriptive aspects needed for modern analytics. Gartner estimates that by 2020, predictive and prescriptive analytics will attract 40% of enterprises’ new investment in business intelligence and analytics. Indeed, enterprises are now spending more incremental dollars on such AI-driven data science platforms compared to the legacy traditional BI (Business Intelligence) platforms (Figure 14). It is worth mentioning that not all analytic tools / techniques that make up a data science platform are driven by machine learning (other analytic techniques include graph algorithms, knowledge extraction, and reasoning systems).

Figure 13: Types of AI applications that organizations have deployed or are planning to deploy

Source: Adapted from Gartner. Results reflect response to the question: What type of artificial intelligence initiatives is your

organization investigating or developing, or has your organization deployed or is planning to deploy?

11%

18%

18%

40%

64%

74%

0% 10% 20% 30% 40% 50% 60% 70% 80%

Other

Self-learning mechanical robotics

Embedded AI in products

Virtual personal assistant/chatbots

Process automation

Solutions for decision making / recommendation

26



Figure 14: Market size comparison: Traditional BI platforms vs. Data science platforms$ (bn)

Source: Gartner.

While some enterprises are deploying resources to develop machine learning-driven data science capabilities internally (wherever these projects are deemed viable, valuable, and vital), most data science delivery models support both on-premises and cloud-based deployments (given that most large organizations have data residing both on-premises and in the cloud).

Some of the leading data science platform vendors include players such as IBM, SAS, SAP, Knime, RapidMiner, Microsoft, Dataiku, Domino Data Lab, Alteryx, Fico, and Teradata. Gartner estimates data science platform market to grow from $2bn in 2016 to $5bn by 2021, growing at 15% CAGR over this period.

5 - Remember (Knowledge): Data Discovery and Integration

Data integration and data discovery are important features of an AI’s ability to function. These correspond to a human’s ability to remember, store and also retrieve information. Without access to memory, and without the ability to retrieve information, an AI's ability to locate, process and act on information is restricted. Understanding data relationships, data structures and storing these relationships, just as the brain encodes relevant information into synapses during memory formation, is an iterative and dynamic process.

Data integration tools play an important role in the extraction, warehousing, population, migration and management of data before machine learning can be effectively applied. This ultimately supports the construction and implementation of access to data as well as supporting the delivery infrastructure for a range of data scenarios. According to Gartner, leading data integration providers include: Informatica, IBM, SAP, Talen and Oracle among others.

Data discovery tools can be split into four categories: Visual Data Discovery, Smart Data Discovery, Search-Based Data Discovery and Self-Service Data Preparation.

Visual Data Discovery is based on an architectural structure that blends data from a range of sources into an in-memory store, which is linked to an interactive visualization layer.

Smart Data Discovery enables the user to automatically locate, project and curate findings which are relevant to them (e.g. anomalies, correlations, patterns, clusters and forecasts), without the need to write algorithms or build models. Smart data discovery essentially encapsulates interactive data visualization,

7.3 6.6

6.3 6.2 6.1 5.9 5.7 5.4

2.0 2.2 2.4 3.0

3.4 3.9

4.4 4.8

-

2.0

4.0

6.0

8.0

2014 2015 2016 2017E 2018E 2019E 2020E 2021E

Traditional BI platforms Data Science Platforms

27



predictive analytics, pattern matching and machine learning to provide autonomous decision making support.

Search-based Data Discovery maps both structured and unstructured sources of data into a categorized architecture of dimensions, from which users can explore through a search interface.

Self-service data preparation tools are often combined with visual data capabilities to support (through profiling, structuring and enriching) the data discovery process with the ability to integrate data.

Figure 15: Smart data discovery flow

Source: Gartner

The multiple styles of data discovery should soon converge as their unique features become requirements for all data discovery, as per Gartner. IBM’s Watson embeds smart data discovery into its analytics capability. Cambridge Semantics solution, Anzo Smart Data Discovery, offers dynamic access to both structured andunstructured data, and visualizes this through an in-memory Graph Query Engine.

Figure 16: Forecasts: Data Discovery Categories ($m)

Source: Gartner (April 2017)

- Algorithms identify schemas and profile data- Recommendations for data quality enhancement- Capable of use on multi-structured data

- Natural-language search- Algorithms leveraged to identify patterns in data- Relay findings to user in their language- Accessible to the everyday data user

- Natural language is able to explain results to user- Visualisation support, and collaboration

Prepare Data Locate patterns in the data Represent and share findings

1,465

451277

642

3,019

1,858

677

1,498

0

500

1,000

1,500

2,000

2,500

3,000

3,500

Visual Data Discovery Smart Data Discovery Search-Based Data Discovery Self-Service Data Preparation2016 2021

28



The AI application ecosystem

While the gamut of available AI applications is enormous, it is useful to think of the entire application ecosystem as one that is built on core AI components such as machine learning, computer vision, natural language processing / generation, augmented analytics, etc. AI software platforms are built on one or more of these components depending on the targeted application.

Broadly speaking, AI applications can be categorized into three key segments –Product, Process, and Insight.

Figure 17: Simplified AI application ecosystem

Source: J.P. Morgan. *Not an exhaustive list.

1. Product applications: Incorporating AI in a product or a service to provide end-customer benefits. Examples include: Amazon Echo (a wireless speaker equipped with Alexa, a virtual personal assistant), Netflix's recommendation engine, Google’s search engine, iPhones (equipped with Siri), autonomous cars (and smart robots ingeneral), etc. From a solution-provider's perspective, offering AI embedded in a product platform or via a cloud-based service not only enhances the value of the underlying offering but also creates new business opportunities that can generate high-margin & recurring revenue streams (Software-driven revenue streams are much more recurring in nature (pay-as-you-use model or subscription-based) and carry higher margins vs. hardware-based revenue streams).

Further, traditional hardware-focused companies are now facing stiff margin pressure driven by the rise of low-cost companies based in EMs such as China, the high cost of innovation (such as investments in truly flexible display), and uncertainty around end-demand. Against this backdrop, conventional hardware companies are racing ahead with investments in AI-related applications (either driven by in-house development or acquisitions of tech start-ups – Samsung’s Viv acquisition is one such recent example).

The segmentation provided here

is based on a Deloitte study of 100+ organizations across 17

industry sectors that

implemented / piloted applications enabled by AI

29



2. Process applications: Incorporating AI into a business's process flow to improve productivity or automate tasks. Examples include deploying AI-enabled systems for fraud detection, processing credit card applications, customer service (via chatbots / virtual conversational agents), surveillance, warehouse management, automated investment advisors, etc.

3. Insight applications: Aiding operational and strategic decision making based on insights generated by AI-enabled systems. Examples include customizing sales & marketing efforts based on insights gleaned from prior customer buying behavior, automated threat intelligence, predictive maintenance (to reduce operational downtime) enabled by AI-driven systems, etc. This remains one of the top AI application initiatives undertaken by enterprises.

30



What does AI bring to the Enterprise?

Artificial Intelligence is a buzzword in every boardroom, and enterprises want to use AI for multiple purposes; mainly:

Drive sales and customer engagement. AI can improve the overall customer experience in a multi-channel world. Applications include recommendation systems, virtual assistants, chatbots and voice bots. AI-managed marketing platforms can automate digital marketing and target high-value customers (for example, when launching a new product, AI can identify characteristics of previous high-value customers and the products purchased, plus feedback or other data, to target the highest probability customers for new products). AI assistants or agents can also handle higher volumes of customer service issues (especially repetitive or routine tasks), thus improving customer satisfaction and supporting customer intimacy overall.

Operational efficiencies. Within the organization, AI functions are enhancingquality control, predictive maintenance and prescriptive responses. Efficiencies can range from reducing operational costs and churn and to increasing legal/regulatory compliance and improving machine and process performance.

Enhancing products with embedded AI. Incorporating AI into a product or service can provide end-customer benefits.

Generating new insights, and enabling new business models. Better data analysis is allowing companies to think differently and often more creatively. Employees are spending less time in routines, and more time thinking of new products, new go-to-market strategies and engaging with and understanding customers at a higher level.

Is AI a competitive advantage?

We believe that earlier adopters of AI functionality will see significant benefits, which also speaks to the current urgency in the market for enterprises and governments to consider their approach and strategy around AI.

Whereas last year most organizations were testing proof of concepts in AI, 2017 has seen more AI applications moving into production. It is still early stage though, and we expect to see continued enhancements and fine tuning for these applications to find their full value. Today many AI applications are automating existing repetitive tasks, but we see the future as very much about whole new processes or even business models being created because of the capabilities of AI.

Not only does more automation free up time/effort for businesses to spend on the more creative aspects of its business, but the insights gleaned from AI can also help identify new opportunities, new products, new customers and new channels to market. Over the longer term, AI will be embedded into numerous applications and computer functions and nearly all enterprises will see some benefits, but we would argue that the early adopters and fast followers will take more market share, while the late adopters will implement AI just to keep up and non-adopters become far less competitive.

AI is being quickly embedded into numerous applications and computer functions. Early adopters should see competitive advantages.

AI will lead to whole new business models.

31



How are enterprises thinking about AI? Build, buy or outsource?

As with most IT solutions, enterprises typically have the choice of several ways to obtain AI capabilities: build, buy or outsource. The method will be dependent upon several functions: internal capabilities (the biggest challenge today being the availability/cost of data science skills), availability and pricing of packaged solutions currently available, urgency and desired time-to-market, amount of agility and customization needed, perceived uniqueness/differentiation of the solution envisioned, and other factors in ROI calculations.

Build

Organizations are likely to build their own AI solutions if they believe it will create significant differentiation for the business, few off-the-shelf alternatives exist in the marketplace, they need the higher level of agility and control, and the company has existing internal IT and data science resources to utilize for the project. We think mainly larger enterprises will pursue the build option, and some smaller companies which have few other options than to build.

Buy

Over the mid-term, we believe the bulk of organizations will buy existing packaged applications for AI. The cost and time-to-market are more favorable; however,customization (and differentiation) will be lower. Given the early stage of the market, only a handful of AI solutions are truly available off-the-shelf today, but we believe software vendors and start-ups are working quickly to fill this gap. In addition, some organizations will have no choice but to purchase AI functionality because it may be tied to proprietary data owned by a third-party, or perhaps the computing power is being rented from a cloud provider (which requires the use of its own AI tools).

Outsource

Outsourcing is typically used by organizations which need a more bespoke solution, but do not have their own IT resources or appropriate data science skills sets in-house to build the AI solution. Accenture, Capgemini, Atos and other IT services vendors are keen to engage their customers on bespoke projects (and for that matter, to implement packaged solutions, too, where available).

Enterprise AI adoption still in its infancy; growing digitization and awareness to spur growth

Contrary to the amount of media coverage surrounding AI and its associated benefits, surveys from several research firms (such as Gartner, McKinsey, etc.) suggest that enterprise AI adoption is still in its infancy. Many organizations are in the knowledge gathering / pilot phase of AI adoption, with very few actually using AI in 'live' enterprise applications / products.

Figure 18: Current stage of enterprise AI adoption

Source: Gartner survey. Responses: What is the current stage of artificial intelligence solutions adoption within your organization?

59%

25%

6% 6% 4%

0%

10%

20%

30%

40%

50%

60%

70%

Knowledge gathering/ investigating /

developing strategy

Piloting Implementing Deployed / in usetoday

Have plans to deployby 2018

32



Some of the key challenges impeding the adoption of AI across enterprises include a shortage of available ‘AI-talent’ (data scientists, machine learning experts, etc.), difficulty in porting existing infrastructure to AI-ready infrastructure, challenges in quantifying value derived from deploying AI solutions, and lack of a strategic directive from top management (C-level management, business unit heads, etc.). In addition, there are concerns surrounding the ‘black-box’ nature of AI algorithms (the way ML algorithms interpret data to produce an outcome or a set of outcomes is often not known), which hinder its adoption (especially in highly regulated industries, such as financial services).

Figure 19: Key challenges for enterprise AI adoption

Source: Gartner survey results, November 2017. Survey results in response to the question: What are the top three challenges to the

adoption of artificial intelligence within your organization?

Having said this, we believe that growing enterprise digitization, increasing awareness of AI solutions and their direct/indirect benefits, and the rising sophistication of AI algorithms will help drive enterprise AI adoption. Further, as the market matures, we expect consolidation across the AI software vendor landscape and this should help enterprises select vendors more effectively for AI-related initiatives, aiding adoption of AI. While adoption may be a point of discussion at this stage, we believe that AI will be ubiquitous across all the industry verticals and all digitally mature enterprises in the future.

5

11

13

18

23

27

30

35

35

37

54

- 10 20 30 40 50 60

Other

Understanding what AI is

Governance issues

Potential risks / liabilities

Determining how to measure value from using AI

Complexity of integrating AI with existing infrastructure

Security or privacy concerns

Identifying use cases for AI

Funding for AI initiatives

Defining AI strategy

Lack of necessary staff skills

33



Leading Enterprise AI Use Cases by Industry

Apart from large technology firms (wherein AI is embedded in the core product / service offering), industry verticals leading AI adoption include Communication, Media, and Services (customer analytics, video analytics), Banking & Securities(Fraud analytics, automated threat intelligence and prevention, automated advisors, and discovery of investment opportunities), Healthcare (automated diagnostics, real-time monitoring / pre-emptive warning systems, and drug discovery), Retail (Supply chain management, warehouse automation, customer analytics, and merchandizing), and Manufacturing (quality management, yield improvement, and predictive maintenance). This comes as no surprise as these industries share the common traits of relatively high digital maturity, access to troves of data, a desire to glean patterns from historical events as a guide for future decisions, and having a large portion of tasks that are amenable to automation.

We expect adoption of AI to proliferate across other industry verticals as well, over the next three to five years – Transportation & logistics, Public safety & Surveillance (smart cameras), and Education (customized training for individuals) are some of the key industry verticals that we think likely to witness rapid AI adoption in the coming years.

Table 4: Sample AI Use Cases Across Different Industry Verticals

Industry vertical

Industry as a % of total global IT

spend* Sample AI Use Cases

Banking & Securities 19%

Automated trading & investment discovery, trading strategies, robo-advisors, voice-based commerce, customer behavior analysis, chatbots for customer services, identity verification, fraud detection.

Government 17%Smart surveillance, threat detection, Smart Cities and Utilities, AI-enhanced and personalized education and training, chatbots for info distribution and citizen engagement.

Manufacturing & Natural Res. 17%Predictive maintenance, machine learning driven insights for yield improvement, optimization.

Comms, Media & Services 16%Customer analytics, forecasting and customer demand trends, video analytics, computer vision interactivity (e.g. in video games and other immersive media).

Retail 7%

Customer analytics, forecasting, anticipating demand trends, reducing revenue churn, supply chain management, warehouse automation, chatbots for customer services, conversational commerce.

Insurance 7%

Claims management and fraud detection, analyzing customer behavior and reducing revenue churn, automated underwriting, pricing, conversational platforms for customer services, complying with regulations, trading strategies.

Utilities 5%

Enhanced supply-demand management based on AI-driven analytics, predictive maintenance, dynamic pricing based on consumption analytics (provided by smart meters, for example), chatbots for customer service.

Healthcare Providers 4%Diagnostics, image analytics for early disease detection, drug discovery, patient monitoring (pre-emptive warning systems), personalized medicine and treatment.

Transportation 4%Self-driving vehicles, Advanced driver assistance systems, personalized content delivery / productivity enhancement tools used by providers of transportation services

Education 2%Customized / adaptive learning programs, skill upgrade programs based on real-time insights gleaned from job market trends.

Wholesale Trade 2%Warehouse automation, inventory management based on insights gleaned from demand analytics, autonomous delivery.

Source: J.P. Morgan, *Gartner, Capgemini.

34



The AI-spending opportunity

It is important to note that AI is not a product offering in itself – it is essentially a model (a set of software tools / algorithms) that helps identify patterns and associations in large amounts of data. Further, as we have discussed previously, AI can either be embedded in hardware products or software platforms or be deployedas a component in a larger software process implementation. Hence, rather than focusing on AI, it is helpful to talk about the markets / applications enabled by AI, which can then be assigned a dollar value. Current AI market forecasts vary widely depending on how one defines the scope of applications / end-markets considered for market sizing. However, the common theme underpinning all these forecasts is the high-growth nature of AI-enabled markets.

AI-focused spending to witness sharp growth in the coming years; Software / Services to take the biggest slice

IDC forecasts spending on AI-focused hardware, software, and services to reach $58bn by 2021, up from ~$12bn in 2017, making this one of the fastest-growing technology segments (growing at nearly 50% ’17-’21 CAGR). The components of this AI-focused spend include:

AI applications - applications that learn, discover, and make recommendations / predictions or core AI components.

AI software platforms - tools built on core AI components that enable AI-driven use cases.

AI-related IT & business services – for example: consulting / implementation services provided to an enterprise for deploying AI-related technologies.

AI-dedicated server and storage spending (hardware).

It is important to note that this spending figure excludes internal enterprise R&D investments & AI deployments, capital spent on talent acquisition, M&A, and private financing. These excluded components are meaningful in size. As noted, McKinsey highlighted that roughly $26-39bn was spent on AI-related technologies in 2016.

Figure 20: Global AI-focused spending* ($, bn)

Source: AI-spending estimates from IDC. *Includes AI-focused spending on hardware, software (applications + software platforms), and services (IT consulting & system implementation).

5.3 7.5 12.0

57.6

-

10.0

20.0

30.0

40.0

50.0

60.0

70.0

2015 2016 2017E 2021E

AI-focused spending*

48% CAGR

35



To put spending size into perspective, AI growth projections of ~50% CAGR through 2020/21 are more than twice the growth rate of other high growth tech sub-sectors (such as Big Data 23% and Cloud 20%). By 2020, AI-focused spending will be about the same size as the security software market.

Figure 21: AI market growth compared to other high-growth technology segments

Source: Bloomberg, IDC, J.P. Morgan.

As highlighted in the figure below, software (applications and software platforms) accounts for ~50% of the total AI-focused spending. Services (IT consulting and system implementation) represent the second largest spending category and IDC expects spending in AI-focused services to continue growing at 50%+ CAGR over the next four years. On the other hand, hardware (compute & storage infrastructure) is expected to remain the smallest portion of total AI-focused spending. Of note, we expect growth in hardware spending to continue to be impacted by component price deflation over the long term (i.e. decrease in per unit storage and computing costs). We believe that the AI-spending opportunity opens up meaningful avenues of revenue growth for enterprise software and IT services vendors who take the initial lead in addressing enterprise AI needs.

Figure 22: Components of AI-focused spending (2017E)$ (bn)

Source: Adapted from IDC.

AI-focused spending, 50%Big data, 23%

Total cloud, 20%

Security, 8%

0

50

100

150

200

250

0% 10% 20% 30% 40% 50% 60%

Approximate 2016-2020 CAGR

App

roxi

mat

e 20

20 m

arke

t siz

e ($

, bn)

3.7

2.3

3.7

2.3

12.0

-

2.0

4.0

6.0

8.0

10.0

12.0

14.0

Softwareapplications

Software platforms Services Hardware Total

Software spending comprises bulk

of the total AI-focused spending

36



AI-focused spending still a small portion of overall enterprise IT spending

The enterprise AI-focused spending, though witnessing very strong growth, accounts for a very small portion of the overall enterprise IT spending, growing from ~0.3% in 2016 to ~1.8% by 2021. This makes sense, as the bulk of the enterprise IT spending is dedicated toward ‘run-the-business’ traditional applications and digital transformations (cloud, traditional analytics, etc.) that are not AI-driven. However, we believe that the AI-portion of enterprise IT spending will expand in the future as AI-led innovation transforms the way enterprises operate, driving an increasing number of applications (a ‘transformative AI application’ today is tomorrow's ‘run-the-business’ application).

Figure 23: AI-focused spending as % of overall enterprise IT spending

Source: AI-spending estimates adapted IDC. *Includes AI-focused spending on hardware, software, and services. **Enterprise IT

spend data sourced from Gartner.

Will AI boost overall IT spending?

We believe AI along with many other aspects of digital transformation (like cloud, IoT, big data, blockchain, etc.) are boosting IT spending by corporates, governments and even individuals. However, this is likely already captured in estimates for IT spending growth. Gartner predicts Global Software spending will grow at an 8.3% (constant currency) CAGR through 2021; which is around 2.5x global GDP growth forecasts (JPMe 3.2%). IT Services are also expected to grow faster than GDP at 4.6% CAGR. Even including slower growth subsectors, global IT spending is forecast at 3.0% CAGR to $4.5 trillion by 2021. See Table 5 below.

Table 5: Global IT Spending Forecast by Gartner (constant currency)

(USD billions)2017 IT spend

2017 Growth

2018 IT Spend

2018 Growth

2021 IT Spend

CAGR (%) cc 2017-2021

Software 392 8.5% 425 8.5% 539 8.3%IT Services 1,038 4.3% 1,083 4.4% 1,243 4.6%

Devices 764 5.4% 793 3.8% 817 1.7%Data Center Systems 194 1.8% 196 0.9% 197 0.4%Communication Services 1,593 0.8% 1,613 1.3% 1,686 1.4%Total Global IT Spending 3,980 3.3% 4,110 3.3% 4,481 3.0%

Source: Gartner, October 2017.

0.2%0.3%

0.4%

1.8%

0.0%

0.5%

1.0%

1.5%

2.0%

2015 2016 2017E 2021E

AI-focused spending as % of total enterprise IT spending

37



We would argue, however, that AI functionality (direct or embedded) will take market share within the IT budget, and therefore vendors that are ahead of the curve in embedding AI will benefit. The AI platform vendors have taken an early competitive lead; however, we believe there is ample opportunity both for established traditional IT enterprise vendors to embed AI into their core offerings and for a variety of new, emerging players to offer creative solutions(from niche and vertical-specific to even broad-scope industry-disrupting ideas). Think how Salesforce promoted cloud well before the incumbents, or how Amazon is changing retail, or Uber is disrupting transportation. We expect AI will stimulate similar opportunities for innovative industry-disruptors.

Furthermore, we believe that the traditional enterprise IT budget (from the CIO's office) does not capture the entire portion of corporate spend on IT. Increasingly the lines of business are directly purchasing IT in the form of vertical applications. The marketing department and huge budgets around digital marketing are probably the most obvious example, but payments, customer service and areas of automation are other examples. In fact, one could argue that nearly all companies are becoming IT companies at their core.

Leaders in digital transformation will also be leaders in adoption of AI

Across industries, we believe those corporates addressing their own digital transformation now will also be the leaders in the adoption of AI. Certainly, the puredigital companies have already shown a lead in using AI (Google, Amazon, Uber, Netflix, etc.), and the faster other industries digitize, the faster they can take advantage of the insights and efficiencies of AI.

AI functionality to take market share.

IT spend comes from the central IT budget as well as from Lines of Business.

AI and Digital Transformation will go hand-in-hand.

38



Mapping the Vendor Landscape

In the section, we will discuss a variety of vendors utilizing AI and ML; however, we acknowledge that the field is so wide and so quickly advancing, that it is impossible to do it full justice. There are already over 2,000 companies calling themselves AI vendors, and Gartner estimates that by 2020 nearly every new software product and services will feature artificial intelligence.

Within our global technology/internet stock coverage, we believe the following companies are best positioned to benefit from Artificial Intelligence.

Figure 24: J.P. Morgan coverage companies which stand to benefit from the shift to AI

Region Technology SectorCompanies best prepared for the shift to AI

AI function / use Analyst

Europe

European Software SAP, Dassault Systèmes, HexagonAI platforms, AI-based analytics, object/pattern recognition, predictive maintenance, smart cities.

Stacy Pollard

European IT Services Capgemini, AtosAdvisory & education, creation and solution design, deployment and integration.

Stacy Pollard

European Industrial Tech SiemensMindSphere cloud IoT, digital factory, robotics, industrial optimization.

Andreas Willi

America

US Enterprise Software Salesforce, Microsoft, AdobeAI platform, predictive analytics, automation bots, speech recognition and NLP, image recognition.

Mark Murphy / Sterling Auty

US IT Services IBM, Accenture, CognizantAI platform, analytics/data discovery, chatbots, natural language, computer vision, advisory & implementation.

Tien-Tsin Huang

US Applied & Emerging Tech NICE, Cognex Robots as digital assistants, real-time customer data to chatbots, machine vision.

Paul Coster

US Internet Google, Amazon, FacebookSearch, cloud, home assistants, autonomous vehicles, photos, news feed…many others.

Doug Anmuth

Asia Asian Internet / software Baidu, Tencent, AlibabaAI Platforms, autonomous driving, conversational (NLG, voice), computer vision, customer analytics, etc.

Alex Yao / Gokul Hariharan


Also see Appendix 2, page 68 for more company profiles.

AI: Is this a winner-takes-all market?

Not at all. AI adoption is evolutionary, and although vendors like Google, Facebook, Baidu, Apple, Microsoft and Amazon may have taken early advantage of their large data sets and established platforms, the large enterprise software vendors (like SAP, IBM, Oracle, Salesforce, Sage, etc.) are gradually embedding AI into multipleaspects of their offerings. In addition, we expect certain revolutionary aspects of AI to bring to market new and/or specialist vendors (hundreds are already in the marketplace, and we expect to see many more as individual AI functions mature).

Existing software vendors are gradually embedding AI functions into various aspects of their technology. Large enterprise vendors like SAP, IBM, Microsoft, Oracle, Salesforce, Sage and others already have a handful of AI offerings, and these look sure to multiply over time. We believe they are well positioned to promote AI functionality within enterprises because: they are known brands, which already understand and can contextualize the enterprise-generated data, and they can link automation back into the workflow.

Business Intelligence and Data Analytics will evolve to become increasingly AI-driven, and indeed, we believe that AI is transforming the BI and Analytics market. We expect to see the gradual decline of traditional BI, in favor of modern BI and Analytics platforms and functions which include predictive and prescriptive analytics.

Evolve with AI capabilities, or risk being left behind

39



Traditional CRM vendors are at risk if they are not evolving with AI (and advanced customer understanding and engagement). The same goes for nearly every category of enterprise software, from BI and CRM to supply chain and ERP (for example, automated expense matching).

AI platforms (including AI PaaS) have around 20% (Figure 22) of AI spending share today and are frequently the building blocks for companies designing their own AI applications.

Industry-specific and function-specific AI are likely to be of great interest to clients looking for productized or nearly out-of-the box solutions. The early days of most technology cycles are characterized by broad innovation and fragmentation – which is where we are today (as opposed to consolidation and vendor dominance by a small number of players, which happens later cycle).

Build, buy or partner?

Apart from in-house development of AI capabilities, vendors across different industry verticals have also been active in acquisition of AI-related IP and talent(often at a meaningful premium) in order to complement in-house efforts and accelerate the time-to-market for AI related offerings. This is especially true for markets that are witnessing rapid disruption driven by AI, such as technology, automotive, and healthcare. As per CBInsights, the 10 largest tech companies in the US have invested in 80 AI startups and acquired 50 AI companies in the last five years. AI-related deal activity is at an all-time high, with an estimated $7bn in annual funding in 2017 (up from ~$5bn in 2016). Not surprisingly, Google, Apple, Intel, Microsoft, Facebook, salesforce.com, and Twitter feature among the top AI acquirers in the recent past (see Appendix 3 for some recent AI acquisitions). Furthermore, we expect to see numerous partnerships around artificial intelligence, with different firms leveraging their relative competitive advantages. For example Amazon and Microsoft recently partnered around Gluon (an open-source software tool for developers to build AI and ML apps), and SAP and NVIDIA are collaborating using NVIDIA’s DGX-1 (an integrated hardware and software supercomputer) to offer machine learning in SAP’s applications (such as SAP Brand Impact, Accounts Payable, and Customer Satisfaction).

Figure 25: Artificial Intelligence: Annual Funding History

Source: Adapted from CBInsights.

0

100

200

300

400

500

600

700

800

0

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

2012 2013 2014 2015 2016

Disclosed Funding ($m) Number of Deals

40



Categorizing AI vendors

We have tried to categorize vendors into several functional areas of AI (while noting that many vendors cross functions and thus may fit into multiple categories). We also acknowledge that this is not a comprehensive list, but rather a substantial sampling to make investors aware of the many vendors and types of function and application that are involved in the AI arena.

Table 6: List of AI Vendors, Products and Software

Category Description / Application Sample Vendors

AI Platform-as-a Service Still in early adoption phase; with top three vendors dominating -tough to challenge the hegemony.

Amazon Web Services (c.40% mkt share), Microsoft Azure Machine Learning, and Google, clear public cloud leaders. Others include IBM, Alibaba Aliyun, Oracle, Baidu, etc.

AI Platforms and deep learning frameworks/libraries

AI platforms provide users with a toolkit to build intelligent applications. The platforms may combine decision-making algorithms with data.

TensorFlow from Google, IBM PowerAI IBM Watson Platform, Google Cloud Prediction API, AWS/MS Gluon, Microsoft CNTK, Amazon ML, Infosys Mana, Wipro HOLMES, Rainbird, Ayasdi

Traditional enterprise software vendors

These vendors seem likely to embed AI into multiple aspects of their software offerings.

SAP (with Leonardo), Microsoft, Oracle, Salesforce (and Einstein)Adobe Sensei and Cloak.

Data Analytics Vendors Business intelligence and analytics vendors are creating AI/ML tools to modernize their analytics solutions, especially on areas such as predictive / prescriptive analytics. Data integration is also an important component.

BI vendors: SAS, SAP, Oracle, IBM, Microsoft, Tableau, MathWorks, Qlik, Palantir, Fico, MicroStrategy, etc.Data integrators: Informatica (data integration, quality, and management), IBM InfoSphere, SAP, Talen, Oracle, SAS, Microsoft, Cisco, Denodo, etc.

Specialist AI / Vertical-focused / Niche players

Vertical-specific expertise will be a clear differentiator. We believe that significant numbers of niche and vertical-focused will come to market in coming years, and will attract relatively smaller, but loyal user groups. Clients will choose these vendors/products for their best-of-breed functionality, and likely off-the-shelf availability and quick implementation and ROI.

Clarifai (automated photo tagging), Creative Virtual and many other virtual customer assistants, Ravel Law (legal assistant), x.ai (bots that negotiate via email to find a time/place for a meeting), Dassault Systèmes (in CAD and also fleet mgmt.), Sage’s Pegg accounting bot, and numerous others.Uber and Waze for real-time traffic and routing.

Industrial applications, IoT+AI, IT/OT integration and robots

AI and IoT are symbiotic, because IoT provides the large volumes of data that AI technologies require, and without AI apps, IoT data is far less useful. These applications typically focus on operational optimization (quality control, predictive maintenance - leading to prescriptive responses from AI).

ThingWorx from PTC, Siemens MindSphere, SAP Leonardo, Oracle IoT Cloud Service, Mosaic from LTI, ABB Industrial IoT, GE Predix, Honeywell Industrial IoT, Schneider Electric, Hexagon, Software AG (Cumulocity IoT platform), Amazon IoT, MS Azure IoT Suite.

AI-enhanced customer-facing channels, Conversational platforms and Chatbots

These applications are focused on improving the overall customer experience – in a multichannel world, they aim to be the gateway to the customer (potentially disintermediating others).

Kore.ai (all-in-one chatbot PaaS), Openstream, NextIT, Aspect, Amelia from IPsoft,

IT Services IT Service vendors design and implement AI functions for enterprises, especially custom projects. Ultimately, IT Services vendors will democratize and facilitate the proliferation of AI knowledge and best practices.

Accenture, IBM, Capgemini, Atos, Cognizant, Tata, etc.

Hardware AI specialized compute / networking / storage Key AI chip vendors include Nvidia (GPUs), Intel (x86 platforms with accelerators), AMD (CPUs/GPUs), Xilinx (FPGAs), Broadcom (ASIC), etc. Some of the prominent AI chip startups include Graphcore, Cambricon, Wave Computing, Mythic, and Cerebras. Apart from compute, AI applications are also impacting the networking and storage elements of the datacenter, with the focus on high performance and low latency.

Source: J.P. Morgan; company data from many various sources.

41



AI Platform-as-a-Service

AIPaaS is still in the early adoption phase, but progressing quickly. The top three vendors (Amazon, Google and Microsoft) dominate today, and it may be tough to challenge the hegemony.

AI Platform-as-a-Service Vendors

Amazon Web Services (c.40% mkt share), Microsoft Azure Machine Learning, and Google are the clear public cloud leaders.

Others include IBM, Alibaba Aliyun, Oracle, Baidu, etc.

AI Platforms and deep learning frameworks/libraries

AI platforms provide users with a toolkit to build intelligent applications. The platforms may combine decision-making algorithms with data. Some require knowledge of complex AI structuring and coding, while others offer pre-built algorithms and simple workflows and can include drag-and-drop modeling and visual interfaces. IDC valued the AI software platform market at $1.6bn in 2016, going to $8bn by 2021.

Figure 26: Cognitive/AI Software Platforms market share, 2016 (mkt size $1.6bn)

Source: IDC.

Deep learning frameworks are symbolic math libraries, systems for building and training neural networks to identify patterns and correlations. These frameworks are mainly open source, and there are many iterations and layers.

AI Platforms and Framework Vendors

IBM PowerAI IBM Watson Platform and Analytics, IBM Caffe

Palantir data analysis (and visualization)

GoogleTensorFlow, Google Cloud Prediction API, API.AI (natural language interactions, acquired by Google)

Amazon: Amazon ML, AWS Machine Learning API, DSSTNE (Amazon’s Deep Scalable Spares Tensor Network Engine), AWS/MS Gluon

Microsoft Azure Machine Learning workbench and model management, CNTK (computational network toolkit), DMTK (MS Distributed Machine Learning Toolkit), Gluon (with AWS)

Apache Spark MLlib, Apache Mahout, Apache Singa, BigDL focused on ApacheSpark and only works on Intel chips.

IBM10% Palantir

4%Google

2% Digital Reasoning

1%IPsoft1%

Nuance Comms1%

CognitiveScale1%

Expert System1%

CustomerMatrix1%

Other78%

42



NVIDIA: DIGITS (NVIDIA Deep Learning GPU Training System), NCCL (NVIDIA Collective Communications Library), NVIDIA Caffe

Others: Digital reasoning, IPsoft, Nuance, CognitiveScale, Expert System, CustomerMatrix, Alibaba Aliyun, Baidu Paddle, Caffe2, Caffe OnSpark, Infosys Nia, Wipro HOLMES, Rainbird (visual user interface, natural language processing, analytics & insights), Premonition.ai - litigation database, Wit.ai Bot Engine, Ayasdi (big data analytics and modeling), Vital A.I – agent platform, Meya Bot Studio, MindMeld (conversational AI for voice/chat assistants), KAI conversational AI, Receptiviti (emotional intelligence, bots, marketing and engagement)Veles, H2O.ai, Neon, Wise.io, Deeplearning4j (released under Apache License 2.0), ND4J, Torch, Pytorch (open sourced by Facebook), DyNet, MxNet, OpenBLAS, Distributed Frameworks, Bazel, etc.

Software coding languages:

Typically open source, these are some of the programming languages most frequently used in the AI field.

Python, Java, Lisp, Prolog, C++, MATLAB, Scala, Julia, R, Haskell, AIML, among others.

Traditional enterprise software vendors

Traditional enterprise software vendors are embedding AI capabilities into multiple aspects of their software offerings. They are well-positioned to promote AI functionality within enterprises because: they are known brands, which already understand and can contextualize the enterprise-generated data, and they can link automation back into the workflow.

SAP (Leonardo)

Microsoft

IBM (Watson)

Oracle

Salesforce (Einstein)

Adobe Sensei and Cloak

AI-based features already include virtual assistants, bots, advanced analytics and numerous other specialist features. See individual vendor profiles for details.

Figure 27: Global Enterprise Software Market Share, 2016

Source: Gartner.

Microsoft16%

Oracle9%

IBM8%

SAP6%

Salesforce2%

Vmware2%

Adobe2%

Others55%

43



Data Analytics and BI Vendors

Data analytics is one of the most natural places to embed AI. Business intelligence and analytics vendors are creating AI/ML tools to modernize their analytics solutions, especially on areas such as predictive / prescriptive analytics. Traditional BI functions will see declining market share, and vendors who fail to modernize will quickly be left behind.

Figure 28: Largest Business Intelligence Software Vendors

Source: Gartner.

Within the BI & Analytics Software Market, Data Science Platforms are set to grow around 13% CAGR through 2021, while the overall BI Software Market should grow 8% (see Figure 29 below).

Figure 29: Data Science Platforms are growing faster than overall BI & Analytics SW

Source: Gartner.

SAP16%

Oracle12%

SAS11%

IBM9%

Microsoft9%

Tableau5%

MathWorks4%

Qlik4%

Palantir3%

Fico2%

MicroStrategy2%

Others23%

44



Data and its role in shaping the competition landscape

As a reminder, we refer to ML-driven intelligence as AI in this report. In this context, data is of paramount importance for AI algorithms and applications. By its very nature, ML is very data-hungry, as it entails fine tuning of several model parameters which require vast and diverse training datasets to yield an effective model output. An ML-driven algorithm is only as good as the data it is trained on. Without sufficient, clean, and diverse datasets, ML algorithms are of limited use (and in fact, using ML algorithms trained on ‘immature’ datasets can result in sub-optimal / meaningless output).

Given the importance of data in the emerging AI-driven economy, it is natural to assume that platform owners (such as Amazon, Google, Apple, Facebook, Microsoft, Baidu, Alibaba, etc.) have an edge compared to competition. We see increased evidence of these platform owners exerting influence in areas outside their primary domain of operation and we believe that data lies at the heart of this transformation. Leading platform owners benefit not only from the learning-curve effect driven by better data access (which translates to better AI capabilities) but also by positioning their businesses around technologies that leverage the value generated from this data.

Data storage (database) vendors are benefitting from the massive increase in data generation – in parallel to AI. Several top DB vendors include:

Oracle 12c and MySQL

Microsoft SQL Server

IBM DB2 and Informix

SAP HANA and Sybase

Teradata

Amazon SimpleDB and Dynamo DB

The Hadoop and Spark frameworks for big data analytics are also worth a mention. A few top vendors include: Cloudera, Hortonworks, Pivotal, and numerous others.

For a longer list and description of technology tools used in Big Data Analytics, we recommend a read of our colleagues' report: Big Data and AI Strategies by Marko Kolanovic and Rajesh Krishnamachari. This report dedicates around 80 pages to explaining the granularity of Machine Learning Methods, 20 pages to data technology solution vendors (including databases, ETL, Hadoop, infrastructure, management, security, and ML tools, and another 50 pages listing various big data sources by industry.

https://www.jpmm.com/research/content/GPS-2345119-0

45



Data integration tools play a role

We also note that data integration tools will play a role in the harvesting of big data for AI functions. While the data integration tool market is beyond our scope for this report, it is useful to be aware of some of its primary functions and vendors.

Data integration tasks

ETL – Extraction, Transformation and Loading.

Data migration – moving data to new platform or upgraded application.

Data acquisition for business intelligence, analytics and business warehousing -extracting data from operational and other systems for the purpose of analytics, predictive modeling in-memory data base management systems (definition, creation, querying, update and administration of databases), and data warehousing (central repositories of integrated data, often from multiple different sources).

Sharing data across multiple organizations, for example, through a supply chain or with customers and other business partners. Also, data integration tools for frequently need to manage enterprise data presiding in multiple locations, such as pivoting between cloud and on-premises.

IoT and enterprise digital transformation are boosting the need for data integration and analytics.

Leading Data Integration Vendors

Informatica (data integration, quality, and management), IBM, SAP, Talen, Oracle, SAS, Microsoft, Cisco, Denodo, Information Building, Attunity, Adeptia and others.

Behind the scenes, Hadoop (open-source software) supports the processing and storage of large data sets in distributed computing environments. Cloudera and HortonWorks are the main big data vendors, as well as AWS Elastic MapReduce Hadoop Distribution, Microsoft Hadoop Distribution, MapR Hadoop Distribution and IBM InfoSphere Insights.

Specialist AI / Vertical-focused / Niche players

Vertical-specific expertise will be a clear differentiator in our view. We believe that significant numbers of niche and vertical-focused will come to market in coming years, and will attract relatively small, but loyal user groups. We expect clients to choose these vendors/products for their best-of-breed functionality, and likely off-the-shelf availability and quick implementation and ROI. There are numerous vendors offering specialist applications, and we expect to see continued significant innovation by hundreds, if not thousands, of start-ups and existing AI vendors.

Some interesting vendors

Clarifai (automated photo tagging),

Creative Virtual and many other virtual customer assistants,

Ravel Law (legal assistant),

x.ai (bots that negotiate via email to find a time/place for a meeting),

Dassault Systèmes (in CAD and also fleet mgmt.),

46



Sage’s Pegg accounting bot, and numerous others.

Uber and Waze for real-time traffic and routing.

Industrial applications, IoT+AI, IT/OT integration and robots

AI and IoT are symbiotic, because IoT provides the large volumes of data that AI technologies require, and without AI apps, IoT data is far less useful. These applications typically focus on operational optimization (quality control, predictive maintenance - leading to prescriptive responses from AI). See also the separate section on AI Brings Life to IoT.

Sample vendors:

ThingWorx from PTC

Siemens MindSphere

SAP Leonardo

Oracle IoT Cloud Service

Mosaic from LTI

ABB Industrial IoT

GE Predix

Honeywell Industrial IoT

Schneider Electric

Hexagon

Software AG (Cumulocity IoT platform)

Amazon IoT

MS Azure IoT Suite

The IoT Platform vendor landscape is crowded, and as yet, there appears to be no clear – differentiated – market leader. The market is indeed still emergent and its evolution and leadership will depend on the ability vendors have to provide clear differentiation through a complete horizontal solution, specialist vertical market knowledge, and a clear ecosystem of partners. We have outlined below a non-exhaustive table of IoT platform providers.

47



Table 7: Sample IoT Platform Providers

IoT Platform Vendor Description/Components Vertical Markets Areas of Application Go-To-Market Partners (Examples)

Amazon - AWS IoT

Builds on its IaaS capability and extensive cloud infrastructure. Services in the stack: AWS message broker (or Device Gateway), AWS IoT rules engine, Device Shadows (digital twin implementer), Device Registry and IoT Console Services.

Automotive,Healthcare,Smart City

Asset Management, Predictive Maintenance, Personal Device Integration

Intel, Samsung – HardwareTwilio, Splunk – SoftwareAccenture, Deloitte – Services

GE - Predix

Built on a Pivotal foundation/Cloud foundry, supplemented via a microservice architecturalmethod. Leverages interconnectivity, ‘intelligent’ machines, as well as analytics for industrial markets. GE is aiming to roll out vertical specific applications.

Energy, Manufacturing, Transportation

Industrial Internet Applications, Asset Management, Operations Optimisation

Cisco, Intel – HardwareAT&T, Vodafone – CommunicationsApigee, Pivotal – SoftwareEquinix – InfrastructureGenpact, HCL Technologies - Services

IBM - Watson IoTPlatform

IBM has evolved from its Bluemix IoT Foundation (launched in 2014) into today’s Watson IoT portfolio and sees IoT value in the cognitive analytics. The stack begins with Bluemix IoT Zones, and leads onto various options, such as solutions for manufacturers (e.g. Product Line Engineering and Rhapsody, among others). Asset management solutions (e.g. Trirga & Maximo Asset Management), and cloud capabilities (e.g. SoftLayer - IaaS, IBM Watson -cognitive analytics).

Manufacturing,Automotive, Building Management

Asset Management, Manufacturing

Cisco, Arrow Electronics– HardwareOrange, Verizon – CommunicationsCloud Foundry, Cisco Jasper – SoftwareIBM GBS, HCL Technologies/Deloitte -Services

Microsoft - Azure IoT Suite

Built from an initial operating system and includes a range of technology components: Azure IoT Hub (designed specifically for IoT), Azure Stream Analytics, Azure Data Lake, Azure Machine Learning, Gateway SDK, Microsoft Dynamics & Power BI. Supports end-to-end model with the Azure Logic Apps integrating back-end applications (e.g. Oracle, SAP, Salesforce) with the Azure IoT Suite.

Manufacturing, Smart Buildings, Smart Transportation

Predictive Maintenance, Device Management, Remote Monitoring

Intel, Liberum – HardwareHadoop, Cisco Jasper – SoftwareAccenture, Avanade - Services

PTC - ThingWorx

PTC’s capability evolved through internal R&D and acquisitions (ThingWorx, ColdLight, Axeda, Kepware & Vuforia). Comprised of core modules: Foundation, Analytics, Utilities & Studio integrated into the ThingWorx Thing Model (digital twin) – the data store holding business logic for assets and data streams.

Manufacturing, Smart Cities,Supply Chain

Asset Management, Predictive Analytics, API management,

ARM, National Instruments – HardwareAeris, Verizon – CommunicationsGE, SAP – SoftwareAmazon, Microsoft – InfrastructureCSC, ITC Infotech - Services

SAP – SAP HANA IoT Foundation

The stack begins with Hana Cloud Platform, including Hana Integration Services and SAP Process Orchestration. IoT Client & IoT Intelligent Edge also supports the platform’s capabilities. SAP leverages IoT to enhance its core business (S/4 HANA, SAP ERP) applications and networks (Ariba) with IoT generated data & analytics. Applications are built to support digitalization of the enterprise (e.g. SAP Predictive Maintenance, SAP Networked Logistics Hub, SAP Asset Intelligence Network).

Manufacturing,Transportation,Logistics

Connected Assets (e.g. Fleet Management), Business Networks

OSIsoft, Telit – Device Data ManagementCisco Jasper, Vodafone – Connectivity ManagementHadoop, Apigee – API ManagementAccenture and others – System Integrators

Source: Gartner, J.P. Morgan

48



AI-enhanced customer-facing channels

These applications are focused on improving the overall customer experience – in a multichannel world, they aim to be the gateway to the customer (potentially disintermediating others). They can include conversational apps, recommendation systems, virtual customer assistants, chatbots, and voice bots.

Key chatbot components include: Front-end interface, NLP / ML framework, Dialogue management, and Back-end interface.

Customer facing AI systems allow people to more naturally interact with the digital world, and this is progressing even further with augmented reality, virtual reality and digital twins – frequently linked to AI computer vision (for example, to help guide mechanics fixing machines or stroke patients to perform their physiotherapy exercises).

Customer-facing AI vendors: Kore.ai (all-in-one chatbot PaaS), Openstream, NextIT, Aspect, Amelia from IPsoft.

Recommendation and search engines: Netflix, Amazon, Facebook newsfeed, Google advert selection, Baidu Search, etc.

Digital assistants, often embedded in devices, search engines or commercial channels, include: Apple Siri, Amazon Echo (Alexa), Samsung Viv & Bixby,Microsoft Cortana, Apple HomePod, Google Now, Google Assistant, Google Home, Baidu Duer and Baidu Brain (voice & facial recognition).

Messaging channels with AI chat include: Facebook Messenger, Slack, Hipchat, WhatsApp, Tencent WeChat, Twitter, Kik, LINE, Skype for Business.

Chatbots from enterprise vendors such as SAP, Oracle, Salesforce, and others are also introducing chatbot capabilities. Sage has accounting chatbot Pegg.

IT Services

IT Service vendors assess, select, design and implement AI functions for enterprises, especially custom projects. The consulting and SI spending for AI, including Smart RPA (robotic process automation), is today valued at around $2-3bn, but growing quickly to an estimated $29bn by 2021 (source: Gartner). Consulting is the largest component of spend today (approx. 2/3 of spend), due to the early stage of technology, and typical need to create bespoke solutions. However, as the industry grows and matures, Implementation should be around 2/3 of spend by 2021, while Consulting drops to just 1/3 – although still probably 4x larger in absolute terms.

Ultimately, we expect IT Services vendors to democratize and facilitate the proliferation of AI knowledge and best practices across enterprises. They play several roles along the way: advisory & education, bespoke creation and solution design (including security and risk management), deployment and integration into the existing software ecosystem, then ongoing maintenance and enhancements.

Some key IT services vendors focused on AI include:

Accenture

Capgemini

Atos

49



Infosys

Wipro

IBM

TCS - Tata Consulting Services

Cognizant

Intelligent things/devices

We believe that broad-based diffusion of AI applications will result in the migration of ‘intelligence’ from the datacenters to end-devices (such as smartphones, PCs, cars, home appliances, speakers, or IoT end-points in general) as not every AI-enabled decision will be computed in the datacenter. For example, an autonomous car needs capabilities of instantaneous decision making based on real-time feedback of sensory inputs (cameras, radars, etc.); or smart surveillance via a network of cameras will need capabilities of facial recognition / video analytics for near-real-time decisions. We expect this migration of intelligence from the ‘core’ of the network to the ‘edge’ to create meaningful opportunities for hardware vendors. Some examples of such intelligent devices include security cameras from HikVision, smart cameras for industrial applications (from vendors such as Basler, Cognex, etc.), cars equipped with autonomous driving capabilities, smart speakers, smartphones, consumer and industrial robots, etc.

50



Figure 30: Softbank's Pepper

Source: Softbank

Figure 31: Sawyer from Rethink Robotics

Source: Rethink Robotics

Figure 32: Relay – A hotel butler robot from Savioke

Source: Savioke

Figure 33: Amazon Echo with Alexa imbedded

Source: Amazon

Migration of intelligence from the core to the edge creates meaningful opportunities for semiconductor vendorsAny AI-enabled IoT device (be it smartphones, cars, cameras, or speakers) will need capabilities to sense its surroundings, process the sensory data to arrive at an actionable output, and / or communicate with other devices or the cloud for further action. This creates an enormous opportunity for semiconductor devices as each of these ‘sense’, ‘process’, and ‘communicate’ functions require semiconductors. Examples of sensing systems include 3D cameras, radars, lidars, and other non-optical sensors (for measuring parameters such as pressure, temperature, etc.). Semiconductors used for processing can be low-ASP devices such as MCUs to high-ASP devices such as sensor fusion chips (which will include accelerators for machine learning inference, in addition to general purpose processing chip). Communication semiconductors include devices such as Bluetooth / WiFi / RF chips.

We estimate the demand for IoT semiconductors to grow at 20% CAGR during 2017-25E, reaching ~$72bn in 2025E, up from ~$17bn in 2017E.

51



Figure 34: Semiconductor* market for IoT devices$ (bn), %

Source: Gartner, J.P. Morgan. *Includes semiconductor revenue from sensing, processing, and communication systems.

As a corollary, the migration of intelligence to IoT end-points will also lead to a surge in demand for semiconductor memory as the memory content in these intelligent end-points will go up.

Specialized Infrastructure for AI

Deploying machine learning algorithms for any application requires the initial key steps of training and testing the algorithm. Training the ML algorithm requires a lot of processing power and is particularly suited to a class of compute engines that specialize in highly parallel processing, i.e. GPUs (Graphics Processing Units). A traditional general-purpose CPU (such as Xeon chips by Intel) consists of a relatively small number of processing cores that are optimized for serial processing. On the other hand, a GPU typically consists of thousands of cores (albeit, these cores are less powerful than the traditional CPU cores) and are much more suited to handle multiple tasks simultaneously (for example: computing the activation state of multiple neurons within a neural network layer). Advances in ML have, in part, been led by the adoption of GPUs for training complex and deep neural networks that would otherwise consume a lot of time on traditional CPU-only systems. Nvidia has the leading provider of GPUs for machine learning training in the datacenters and has witnessed rapid growth in revenues associated with this segment.

Figure 35: CPU vs. GPU - Number of processing cores

Source: Nvidia

Figure 36: CPU vs. GPU: Training time for AlexNet (ML algorithm)

Source: Nvidia. Note: Titan is the brand name for Nvidia's GPUs.

23%24%

27% 26%24%

21%19%

16%13%

12%

0%

5%

10%

15%

20%

25%

30%

-

10

20

30

40

50

60

70

80

2015 2016 2017E 2018E 2019E 2020E 2021E 2022E 2023E 2024E 2025E

IoT semiconductor revenue ($, bn) YoY growth (%) - RHS

52



Figure 37: YoY growth in Nvidia's datacenter revenue

Source: Company data.

While GPUs have been the mainstream industry choice for training ML algorithms, alternative heterogeneous computing architectures have emerged which use semiconductors such as FPGAs (Field Programmable Gate Arrays) or customized chip designs such as Google's Tensor Processing Units (TPUs) to handle the inference side of the workload. Intel (a traditional CPU vendor) has also accelerated its investments in heterogeneous computing via acquisitions (Altera in 2015, Nervana and Movidius in 2016) – The recently announced Intel Nervana Neural Network Processor is an example of Intel’s AI-focused processor offering. In addition, there are also new chip architectures that are being trialed for future AI-focused applications. Some of the prominent names here include: IBM (True North chip), Qualcomm (Zeroth), BrainChip, etc.

Apart from computing, growing adoption of ML applications is also impacting other elements of the datacenter infrastructure (namely, storage and networking). On the storage front, not only are we witnessing rapid adoption of SSDs (Solid State Drives) in the datacenter, but also emergence of a new class of storage devices, i.e. Storage class memory (such as 3D Xpoint, ReRAM, etc.). These semiconductor storage devices bring the benefits of lower latency, which is crucial for data-centric machine learning applications.

Cybersecurity—immediate benefactor to machine learning

The most pervasive use of artificial intelligence, or to be more specific, the use of machine learning to improve products is found in our cybersecurity universe. Here, as far back as the early 2000’s the industry was attempting to use computer modeling labeled heuristics to determine what constituted normal behavior within computer environments. Things identified outside of determined thresholds were considered outliers and flagged for further investigation. Now, with the ability to ingest, store, manage, and analyze massive amounts of data we are seeing true machine learning being applied to several segments of cybersecurity ranging from endpoint malware detection/prevention to high level security intelligence and analytics.

Every one of our cyber security companies under coverage has numerous open positions looking to hire machine learning and AI professionals, showing the level of interest and investment being targeted to the segment. See profiles of our covered companies in Appendix 2.

54%

-13% -8%10%

63%

110%

193%205%

186% 175%

109%

-50%

0%

50%

100%

150%

200%

250%

1Q15 2Q15 3Q15 4Q15 1Q16 2Q16 3Q16 4Q16 1Q17 2Q17 3Q17

Nvidia datacenter / HPC Cloud revenue growth YoY

53



Appendix 1: Artificial Intelligence 101

In simple terms, artificial intelligence is the simulation of human intelligence by machines. For example, the development of computer systems with the ability to learn, reason, discover meaning, perceive environment, learn from experience, and interact.

In practice, AI is a group of technologies that helps facilitate the discovery and analysis of information for the purpose of making predictions and recommendations, support decision making, facilitate interactions and automate certain responses. AI tools will be intimately linked with the overall digital transformation going on now within businesses, and AI is likely to be embedded in numerous technology applications within a few years.

AI is different from traditional software programs in that it extracts knowledge from data and can alter its behavior (or learn) without being specifically programmed. Traditional software pre-defines the logic, whereas AI discovers the patterns and logic.

Historical foundations (philosophy and mechanics)

The concept of Artificial Intelligence has been around since … well, ancient Greece, if you trace the concept of intelligent machines to Hephaestus, the mythical blacksmith who manufactured mechanical servants, or more legitimately to Aristotle, who formulated the first deductive reasoning system.

The Antikythera mechanism (dated from around 70 BC) used mechanical gears to program and predict astronomical positions, and we see such technology used by clockmakers again in the 16th century. Along the same time period, Leonardo DaVinci invented numerous mechanical devices – many weapons, but also a famous walking lion (or robot, so to speak, although the word robot didn’t appear in English until 1921). In 1642, Blaise Pascal invented the first digital calculator – to help with tax accounting, of all things. Thirty years later, Gottfried Leibniz improved on Pascal’s calculator to execute all four arithmetic operations with a device called the stepped reckoner; and he followed this with a cipher machine and then the Integraph to solve differential equations. Meanwhile, Samuel Moreland invented a machine that made trigonometric calculations.

Renaissance philosophers also expanded upon the ideas of machines and intelligence, with Descartes' "Cogito; ergo sum" (I think, therefore I am), and his idea that the "bodies of humans and brutes were complex machines made by the hands of God". Interestingly, Descartes proposed that a perfect machine that resembled the organs and outward form of any “irrational animal" would be indistinguishable from the real animal. However, humans could be distinguished by two tests to prove they were really human: 1) humans could use words/language comprehensibly, and 2) that humans would act from knowledge and reason and some sense of morality. Descartes was differentiating humans from animals (obviously our knowledge of animals, evolution and ourselves is more sophisticated today) but, nonetheless, his machine/human test remains relevant even today when it comes to identifying a human vs. a machine.

By the 18th century, mechanical devices were more commonplace, from music boxes (with the musical code carved on a metal disk) to toys and automata (human-like

54



machines). There were even AI hoaxes – one of the more infamous being the Mechanical Turk. This chess-playing machine was exhibited as an intelligent automaton (beating many chess players in the Austro-Hungarian court and throughout Europe – including Bonaparte and Benjamin Franklin), but was later revealed actually to be worked by humans hiding inside. Today Amazon’s Mechanical Turk is a marketplace for humans to perform tasks that computers can’t currently perform (but their work is also helping to train the machine).

In 1801, Joseph Jacquard programmed a weaving loom using instructions on punch cards (the punch card method was used by the first IBM computers as well, right up through the mid-1970s). Charles Babbage came up with the concept of a digital programmable computer – although he used a mechanical computer (with levers and gears, rather than electronic components). George Boole came up with a form of algebra in which all values are binary: either true of false (0 or 1 for the computer bit). Today, Boolean search incorporates "and", "not" and "or" to show relationships between the search words and to drive more relevant results.

Finally turning to the 20th century, Alan Turing was a pioneer of modern computing and artificial intelligence (he also contributed to the computer that decoded the Enigma in WWII). The Turing Test (or imitation game) was designed to determine whether a computer could demonstrate intelligent behavior (via typed words on a computer screen) that was indistinguishable from that of a human. Fresh medical research at the time had shown that the human brain was an electrical network of neurons, and computer scientists were contemplating ways to replicate neural networks. Meanwhile, in the realm of science fiction, writer Isaac Asimov wrote the Three Laws of Robotics: 1) a robot may not injure a human, 2) a robot must obey humans, unless it conflicts with law one, and 3) a robot must protect its own existence, unless it conflicts with laws 1 or 2.

In 1956 the Dartmouth Conference brought together a dozen leading scientists from academia and industry over a 6- to 8-week brainstorming period in Hanover, New Hampshire, to discuss the idea that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” Artificial intelligence was born!

Numerous projects and funding (from MIT, DARPA, US National Research Council, and others) optimistically flowed into the field of AI for the next decade, only to hit a wall in the 1970s, largely due to limited computing power and a general lack of progress.

A rejuvenation of interest occurred in the late 1990s, and many advances were made in machine learning, multi-agent planning, scheduling, data mining, natural language understanding and translation, imagery, virtual reality, and games. Famously, the Deep Blue chess machine from IBM beat the world chess champion, and by the 2000s there was a new proliferation of “smart” machines, games and toys (from robopets and autonomous vacuums to NASA’s Opportunity Mars Rover and Google's self-driving car). By the end of the decade, Xbox 360 was using 3D cameras and infra-red detection to capture human motion to activate game figures on screen. Nevertheless, HAL9000-type AI autonomy certainly did NOT exist by 2001, or even today (reference to 2001: A Space Odyssey).

In 2011, IBM’s Watson beat two Jeopardy champions in the complex TV-quiz show, and that same year Apple's Siri was launched, the first of several intelligent personal assistants utilizing a natural-language user interface to answer questions, make

55



recommendations and perform actions. Then, in 2015, Google's DeepMind AlphaGo beat the European and then world champions at Go (a game with more potential combinations than there are stars in the sky). This year, the Libratus AI program designed to play poker beat four top players in Texas hold'em, using not a built-in strategy but rather an algorithm that computes the strategy.

On the philosophical/political front, in early September, Russian president Vladimir Putin said "the one who becomes the leader in this sphere [AI] will be the ruler of the world.” Meanwhile, China has announced its ambition to become the global leader in AI research.

Untangling the jargon

Despite remarkable advances over the last several decades, computers struggled with tasks that a six-year-old human could perform with ease (such as identifying images, understanding context, etc.). Such tasks required a new approach to programming. Instead of explicitly specifying rules to model the output, programmers fed a few starting rules combined with a framework that enables the computer to "learn" (much as a human baby would learn – through training). This approach, called machine learning, has led to breakthroughs in the fields of image recognition, natural language processing (the ability to understand language), and predictive analytics, to name just a few. The idea of enabling machines (or computers) to learn is not new; however, we have made rapid advances in this field in the current decade.

While the success of such self-learning computers in games like Jeopardy, Go, and poker captures headlines, we often overlook their presence in our everyday lives: customized Facebook feed, recommendations on Netflix / Amazon, spam filter in Gmail, Google search results, and Apple's Siri are all in some way or another driven by machine learning.

More often than not, terms such as machine learning, deep learning, and artificial intelligence are used interchangeably to describe the mechanism driving such self-learning computers. The popularity of these terms has moved in-sync over the past several years.

Figure 38: Search hits for machine learning, deep learning, and artificial intelligence over the past 5 years

Source: Google Trends. Search hits for machine learning, artificial intelligence, and deep learning indicated by blue, yellow, and red lines respectively. Updated as of 20 Nov 2017.

Although linked to each other in some form, these terms are not equivalent and differ from one another in several key aspects.

56



Man vs. machine – the intelligence perspective

The term Artificial Intelligence (AI) often conjures up images of popular fictional characters such as C-3PO (Star Wars), Rosie (Jetsons), or Terminator – machines that are capable of human-level reasoning / intelligence. However, in reality, AI is much broader than the confines of these sentient machines. AI, a term coined in the 1950s, has evolved in meaning and perspective over the last several decades. A simple pocket calculator or an autopilot controlling a commercial airplane (or a car) can be termed as having AI (essentially, these devices possess capabilities to achieve certain goals – such as multiplying two numbers or controlling an airplane / car).

The common theme underlining the different AI applications that we see around us (be it a Netflix recommendation engine or a computer playing Go) is that they are narrow in scope – meaning, these AI-driven systems can perform one (or a few) tasks very well, but the AI cannot be applied to other tasks in general. A Netflix recommendation engine, for example, will not be able to drive a car or recommend an investment opportunity. A Netflix recommendation engine will do what it is designed / trained to do – give movie recommendations.

The AI that we encounter today can be clubbed into a category termed Artificial Narrow Intelligence (ANI). The AI that can emulate human-level reasoning and thinking is termed Artificial General Intelligence (AGI). As it stands, we are still a couple of decades away (most optimistic scenario) from realizing computers capable of AGI.

AI vs. machine learning vs. deep learning

As NVIDIA views it, the easiest way to visualize the difference between these terms is to imagine three concentric circles, with AI as the biggest circle, followed by machine learning and then deep learning. Machine learning is a technique that enables intelligence in computers (or an AI-enabler) and deep learning (roughly modeled after the way human brain works) is one of the several machine learning techniques that is currently gaining momentum. We will explain machine learning and deep learning in detail in the following sections.

Figure 39: AI vs. machine learning vs. deep learning

Source: Nvidia.

Narrow AI vs. General AI

Narrow AI (ANI), also known as "applied AI" or "weak AI" is

purpose-built for a single/narrow

task.

General AI (AGI), also known as

“strong AI” is the pursuit (still many

years away) of a machine that can emulate human or higher-level

general reasoning across nearly

unlimited functions.

57



Why do we want machines to learn?

Conventional computer programs are rules-based, meaning that the programs contain specific instruction sets or rules (such as “if X then Y, else Z”, etc.) which are used to process input(s). This works well in a deterministic environment, where the model parameters are known beforehand and can be structured in a logical framework. For example, instructing a car to travel from point A to point B (assuming no obstacles), as shown in the figure below, can be easily achieved using a conventional rules-based program (start at point A, travel 100m straight, turn right, travel 100m, turn left, travel 100m, arrive at point B). However, to program the same route in the presence of several obstacles (other vehicles moving in the same / opposite direction, pedestrians crossing the road, traffic signals at several intervals, etc.) would be difficult as the variable parameters are not known beforehand. (Using a conventional rules-based approach to model this environment would entail writing a very lengthy code taking into account all possible scenarios.)

Figure 40: Sample car-route problem of traversing from point A to point B


Machine learning – An alternative to rules-based programs

An alternative approach to achieve the desired output in the above scenario would be to "train" the computer to recognize different obstacles and act appropriately rather than attempting to model the entire route (with all the possible permutations and combinations). Conventional computer programs face similar roadblocks when dealing with issues such as fraud detection, pattern recognition (facial, voice, molecular structure, etc.), language processing, and predictive analytics to name just a few examples. One common theme underlies all these problems: The variables involved are fairly indeterministic (as was the case with the car-route problem highlighted above) and hence difficult to structure using conventional rules-based programs.

‘Hey Siri, what is machine learning?’

Put simply, machine learning (hereafter referred to as ML) refers to a process flow that enables systems (computers) to learn from a set of inputs (or known scenarios) without specifically instructing the system what to do (using rules-based software routines). In the example of the car-route problem highlighted above, an ML-driven system would be trained on several hours (possibly running into millions) of driving videos (and even still images for obstacle-identification), enabling the computer to recognize different obstacles and act in a desired way in a given situation. Contrast this with the herculean task of trying to tell the system how to act for every possible situation that could arise with a given set of variable parameters.

Machine learning is the process of using algorithms to learn from input

scenarios (data) and make a

prediction / determination about an unseen scenario, based on that

learning. This contrasts with the

conventional rules-based approach to programming, wherein specific

instructions have to be coded to deal

with unseen scenarios.

58



Figure 41: Machine learning* vs. traditional computing

Source: J.P. Morgan. *Supervised learning

Automating the learning process

The goal of ML algorithms is to enable automated learning without human intervention (a computing paradigm referred to as programming by example). ML greatly simplifies modeling complex environments with several, and often indeterministic, variables. There are three key components in any ML algorithm:

Model: Helps represent the problem statement (knowledge representation) in a way that can be understood by computers (examples include graphical models, decision trees, neural networks, etc.). This is the ‘brain’ of the ML algorithm where the relationships between different parameters are established to produce an output (or prediction).

Parameters: These comprise different factors (or features) that are used by the model to arrive at an output.

Learner: This is the central component of any ML algorithm. Its primary function is to learn from the discrepancy between actual results and model output and tweak the model parameters (and how different parameters relate to each other) in order to arrive at the best possible output (the one with the lowest possible discrepancy).

The process involved in arriving at the best possible model is referred to as trainingthe ML algorithm. Once trained, the ML algorithm can then be run on a test dataset to gauge its effectiveness (a process referred to as testing). Post testing, the ML algorithm is ready for implementation (also referred to as the inference step) to be applied on “unseen” data.

59



Figure 42: Basic process flow* involved in training ML algorithms

Source: Figure adapted from Google presentation, J.P. Morgan. *A sample case of supervised learning presented here.

Types of learning and application areas

There are three primary types of learning: Supervised (also known as inductive / classification learning), Unsupervised, and Reinforcement learning.

In supervised learning (which is the dominant form of ML in use today), the training dataset is labeled / tagged with the ‘output’ being modeled. For example: Training an ML algorithm to identify spam messages using labeled messages ("Spam" or "Not spam") to help the algorithm recognize features that define or characterize a spam message. Other examples using supervised learning include evaluating an individual's creditworthiness (based on tagged data for creditworthy and delinquent customers), facial / voice recognition (based on tagged data of different faces / voices), etc. Supervised learning is typically employed for problems wherein the ‘right answer’ or the output is known for a large number of similar instances.

Another technique, unsupervised learning, involves training the algorithm on a dataset but without telling it what to look for (contrast this with the supervised learning approach, where the training dataset is labeled with the output being modeled). Under this approach, the algorithm learns to recognize different features from the training dataset provided and groups together examples with similar characteristics (clustering). This approach is useful for cases wherein the different categories or classes within the dataset are not known beforehand. As an example of unsupervised learning, back in 2012, Google’s ML algorithm learned how to detect cat faces after Google fed 10mn YouTube video stills to the algorithm – Google programmers did not write the code to detect cat faces nor did they include specific labels to identify cats (as would be the case in supervised learning) – the algorithm learned this by itself.

A popular application of unsupervised learning is for making recommendations (be it Netflix or Amazon). As an example, Amazon can use data on consumer purchases to train an ML algorithm to recognize hidden patterns or links between different consumers and relate the different consumer clusters to the items they typically

For a more detailed description of ML models, techniques, classifications and comparison of algorithms, see JPM note: Big Data and IT Strategies by Marko Kolanovic and Rajesh Krishnamachari.



60



purchase. This forms the basis of purchase recommendations (the same principle applies to movie recommendations on Netflix).

Lastly, reinforcement learning is an approach wherein the learning process is contingent on the reward provided for the model output. As with unsupervised learning, there are no explicit instructions or labels provided in the training dataset; however, unlike unsupervised learning, the model output is tweaked based on the reward provided. Here, the objective of the ‘Learner’ is to maximize the reward by tweaking the model parameters / structure. AlphaGo, a program by Google's DeepMind, which beat world Go champion (Lee Sedol), in 2016, is based on this learning technique.

Neural networks and deep learning

The ImageNet tipping point for deep learning

While there are several classes of ML algorithms in use, one class has been in focus recently: Deep Learning. The seeds of deep learning algorithms were planted decades ago; however, this technique came into the limelight in the current decade after a breakthrough in error rates was achieved in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) using a deep learning algorithm in 2012.

ImageNet is an online database containing millions of hand-labeled images. The annual ImageNet challenge calls upon researchers in the field of computer vision to compete and measure the progress in detection across a wide variety of objects. The algorithms are first trained using a set of labeled images and are then asked to identify labels for unseen images. The winners are ranked based on the error rates in labeling a new set of previously “unseen” images. In 2010, the error rate of the winning algorithm was 28% (vs. the human error rate of roughly 5%). In 2012, a team led by Geoff Hinton (part-time Googler and a professor at the University of Toronto), achieved a meaningful reduction in error rates (to 16%) using deep learning. This sparked further interest in this technique, leading to advances in the field and increased sophistication of deep learning algorithms. The winning algorithm (based on deep learning) in 2015 achieved an error rate of less than 5% (beating humans), and in 2017 the error rate of the winner had dropped to 2.3%.

Figure 43: ImageNet Large Scale Visual Recognition Challenge (ILSVRC) error rates

Source: ImageNet. Other factors beyond classification were tested: Localization, Object Detection (still & video).

Deep learning algorithms have enabled rapid progress in the areas of Natural Language Processing (NLP), Pattern Recognition, Autonomous Driving, Fraud detection, among other areas. Deep learning employs the use of large (or deep)

28%26%

16%

12%

7%5%

3.6% 3.0% 2.3%

0%

5%

10%

15%

20%

25%

30%

2010 2011 2012 2013 2014 Human 2015 2016 2017

Cla

ssif

icat

ion

Err

or

Rat

e

61



Artificial Neural Networks or ANNs (this is the ‘Model’ for Deep learning, as highlighted in Figure 42).

Artificial Neural Networks – A simplified model of the human brain

Artificial neural networks (ANNs) are roughly modeled after the human brain – just as billions of interconnected biological neurons give humans the ability to learn / recognize patterns / etc., an interconnected network of artificial neurons enables computers to identify patterns (or learn from input scenarios / data) and store that learning for application in unseen scenarios.

ANNs are composed of interconnected nodes of artificial neurons (which are emulated via a software algorithm). Essentially, artificial neurons act as feature extractors in a deep learning algorithm. Each connection in the neural network is assigned a weight and each node has its own activation function (which determines when the neuron will be fired up or activated). The weights between different neural connections are altered during the training process – this is the step where the algorithm learns what the defining characteristics of the input scenarios are and the weights are adjusted in order to minimize the discrepancy between the algorithm prediction and known outputs for different scenarios.

Input data is fed to the network via the input layer, with the processing (feature extraction) taking place in the ‘hidden layers’ of the ANN. A Deep Neural Network (DNN) is simply an ANN with several hidden layers. DNNs can capture several kinds of complex relationships (on multiple levels of abstraction) and hence it allows users to model complex phenomenon relatively easily (which otherwise would have been very difficult to code using conventional rules-based programs).

Figure 44: An artificial neuron (also called perceptron)


Figure 45: Basic representation of an artificial neural network


62



Figure 46: Feature representation in deep learning


Figure 47: Examples of deep learning process flows (highlighting feature representation / extraction) in object detection, speech recognition, and Natural Language Processing

Source: Adapted from Andrew Ng’s presentation (Link), J.P. Morgan.

Deep learning process flow

The process flow for applying deep learning algorithms (applies to ML algorithms in general) starts with cleaning the input dataset (to remove errors / outliers and to ensure continuity). The input dataset is then split into two or three sub-sets (for training, testing, and / or validation).

Training the deep neural networkInitial reference points for the deep neural network (number of hidden layers, node count, connection weights and activation functions) are provided by the programmer. The training process involves feeding a large number of examples (from the training dataset) to the neural network.

https://www.csee.umbc.edu/courses/graduate/678/spring15/visionaudio.pdf

63



DNNs learn by extracting features from the training dataset (be it patterns in the image pixels or audio / video content). Consider the sample process involved in training a DNN for facial recognition presented in the figure below. Essentially, the DNN breaks down the input images and collects features from the individual pixels. The relationships between these features (connection weights) are then optimized during the training process to arrive at the model with the lowest possible discrepancy with the actual output.

Broadly speaking, a deep neural network is capable of extracting a large number of features (and thus capable of identifying complex shapes / forms) due to the presence of multiple numbers of hidden layers. The training process requires huge computing power and is typically executed with the help of highly-parallel computing chips such as GPUs (Graphics Processing Units).

Figure 48: Feature extraction / representation in a DNN to be used for facial recognition

Source: Adapted from Nature (Andrew Ng), J.P. Morgan.

DNN testing and implementationDuring the testing phase, the optimized deep neural network is fed input samples from the test dataset and output values are compared against the 'right answers' (may be intuitive guesses as in the case of unsupervised learning or rewards provided for the output in the case of reinforcement learning) to gauge the efficacy of the model. Once the DNN has been trained and tested, the model is now ready for use on "unseen" data.

Can machines think and reason?

Now that we know a computer can be trained to learn like a human, can it also be made to "think" and "reason"? The short answer to this question is: Not yet and this probably will not happen for at least the next several decades.

As explained previously, ML algorithms learn from input examples by identifying and stitching together different features that are representative of a particular segment within the input data. These algorithms work well in situations for which there is a large amount of data (with several features or dimensions) available for training and the relationships between different features in the data are more or less steady (or constant). Some of the prominent use-cases for ML algorithms have been in the areas of computer vision, NLP, speech recognition, pattern recognition, fraud detection, and predictive analytics.

64



An important point to note here is that ML algorithms are only as good as the data that is used for its training. Hence, biases / errors inherent in the datasets will more often than not lead to broken ML algorithms that yield incorrect (and often embarrassing) results.

Roadmap to AGI

Artificial General Intelligence (AGI) – the capability to demonstrate human-level reasoning / thinking / learning to perform any task – is considered the holy grail of AI research. While there is no agreed upon roadmap to solve the problem of general-purpose intelligence, there are two key approaches worth mentioning:

1. Players in the revolutionary camp believe that AGI can be realized by using a radically different approach (rather than using existing ML techniques) – a few examples here include Sentient, Numenta and Vicarious. Sentient is employing concepts of evolution (survival of the fittest) to develop algorithms capable of general-purpose intelligence, while Numenta and Vicarious are reinventing the way information is processed.

2. Players in the evolutionary camp believe that AGI can be approached by buildingon the basic ML blocks (such as NLP / computer vision / speech recognition / etc.) already in place (stitching these basic blocks together by something known as transfer learning). There is also heightened interest in applying reinforcement learning techniques. Google’s DeepMind is one key player that is currently using reinforcement-learning-based techniques to further the progress toward AGI.

How soon AGI can be achieved is still a widely debated topic; however, most seem to agree that AGI, or a machine capable of human-level thinking and reasoning, will be achieved some time over the next couple of decades.

The Singularity

The Singularity is typically defined as the tipping point when artificial intelligence is exponentially smarter than humans, to the point that it could trigger the end of humanity. The theory goes that a super-intelligent entity might continue to upgrade itself at such a great pace that humans would seem utterly insignificant – calling into question the necessity or continued existence of humanity. Many argue over if/how/when we get there, and whether humanity can survive or transcend (through genetic engineering, molecular nanotechnology and mechanical/electronic/neural enhancements) or be linked to the singularity. Interesting as it may be, this is beyond the scope of our investment report.

65



Employment in the face of AI: Disruption or Evolution?

An important issue stemming from the widespread adoption of self-learning systems across the consumer and enterprise space is the impact that AI will have on employment. Today, AI-enabled systems can technically replace not only low-skilled jobs but also ones that require a high degree of expertise (such as reading CT scans). Advances in techniques such as Natural Language Generation (NLG) are now enabling systems to take on tasks in the areas of journalism and investment advisory. Intelligent chatbots are disrupting the customer service space. Commercial service robots (capable of operating in unpredictable environments) are automating tasks from merchandizing to security. There is no denying that these self-learning machines can perform certain tasks much more efficiently than humans can, resulting in productivity improvements (As per IDC, AI-applications are expected to yield productivity improvements in excess of $60bn per year for U.S. enterprises by 2020). However, its impact on the job market cannot be ignored.

Predictable / Routine tasks most susceptible to automation

As it relates to impact of automation on employment, it is important to study the impact on tasks that comprise a particular job rather than thinking of automation as a replacement for the entire job (which may include several tasks that cannot be automated using current technologies). Against this backdrop, we note that automation (led by AI-enabled systems) will impact almost all jobs (with varied levels of intensity depending on the type of tasks involved) over the coming years.

McKinsey, in a study titled “Where machines could replace humans – and where they can’t (yet)”, identified seven top-level task groupings and detailed their technical feasibility of automation (or the % of time spent on tasks that can be automated using currently available technologies). Predictable physical work (such as operating a machine or setting up products for display on shelves in a retail store) and data processing / collection are among the tasks that are most susceptible while tasks requiring managerial skills and applying expertise to decision making / creative tasks were ranked as the least susceptible to automation by currently available technologies.

Figure 49: Technical feasibility* of automation for different tasks

Source: Adapted from McKinsey, J.P. Morgan. *Technical feasibility of automation is defined as the % of time spent on tasks that can

be automated using currently available technologies.

As defined by Autor et al in their

2003 paper, “The skill content of recent technological change”, a

task is a unit of work that

produces output. Further, a skill is a worker's stock of

capabilities for performing

various tasks. Workers apply their skills to tasks in exchange

for wages. The paper (and its

update) can found here (2003)

and here (2013).

http://economics.mit.edu/files/9758

http://economics.mit.edu/files/581

http://www.mckinsey.com/business-functions/digital-mckinsey/our-insights/where-machines-could-replace-humans-and-where-they-cant-yet

66



It is worth noting that predictable physical work features prominently in sectors such as manufacturing, logistics (esp. warehousing), retailing, and food / accommodation services. Data processing and collection comprises a meaningful portion of tasks performed by workers in the financial services / insurance / healthcare sectors. As we highlighted previously, these are the same sectors that are leading the adoption of AI (in terms of both investments and deployment).

Further, U.S. workers spend more than 50% of time on tasks that are highly susceptible to automation. This signals a massive potential* for AI-led automation in current jobs (what applies to the U.S. broadly applies to other developed economies).

Figure 50: % time spent in all U.S. occupations

Source: Adapted from McKinsey, J.P. Morgan.

The findings from McKinsey are in-sync with the work done by Osborne and Frey (who estimate that ~47% of U.S. employment is at risk from computerization). To be clear, the impact of automation on employment is not a new phenomenon (and discussions on this topic date back to the Industrial Revolution). In fact, a closer look at the mix of U.S. employment over the last couple of decades demonstrates how employment in non-routine jobs (cognitive and manual) has grown, while employment in routine jobs has stagnated / declined (of note, Figure 51 also includes the impact from offshoring). With the proliferation of AI across several industries, we expect this trend to continue.

Figure 51: The changing composition of U.S. employmentThousands of persons, not seasonally adjusted

Source: Federal Reserve Bank of St. Louis, J.P. Morgan.

Managing others, 7%

Applying expertise, 14%

Stakeholder interactions, 16%

Unpredictable physical work, 12%

Data collection, 17%

Data processing, 16%

Predictable physical work, 18%

10000

20000

30000

40000

50000

60000

Jan-83 Jan-88 Jan-93 Jan-98 Jan-03 Jan-08 Jan-13

Non-routine cognitive Routine cognitive Routine manual Non-routine manual

*As highlighted by McKinsey, there are several factors that determine whether

a particular task will eventually be

automated. These factors include: a.) Technical feasibility, b.) Cost of

developing / deploying both hardware

and software for automation, c.) Labor costs and relative scarcity of skills, d.)

Benefits beyond labor cost savings

such as better productivity and quality, and e.) Regulatory

environment and social-acceptance

issues.

A ‘routine’ task is one that can be coded into algorithms that can be

executed by machines. On the other

hand, a cognitive task is one that requires mental skills and adaptability

to the task at hand.

Examples:

Non-routine cognitive: Managing

people, education/tutoring, writing algos for autonomous driving, etc.

Routine cognitive: bookkeeping,

clerical work, interpreting x-rays, aspects of nursing, repetitive

customer service, etc.

Routine manual: Manufacturing (predictable environment), food prep,

packaging, transportation services,

etc.

Non-routine manual: Hotel

housekeeping, construction, raising

outdoor animals, preparing a meal.

http://www.oxfordmartin.ox.ac.uk/downloads/academic/The_Future_of_Employment.pdf

67



Will AI be a net creator or destroyer of jobs?

In a 1930 essay titled, Economic possibilities for our grandchildren, Keynes described ‘technological unemployment’ as something arising from "our discovery of means of economizing the use of labor outrunning the pace at which we can find new uses for labor”. Increasing sophistication of AI-enabled systems is likely to make more tasks amenable to automation going forward. This may potentially lead to job-polarization (concentration of jobs either in the high-paying, non-routine cognitive or in the low-paying, non-routine manual buckets) and displacement (esp. in certain verticals such as manufacturing, customer service, financial services, etc. that have a higher proportion of automatable jobs).

Net creator or destroyer? The ‘Lump of labor' fallacy

The impact of technological progress on employment is characterized by two key stages: 1) Destruction of jobs – as technology substitutes for labor (wherever economically viable) and 2) Creation of new jobs – as technology improves productivity, industries expand and grow creating new job opportunities. The previous waves of technological disruption have almost always led to net creation of jobs (example: the roll-out of ATMs reduced the operational cost of a bank branch, which in turn led to opening of more bank branches, creating more jobs).

We expect the impact of adoption of AI to follow a similar trend. While it is often easy to identify areas where jobs can be substituted by technology, it may not always be apparent to predict the new job opportunities (albeit, ones that will require new skills) that are created as a result of technological advancement (as an example: no one would have imagined working as a video-game designer back in the 50s). The notion that there are a finite number of jobs, which if replaced will lead to unemployment, is termed by economists as the 'Lump of Labor' fallacy.

While the current wave of AI adoption may lead to new job opportunities in the future, a short-term disruption in the labor markets seems inevitable (especially inDMs, where the adoption of AI will be more prominent. EMs may continue to chug along with traditional jobs as low labor costs may not justify deploying AI). This is where governments and regulatory agencies are likely to step in – to smooth the transition to a new information age. Acquisition of new skills via continued education and overhauling the existing educational system to ensure a robust pipeline of workers ready for the jobs-of-the-future are likely to garner increased focus from the concerned authorities, in our view.

http://www.econ.yale.edu/smith/econ116a/keynes1.pdf

68



Appendix 2: What companies are doing in AI

Given our global stock coverage, one of our main goals is to educate investors on what these companies are doing within the realm of AI.

Accenture (ACN US, OW)

Accenture’s AI practice is embedded across its ACN Consulting, ACN Technology, and ACN Digital and ACN Operations segments. The company has opened an AI R&D lab and incubation hub in Dublin called The Dock, employing ~200 researchers to innovate with clients and ecosystem partners. The company helps its clients build/deploy AI solutions, providing advisory and implementation services. Implementation services typically include offering AI based solutions (custom or standard), training AI robots, data modeling, and SI.

Accenture Artificial Intelligence Engine (AAIE) is a framework that is deployed as a service, for on-demand flexibility, open-source software to deliver modular, reusable, scalable and cost-effective AI automation and augmentation services. ACN implements AI as a part of its platforms (IP) as well as through industry collaborations.

The AI framework consists of:

1. Intelligent Automation. Eliminates repetitive tasks through automation while improving complex problem solving, risk analysis and business decision-makingcapabilities.

2. Cognitive Robotics. Create bots with perception, planning, memory and reasoning leveraging AI across B2C industries like customer service, retail and hospitality.

3. Machine Learning. Provides systems with the ability to learn without being explicitly programmed leveraging combination of integrated analytics and AI algorithms.

4. Robotic Process Automation. Allows organizations to deploy intelligent software systems replace the actions of human users.

5. Biometric Face Testing. An AI application that gives intelligent systems theability to gather data on facial features and demographics.

Tien-tsin Huang, CFAAC

(1-212) 622-6632

[email protected]


69



Figure 52: Application Benefits of Accenture AI

Source: Company website

ACN applies AI technologies across applications like claims processing, fraud detection, drug testing, dynamic process compliance, pharmacovigilance etc.

ACN's Intelligent Automation Platform myWizard augments human workerswith virtual agents to improve data analysis and pattern identification and drive better business outcomes. ACN claims that platform generates upto~60% productivity improvements for its clients.

ACN launched Drishti, which provides smart phone-based assistance using AI technologies such as image recognition, Natural Language Processing and Natural Language Generation capabilities to describe the environment for a visually impaired person.

Intelligent Robot called Pepper at Softbank, embedded with AI technologies to drive customer engagement and experiences.

In 2016, ACN and RoboValley, an innovation hub for robotics located at Delft University of Technology in the Netherlands, announced a 5-year collaboration designed to advance the development of the next generation of robotics technology.

Accenture Robotics Solution is ACN's five-part solution that combines RPA with more advanced AI capabilities.

70



Figure 53: ACN’s RPA spectrum of services

Source: Company website, J.P. Morgan.

Unique Identity Service Platform. ACN developed Unique Identity technologies ranging from biometric R&D to delivery-at-scale, via algorithm benchmarking, device evaluation, prototyping, development of proofs of concept, live piloting, biometric governance & quality assurance, diagnostics & capability assessments and biometric strategy definitions.

Accenture Cyber Intelligence Platform (ACIP). Applies proprietary AI techniques to data in real-time, identifying internal and external threats.

Accenture Intelligent Email Advisor, analyzes emails using natural language processing, understands the requestor’s intent and retrieves relevant data from systems to help the team resolve issues; Accenture Procurement Advisor -prepares reports by combining data mining, text and data-analysis skills to analyze publicly available digital information including financial publications and business news.

Notable Industry Collaborations. 1) With IPsoft, an AI provider, to launch a new practice focused on its virtual agent platform, Amelia. ACN will utilize IPsoft’s Amelia platform to develop go-to-market strategies, solutions and consulting service offerings around deployments of virtual agent technology. 2) With Pega, to leverage Pega’s platforms (including its AI solution, Pega Customer Decision Hub) to provide applications at scale. As such, ACN deployed the scaled solution, Accenture Aviation Experience Accelerator to Transavia, a low-cost, European airline. 3) With Pivotal to form The Accenture Pivotal Business Group (APBG) to leverage and deploy next-gen technologies including artificial intelligence, connected cars and home, and internet of things(IoT), 4) With German Research Center for Artificial Intelligence (DFKI) to embed DFKI’s specialized AI research capabilities with ACN's analytics expertise to further the adoption of new technologies in Germany.

71



Adobe (ADBE US, OW)

The most visible use of AI within our Software Technology universe here in the US comes from Adobe with its Sensei platform. The inputs coming from millions of users—from the enterprise level all the way through to consumers—are a powerful platform for data aggregation to utilize within the growing set of algorithms in the Sensei solution set.

Making product better rather than a specific AI productSensei is the technology codename that Adobe uses to describe the artificial intelligence technology that it is baking into the creative cloud suite of products that range from Photoshop to Lightroom. The solution is able to do a lot of advanced editing capabilities that began with content aware editing. This gives Photoshop the ability to eliminate an object from an image, filling in with what the program is able to understand should be the background to that object. In fact, the company is taking it a step further with an automated deletion capability for video.

Alibaba

Secret recipe to success in AI: CUBA. Alibaba identifies four key elements of AI(the concept of “CUBA”) which includes cloud computing, use case, big data and algorithm. We think the company has competitive edges in all these areas by leveraging its strong cloud infrastructure (No.1 cloud service provider in China), extensive use case (from ecommerce, local service to digital content), rich data resource generated from various use cases and seasoned technology team.

Areas of AI application

Ecommerce

Personalized recommendation. By leveraging its rich data resources and enhanced data analytic capability, Alibaba made a major algorithm upgrade in Sep 2016 and is able to push more relevant personalized recommendations to customers. It essentially expands the presenting inventories, which generates more product exposure and improves traffic distribution efficiency. Meanwhile, the enhanced matching accuracy of recommended products with customer demographics translates to increased click-through and conversion rates, which drives both monetization improvement and GMV growth. Since the algorithm upgrade, the revenue growth rate of China retail business has accelerated from 40% in Sep Q 2016 to 64% in Sep Q 2017. In 2017 Single Day event, Alibaba achieved 100% personalization for content feeds and search result page

Figure 54: Personalized recommendation drivers revenue acceleration

Source: Company reports and J.P. Morgan

39%

24%

35% 35%41%

49%

40% 42% 41%

57%64%

0%

10%

20%

30%

40%

50%

60%

70%

Mar-15 Jun-15 Sep-15 Dec-15 Mar-16 Jun-16 Sep-16 Dec-16 Mar-17 Jun-17 Sep-17

YoY growth of China retail business

Growth accelerated after algorithm upgrade in Sep 2016

Alex YaoAC

(852) 2800-8535

[email protected]


Sterling Auty, CFAAC

(1-212) 622-6389

[email protected]


72



AI powered customer service. The company has applied AI technologies such as machine learning and natural language processing in customer service area and developed AI robot to address customer queries. During Double 11 event in 2017, some 95% of customer service queries were answered by AI robot. Thanks to this application, customer service headcount went down in the past few years despite growing business scale.

New Retail. Hema is able to fulfill online orders within 30 minutes (during non-rush hour) in a 3km radius range. This strong fulfillment capability is powered by AI algorithm that enables quick in-store preparation (product sorting and packaging are generally done within 10 mins with the help of automated equipment) and fast delivery (the system provides optimized route). On the B2B

side, Alibaba’s merchandise sourcing initiative Lingshoutong (零售通) leverages its data analytics capability to help mon-and-pop stores analyze nearby residents’attributes and therefore source potentially fast-selling products.

Financial services

“Pay with Face”. Leveraging its facial recognition technology, Ant Financial launched "Pay with Face" service at one KFC restaurant in Hangzhou this September. Alipay users can make the payment through facial recognition, which uses a 3D camera and live detection algorithm to verify users’ identity. The company offers compensation in the event of account stolen. It plans to expand this service to more use cases when the technology becomes more mature.

Risk management. AntBuckler is a big data and AI based risk management system developed by Ant Financial, which helps enterprises identify potentialfraudulent activities. For instance, coupon and cash rebate is a popular promotional strategy for user acquisition. Fraudsters generally will register hundreds of different accounts to falsely claim this incentive, causing significant loss to these merchants. Leveraging its big data analytics and machine learning capability, AntBuckler is able to identify risky user and device ID, prevent and stop any potential fraud activities.

Logistic services

Route optimization. Cainiao leverages its data analytic capability to help delivery partners to optimize routes therefore increasing deliver speed and lowering cost. As of Sep 2016, Cainiao’s fulfillment network provides same-day delivery in 37 cities/128 districts and counties and next-day delivery in 162 cities/959districts and counties.

Others

Voice assistant. Alibaba launched its first AI powered voice assistant “Tmall Genie” (similar to Amazon Echo) in July 2017. It's able to identify users’ voice and assist them in a number of areas such as controlling smart appliance, shopping, ordering local services and searching for information.

Alphabet (GOOGL US, OW)

Google has rebranded itself from being a “mobile-first” company to an “AI-first” company with AI/ML deployed throughout the organization. While Google has two teams primarily focused on AI—Google Brain, which it started in 2011, and DeepMind, which it acquired in 2014—there are 1,000+ deep-learning projects underway w/applications across Search, YouTube, Android, Gmail, Photos, Maps,

Doug AnmuthAC

(1-212) 622-6571

[email protected]


73



Translate, Waymo, and more. We believe Google will continue to attract and retain talent in the field of AI/ML—which is increasingly challenging—driven by Google’s broader scope and scale.

Search. Since early 2015, Google has incorporated RankBrain, a deep learning technology that helps refine queries & rank web pages, into its Search engine. Since then, RankBrain has become one of the most important ranking factors for Google Search. Separately, Google Brain improved video recommendations for YouTube users, leading to increases in watch time.

Google Hardware. Google is using AI/ML based enhancements incl. Google Assistant in Google Home, Google Lens in Pixel Phones, and translate in Pixel Buds to differentiate its hardware products. We believe voice and image search, with real world application in Google Assistant & Google Lens, will be incremental to the search experience over time, and while we believe these technologies are still early in adoption, Google’s hardware puts them in consumers' hand.

Google Cloud. Some of the AI/ML technology available in Google Cloud include:

TensorFlow. Google’s open-source ML framework originally developed by Google Brain team for conducting ML and deep neural networks research. It is widely used companies like Airbnb, Uber, Sap, eBay, Intel and more.

Conversational API uses ML to recognize intent & context, and helps in building conversational interfaces like messaging platforms, chatbots, etc.

Google Cloud Video API makes videos searchable and discoverable by extracting metadata.

Google Cloud Vision API uses machine learning models to understand the content of an image, detects individual objects incl. printed words and faces within the image, and classifies image into categories.

Google Cloud Speech API uses neural network models to convert audio to text and recognize 80+ languages and variants.

Any Many Others including Natural Language API, Translation API.

Waymo. This autonomous vehicle division within Alphabet has been testing its vehicles on the road since 2009. While the details around commercialization of this division is still unclear and far off, we believe it has the potential to become one of Alphabet's next big driver of growth, and its recent spin within Alphabet and rebranding to Waymo is a clear indication of progress.

AlphaGo. DeepMind’s Go-playing general-purpose AI program AlphaGo’s accomplishment includes its victory over the world champion Lee Sedol. AlphaGo utilizes technology like advance tree search, deep neural networks, and reinforcement learning, and the DeepMind team believes their approach behind AlphaGo can be commercialized and applied broadly to solve general problems in the world.

Alteryx (AYX US, N)

In June 2017, Alteryx acquired Yhat, which offers self-service data science tools for developing, managing, and deploying machine learning models to web and mobile apps. The acquisition will enhance Alteryx’s platform and build on its strategy to help organizations empower data scientists to rapidly deploy and manage advanced analytic models.

Mark MurphyAC

(1-415) 315-6736

[email protected]


74



Amadeus (AMS SM, N)

Amadeus is looking at applying AI to the travel industry in several ways, via predictive analytics and ML, as well as using chatbots and virtual assistants. However, these are more plans than actual products on offer today. Though, within its Travel Intelligence division, Amadeus offers a range of BI tools to support airline decision-making. The company uses big data tech with rich data sources and a dedicated team of talented data scientists. Data is extracted from multiple sources: industry, customer and Amadeus systems, providing in-depth information for airlines to gain real insights into overall market dynamics, their own performance and growth opportunities.

The Travel Intelligence division has a TAM of around €1.3bn pa, but with the expanding capabilities of AI, the TAM could also be increasing. Current products and solutions from Amadeus include:

Amadeus Booking Analytics

Amadeus Performance Insight

Amadeus Schedule Recovery

Amadeus Search Analytics

Future opportunities, with more AI and ML embedded: According to Christophe Ancolio, Innovation Manager at Amadeus, inspiring

people to travel is about showing them relevant content and offers, and AI and ML is helping travel providers target online travelers more effectively (via smart advertising, for example).

AI can also help travel companies create highly-tailored offers based on customers’ needs and preferences, utilizing deep learning algorithms which analyze customers’ online activities (what photographs they’re looking at, or what articles, news and ads they are reading, as well as other shopping preferences, location, time and other parameters).

Figure 55: Creating tailored offers with artificial intelligence

Source: Amadeus materials

Chatbots can be useful in helping travel agencies deal with travelers disrupted by weather or other travel delays - when things need to be rescheduled or rebooked. Finally, virtual assistants could scan its user’s calendar to identify events which

Stacy PollardAC

(44-20) 7134-5420

[email protected]


75



require travel and offer to book flights/hotels proactively, while taking into account the traveler’s preferences and past booking behavior.

We expect to hear more at Amadeus’ CMD scheduled for 13th December.

Amazon (AMZN US, OW)

Amazon was an early adopter of AI, which is prevalent throughout the company’s retail & cloud business segments. Below we highlight 4 key areas where AI is core to Amazon’s offering: 1) Amazon’s product recommendation engine, 2) Amazon Alexa, 3) Amazon Web Services (AWS), & 4) Amazon Go. We note there are many other areas where Amazon utilizes AI, including its supply chain, demand forecasting, capacity planning, fraud detection, translations, & more.

Product recommendation engine. Amazon’s AI framework DSSTNE powers the platform’s product recommendation system, which is an AI-driven information filtering system that can automatically predict user preferences and responses to queries based on past behavior, one user’s relationship to other users, similarity among items being compared, and context. Amazon’s recommendation engine plays an important role in its retail business, with an estimated ~35% (per Techadvisor) of AMZN sales generated through the recommendation engine.

Amazon Alexa. Alexa is the voice assistant, or brains behind Amazon’s Echo devices. Alexa is an intelligent AI system fueled by Natural Language Understanding and Automated Speech Recognition. Since the market debut of Amazon Echo devices in 2014, Amazon has shipped tens of millions of these smart-speakers, which creates an important feedback loop as Alexa constantly learns & improves from human data & interaction. Amazon has made it easy for developers to build Alexa voice applications and has heavily invested in tools for application creation. According to Voice Labs, there are more than 10k+ Skills on the Alexa platform.

Amazon Web Services. Amazon utilizes AI to both increase the efficiency of its cloud operations & also offer products that allow other companies to leverage its capabilities. AWS rolled out its Amazon Machine Learning Service as a competitor to similar products from Microsoft and Google, and made its deep learning software (DSSTNE) available to developers in 2016. AI driven services offered include Amazon Lex, Amazon Polly, & Amazon Reckognition.

Amazon Lex—the technology behind Alexa—provides advanced deep learning functionalities of automatic speech recognition (ASR) for converting speech to text, natural language understanding (NLU) to recognize the intent of the text, and helps developers build natural language into services such as Chatbots.

Amazon Polly is Amazon’s text-to-speech service that uses deep learning technologies to synthesize speech that sounds like human voice.

Amazon Reckognition, the service based on the same deep learning technology used to analyze billions of images daily for Prime Photos, can help detect objects, scenes, text, faces, etc.

Amazon Go. Amazon Go is a new store format Amazon is testing, which relies on computer vision and machine learning to allow customers to walk in, select items, walk out, and have their account automatically charged. Amazon Go represents the potential for a new retail format where there is no checkout process and no cashiers. Amazon Go currently has one Seattle location open in Beta,

Doug AnmuthAC

(1-212) 622-6571

[email protected]


https://www.martechadvisor.com/articles/customer-experience/recommendation-engines-how-amazon-and-netflix-are-winning-the-personalization-battle/

76



though there is potential the technology could roll-out across Amazon’s other physical locations (such as Whole Foods Market) over time.

Atos (ATO FP, OW)

Atos is looking to apply its expertise and technology in supercomputing (through the Bull brand) to the training of algorithms. They are involved with the software required to do the training (image processing, speech recognition), but are not currently developing all the blocks required – rather, they are adding in 3rd party software onto their own where necessary. Atos is also using applications designed toimprove the efficiency of IT service responsiveness.

Atos Security Operations Center (SOC), for example, based on Big Data analytics and Machine Learning technology and powered by Bullion servers, continuously learns from previous threats and orchestrates responses in real-time. Detection and total response/recovery time can be significantly reduced. By analysing and correlating high volumes of structured and unstructured data, the SOC monitors the internal customer network, social networks, deep web and the dark web for full scale screening. Through deep packet analysis, pattern recognition and weak signal detection Atos’ SOC transforms the data into intelligence.

Atos Virtual Assistant (AVA), newly launched through a strategic partnership with CogniCor (Barcelona based AI company with Machine Learning, Cognitive Methods & Natural Language Processing capability), is able to provide support to and expand Atos’ IT Service Desk offerings. AVA delivers machine learning capabilities in the cloud and is able to automatically answer IT service desk and data centerrequests through a virtual agent. As well as answering and responding to problems, AVA is able to anticipate future issues based on collected trend data such as IT tickets, FAQs and user-generated content. AVA is integrated within Atos Digital Transformation Factory, and also a key part of Atos’ help and interaction center.

Atos’ Quantum Learning Machine (QLM) can simulate up to 40 quantum bits, and can be used to develop, simulate and test quantum applications. The QLM is based on a universal programming language driven by a new ultra-compact supercomputer (about the size of an enterprise server). The universal programming language, aQasm, enables connectors, which carry programs from other quantum simulators to be developed. The fact that the machine is the same size as a standard enterprise server means it can easily be installed on-site at all types of clients (industry, government, and research/academic institutions).

Atos Codex, Atos’ analytics platform, is an integrated end-to-end analytics solution which incorporates predictive computing and cognitive analysis, enabling customers to collect, analyze and share data. Atos Codex incorporates high performance analytics, and has pre-loaded business use cases ready to implement. As an example, Atos Codex has helped T-Mobile Australia identify (through demographic, geographic and financial attributes) fraudulent customers in their infancy, resulting in lower potential revenue losses for T-Mobile Australia. Figure 56 below reflects how Atos Codex operates within a utility company’s data framework.

Stacy PollardAC

(44-20) 7134-5420

[email protected]


Source: Atos.

77



Figure 56: Atos Codex

Source: Company reports.

Aveva (AVV LN, OW)Aveva sees opportunities to utilize Artificial Intelligence to analyse real-time big data sets, especially in the area of predictive maintenance. For example, in a plant environment, sensors (such as vibration or temperature sensors) collecting real-time data can use pattern recognition to determine maintenance needs of various equipment. Matching real-time data to the equipment’s digital twin could also allow for automated replacement parts to be ordered and repairs to be scheduled for mechanics or robots. Furthermore, the feedback loop could be implemented into future design iterations.

We spoke to Aveva’s CTO Dave Wheeldon, and he thinks about implementing AI within the framework of the scientific model to develop theories about the world: designing experiments, running simulations, analysing outcomes and designing models to predict empirical evidence (then continually test and refine based on real-world sensory data).

Overall, our view is that full AI is still in the development phase at Aveva, and not yet a routine advanced feature of its CAD or products. However, we hope/expect to see much more with the integration of Schneider Software.

Baidu (BIDU US, N)

Baidu’s AI strategy comprise of 4 major projects:

Autonomous driving platform ‘Project Apollo’,

Baidu cloud,

Conversational-based AI smart device platform ‘DuerOS’.

Internet finance

After attending Baidu’s two AI events in July/Nov, we are incrementally more positive on Baidu's AI strategy as the development of AI has moved from lab research stage to product commercialization stage. Nonetheless, we believe large scale commercialization and product adoption of Baidu's AI strategy seems still a

Stacy PollardAC

(44-20) 7134-5420

[email protected]


Alex YaoAC

(852) 2800-8535

[email protected]


78



couple of years ahead of us at least, let alone monetization. We believe market sentiment on Baidu will improve post the event, however, the stock might continue to trade as a search engine until it gains traction in AI adoption volume

Exporting autonomous driving solutions through ‘Project Apollo’. In April2017, Baidu announced Project Apollo, which opens complete autonomous driving solution to outside partners. Key technologies opened so far include environment awareness, route planning, automobile control and in-car operating system, which cover four major components of autonomous driving: auto, hardware, software and cloud based data service. Baidu now has 1.7k+ eco-system partners for ‘Project Apollo’, which includes OEM, map technology companies, chipset developers, private equity fund, government agencies and research institutes, with large auto companies such as overseas car brands Ford/Daimler AG, Chinese car brands Chery/Greatwall/Changan and car equipment developer Bosch/Continental. In addition, Baidu cooperates with King Long Motor Group, a leading bus manufacturer in China to achieve mass production of the first autonomous driving bus in July 2018 (There will be no wheel/driving seat on the bus).

Baidu Cloud: AI based cloud supported by ‘Baidu Brain’: Baidu announced its cloud strategy, which comprise of ‘ABC’: AI, Big data and Cloud Computing.Compared to Alibaba/Tencent, Baidu Cloud mainly focus on AI based SaaS service (voice recognition, natural language processing (NLP), user portrait and visual recognition), supported by Baidu’s core AI engine ‘Baidu Brain’ (similar to Google Brain). Since 2016, Baidu has provided over 80 core AI capabilities of Baidu Brain to its partners in the form of API. Over 370k partners have joined open platform of Baidu Brain and use various capabilities of Baidu Brain every day. According to Baidu management, daily request for Baidu Brain API exceeded 219b recently.

Conversational-based AI smart device platform ‘DuerOS’: DuerOS is an interaction system that enables devices to communicate with users in voices. The system is an integrated solution of software and hardware that is developed by Baidu and partners. Baidu believes that high quality voice recognition, semantic recognition (understand what user want) and service breadth (satisfy user demand) are the key elements for a good AI system. DuerOS has a wide range of partners, including hardware manufacturers such as Sharp/Sony/Meizu; smartdevice sales channel such as JD/Gome; component developers such as Qualcomm/Inspur and content provider such as iQiyi. Baidu also launched 3 new smart hardware devices as reference designs in the recent Baidu World in Nov 2017:

Raven H, which is an intelligent speaker (similar to Amazon Echo) based on Baidu’s mobile operating system DuerOS 2.0 with a detachable voice input remote control (Price at Rmb1,699). Raven H enables users to search for sports/music/lifestyle related content via voice input.

Raven R, which is a robot-like intelligent speaker with all function of Raven H, and able to stimulate human gesture via voice input.

Raven Q, which is a robot with a series of Baidu core AI technologies, such as facial recognition, visual recognition and autonomous driving.

Internet finance

Baidu’s core AI application in internet finance includes: 1) smart investment advisor, supported by its rich user data and algorithm, Baidu is able to provide tailor-made

79



financial service to each user. 2) identity recognition, supported by Baidu’s capability in visual/voice recognition and OCR (Optical Character Recognition). 3) big data risk management, supported by Baidu’s rich user behavior data and machine learning capability. 4) smart customer service provided by AI based chat-bot.

Baidu’s products of AI-based Internet finance include:

Baidu Wealth Management. Baidu aims to provide fairer, more creditworthy and more transparent wealth management service to investors by leveraging strengths in big data and technology. For its customers, Baidu analyzes their risk profiles using its own artificial intelligence technology. For investment products, Baidu uses Baidu data analytical capabilities to more holistically analyze their potential risks and returns. Furthermore, Baidu has established a team of experienced wealth management professionals to serve its customers.

Baidu Consumer Credit. Baidu Consumer Credit offers education loans and consumer financing in industry sectors such as travel, beauty, home decoration and home rentals, through partnership with a large number of educational institutions and other companies and merchants. Baidu is creating an innovative platform to provide internet financial services, which give its users more convenience and faster approval, with the help of its AI-based risk control technologies including facial and fingerprint recognition, optical character recognition (OCR) of identification documents, and live detection.

Barracuda Networks (CUDA US, OW)

The messaging security that the company provides through both on premise appliances and in the cloud utilizes what the company labels deep machine learning to identify and protect against malware. Especially against modern polymorphic malware that has been tougher to detect over the last couple of years.

Box (BOX US, N)

Box is using AI to automate the process of tagging and categorizing content within photos, audio or video files. In October 2017, Box announced Box Skills, which comes with three out of the box skills: 1) Image intelligence (powered by Google), 2) Audio Intelligence (powered by IBM Watson) and 3) Video Intelligence (powered by Microsoft Cognitive Services). It also announced Box Skills Kit which is an SDK for customers to build their own skills and supports cognitive services APIs from IBM Watson, Azure and Google to power these skills. Box also announced Box Graph recently, which is an intelligent network of content, relationship and activity enriched by Machine Learning techniques and governed by Box’s built in security and permissions. The first experience powered by Box Graph is called Feed, whereby every user can see content recommended to them based on their activity along with trending content popular within their company.

Capgemini (CAP FP, OW)

Capgemini aims to introduce and implement AI-based innovation to its clients where it serves a specific business purpose. Capgemini sees the process of learning, developing and preparing in this area as key to developing an internal ecosystem of knowledge around Artificial Intelligence, Machine Learning, IoT and Smart Services. The know-how will support go-to-market strategies and enable CAP to further engage with both existing and new clients in these areas.


(1-212) 622-6389

[email protected]


Mark MurphyAC

(1-415) 315-6736

[email protected]


Stacy PollardAC

(44-20) 7134-5420

[email protected]


80



Capgemini’s Insights and Data team (I&D), with know-how in AI, ML andpredictive analytics, are working with vendors and customers looking to implement these technologies. Recent projects include using cognitive technologies (IBM Watson Explorer) at Akershus Hospital, University of Oslo, to understand the cumulative impact of radiation on cancer in children as well as leveraging Big Data and predictive analytics to produce a 360 view of customer profiles for French retailer Kiabi.

The Applied Innovation Exchange, Capgemini’s network of innovation hubs (ten centres worldwide) connects startups, partners and universities, to provide an environment for customers to test technologies and develop pilot projects. The network of centres aims to adapt innovation to the specific sectoral needs of each client. As an example, a virtual fashion assistant for a European retailer was created, as well as shelf analytics, visitor management and virtual store solutions for separate retailer.

Capgemini announced (in February last year) it will be partnering with Celaton, and using its inSTREAM software, a cognitive learning platform. Celaton’s technology is able to process structured, semi-structured and unstructured information, and learn from the data. The platform can minimise the need for manual processing and aims to ensure that only accurate, relevant and structured data enters line of business systems. Through processing and monitoring the actions and decisions of human operators and referring ‘exceptions’ to them, inSTREAM is able to continually improve. The solution is able to capture, extract and verify data, as well as determine outcomes (using previous experience, knowledge and business rules), archive and monitor.

Cloudera (CLDR US, OW)

Cloudera is positioning its Spark platform for Machine Learning workloads. Spark is an in-memory based massively parallel processing engine, making it suitable to run the iterative, compute heavy machine learning algorithms. We estimate that about 700 Cloudera customers are today using Spark with 30-50% of them using Spark for Machine Learning use cases. Additionally, Cloudera recently launched the Data Science Workbench, which is a web application that allows data scientist to collaborate and build Machine Learning models. In September 2017, Cloudera also announced its acquisition of Fast Forward Labs, a startup that is an expert in Machine Learning and helps customers to choose the right Machine Learning algorithm or framework based on the use case.

Cognex (CGNX US, UW)

Cognex is a leading supplier of machine vision products, including barcode readers, machine vision 2D and 3D sensors used in factory automation, logistics, life sciences and other applications. The company acquired ViDi Systems in April 2017, a maker of deep learning software for industrial machine vision. The software uses AT to improve image analysis and trains the systems to distinguish between acceptable and unacceptable variations and defects in high-scale manufacturing. Cognex offers three ViDi tools for use with machine vision solutions: Fixturing (find and localize features in an image), segmentation and defect detection (e.g., scratches on a finished surface), and object and scene classification. CGNX has been growing rapidly, its machine vision technology now features prominently in CE manufacturing, including production of Apple’s iPhone X OLED screens.

Mark MurphyAC

(1-415) 315-6736

[email protected]


Paul Coster, CFAAC

(1-212) 622-6425

[email protected]


81



Cognizant (CTSH US, OW)

CTSH’s AI solutions are largely a part of its analytics practice, and are applied across its Digital Business, Digital Operations and Digital Technology. CTSH primarily focuses on deploying AI technologies in B2C applications like trade processing, pharmacovigilance, consumer interface etc. across banking, insurance, retail and healthcare verticals.

CTSH primarily deploys AI through its Intelligent Process Automation practice:

Intelligent Product Practice

Intelligent Automation Solutions. CTSH applies AI and Machine Learning technologies to standardize repeatable tasks. Applications across mortgage document processing, insurance claims processing, medical claims management. CTSH robots based on its own automation technology as well as leverage partnership intelligent automation solutions.

Dynamic Workflow Management, applies image recognition, artificial intelligence, machine learning, and process automation to modernize and automate workflows. Applications across loan origination, servicing, and support functions workflow and claims processing.

Artificial Intelligence-enabled Robotic Investment Advisor. CTSH deployed the platform across a leading BFS client, to enable the client achieve its goal of doubling retail assets under management by 2020

Additionally, CTSH has opened global ‘Collaboratories’ across NY, London and Amsterdam to deploy AI (and other digital) applications.

Notable Industry Collaborations. 1) CTSH has collaborated with Microsoft and Amazon to deploy the latter’s AI and Machine Learning tools to client specific use case and demand. 2) Separately, the company has also collaborated with Artificial Solutions, a specialist in Natural Language Interaction (NLI) to integrate its Artificial Solution’s Teneo’s platform with CTSH’s consulting, SI and support expertise to offer enterprises AI-based conversational channels to interact with customers and employees.

Cornerstone OnDemand (CSOD US, N)

Cornerstone acquired Evolv ($43M in Oct 2014), a machine learning and data science platform for helping customers make workforce decisions, e.g. analyzing data to evaluate skills, experience, and personalities for jobs. In Sept 2015, Cornerstone announced its Insights solution based on Evolv, which predicts risks from employees who haven’t finished a training course, for example, and offers predictive succession planning.

Coupa Software (COUP US, N)

In October 2017, Coupa acquired assets from Deep Relevance, which uses AI-based behavior and relationship analytics to help detect employee and collusive fraud, or other suspicious transactions. The acquisition will help customers create fraud profiles based on analyzing customer and aggregated community data for expenses, purchase orders, and invoices. This profile score and related spend transactions can

Mark MurphyAC

(1-415) 315-6736

[email protected]


Mark MurphyAC

(1-415) 315-6736

[email protected]



(1-212) 622-6632

[email protected]


82



then be used to alert a company’s internal auditors or finance department for further review and action.

Dassault Systèmes (DSY FP, OW)

In our view, Dassault is ahead of the curve with regard to AI. The company has been integrating machine learning into a number of its core functions for years. As early as 2001, Dassault introduced Knowledgeware for PLM into CATIA, an application incorporating AI. Today, AI is truly embedded into Dassault’s 3DEXPERIENCE Platform and even into the data model.

As we look across Dassault’s solution portfolio, more than 80% of products run on the 3DEXPERIENCE platform; a collaborative and unified platform with embedded machine learning (ML) & artificial intelligence (AI). The company is at a stage now where re-programming and optimizing algorithms is part of the regular routine. Arguably, ML & AI is built into the way the company imagines and codes its products, indeed, into the very fabric of its business model. The company’s tag line “If We” even alludes to AI, in a sense.

The edge for Dassault lies in its deep industry knowledge. The company has extensive expertise in Transport, Aerospace, Industrial Equipment (and other sectors), as well as the associated characteristics/specifications of the digital format.

Last year Google X (subsidiary of Alphabet Inc.) chose to partner with Dassault, taking the EXALEAD product, which specialises in big data search of structured and unstructured data. Dassault was able to offer unique contextual understanding of engineering, components and material physics – which Google could not do.

Below we show how Dassault is using ML & AI in its current product set:

CATIA – Dassault’s flagship 3D Computer Aided Design (CAD) software; powered by the 3DEXPERIENCE Platform and tailored to manufacturing organisations. In terms of embedded AI, Design Robots within CATIA continuously analyse what the engineer is doing, predicting what elements might be useful and offering a menu of best design options. CATIA's algorithms are constantly incorporating data (product specification, material availability, input constraints and so on) and using this information to generate designs, as well as suggesting alternatives based on input requirements, viability of materials, etc.The software can also do this at the level of the supply chain – meaning the engineer is encouraged to design using parts which can be sourced (rather than doing design rework when supply shortages are discovered later in the process).

SIMULIA – realistic simulation, finite element analysis, multiphysics and multiscale. Think assessing the safety and reliability of products and materials (e.g. ship interacting with the water) before bringing them to physical prototypes.

DELMIA – Dassault’s digital manufacturing, virtual twin and supply chain & operations solution. Within DELMIA, the Operations Intelligence program is basically pure AI. Its algorithms tap into operational research, factoring in a variety of specifications, from air temperature, to viscosity, material strength, etc. The software's Predictive Analytics solutions also enable real-time analysis and maintenance within the industrial environment. For example, in case of operational disruption (e.g. faulty machine or supply chain constrains), shop floor processes can quickly be re-scheduled or adopted to assure utilization and efficient use of resources.

Stacy PollardAC

(44-20) 7134-5420

[email protected]


83



Quintiq – offers supply chain planning and optimization to help customers adapt to the constant technological, regulatory and environmental changes. The supply chain of the future will be connected, intelligent and sophisticated. Quintiq's solution combines predictive analytics, optimization and cloud-based services on a single system to enable an end-to-end planning experience. The software helps plan and optimize workforces, logistics and production. For example, Quintiq software was used by DHL to develop a dynamic route planning solution that helped the company reduce tour duration by 8% and mileage by 15%. Quintiq also helped Brussels Airport reduce the number of complaints from airlines by 90%.

Information Intelligence Apps - EXALEAD has the capability to apply advanced semantic processing to web-scale data. The solution relies heavily on machine learning and AI. EXALEAD OnePart, a sourcing and standardization intelligence application within the suite, indexes product part information from different sources, sites and projects, and then reconciles, analyses and interprets this information (with advanced semantic technology) in order to generate insights used for prototyping, developing and testing.

In the not too distant future…

Life Sciences. Dassault is already active in this space, for example The Human Heart simulation is an example frequently referred to. When you start to design and simulate virtual drugs that don’t yet exist, on virtual human cells or organs, in order to treat a disease that doesn’t exist, you can see the potential. This type of simulation can drastically reduce costs, testing time (and animal testing) and overall time-to-market.

Virtual Twin / Virtual Worlds. These are becoming more and more precise, and AI helps to test new scenarios in a simpler, cheaper and faster way. When you know you have the capability to regenerate, in a few hours, a different shape of car, a different model specification, a modified engine, and then run it again on anew set of roads which are now wet from the virtual rain (to test the tyre traction), again, you can sense the vast potential of the software. You can simulate every aspect involved in the development of any product, process or system before even thinking about creating a physical prototype. The virtual world has the potential to generate an immense degree of innovation and cost savings. Thanks to enhanced computing power, extensive availability of data and more advanced algorithms, Artificial Intelligence is able to enhance the virtual experience.

Generative Design for 3D printing. Software can be engineered to define a set of constraints. When you combine the ML & AI with these constraints andoperational specifications – it can compute a new shape, using the laws of physics (many ideas of which humans couldn’t comprehend or compute on their own).Engineers can then make small tweaks. Though Dassault will do the generative design part; based on time, cost, size, weight – and compute and compute until it creates a structure which is aligned to the constraints that have been set. Once design, construction, simulation, evaluation and creation from an operational perspective have been completed, Dassault’s software will then be embedded into the machines. Its software will be able to generate software that will be pushed to the robot.

84



Face++/Megvii (Private company)

Established in 2012 by Mr. Yin Qi, a PhD from Columbia University with working experience in Microsoft, Face++ engages in the development of deep-learning based computer vision technologies. In 2017, the company completed Round C financing of US$460 mn, led by China Venture Capital Corporation, Foxconn and Ant Financial.

Face++ has more exposure to financial industry, compared to other players. Its face recognition solution has been incorporated into Alipay, the largest online payment platform in China, for account login and password resetting. Unlike Apple's FaceID, the solutions from Face++ and other competitors are 2D-based by using selfie camera to capture the facial information without mesh modeling. Therefore, they are not qualified to replace password, and have to work with other authentication methods, e.g. input the registered phone number, to complete the online payment. Face++ is actively working with its partners on 3D solutions.

Face++ also developed vision systems for robotics. The service robots equipped with Face++'s solution have been adopted by Bank of China to perform some preliminary client services such as VIP recognition and client guidance.

Face++ has over 250+ R&D employees with a self-developed database of 120mn images of faces and other objects to train the algorithm. Total revenue was above RMB100mn in 2016, up 500% yoy.

Figure 57: Face recognition in Alipay

Source: Company data

Figure 58: Vivo Face Access with Face++ solution

Source: Company data

Facebook (FB US, OW)

Mark Zuckerberg considers AI to be one of the company’s 10-year bets, and believes that AI should replicate—and exceed—human senses such as vision and hearing so that FB is better able to serve users. Facebook has been using AI in photo-tags for facial recognition since 2010. In 2013, FB created its Facebook AI Research (FAIR) team to focus on developing and using AI. In 2015, FB launched its Applied Machine Learning team (AML) to infuse AI into FB products and focus on language tech, core machine learning, computational photography & perception. FB has developed a dedicated Machine Learning Platform “FB Learner Flow”, which has been used for products within Search, Ads and NewsFeed. FBLearner Flow spans

Doug AnmuthAC

(1-212) 622-6571

[email protected]


Not Covered

85



several areas, including infrastructure, algorithms, and applications built on top of the platform. In 2015, FB acquired Wit.ai, a company that builds developer tools for NLP, to integrate NLP tools into front-end of the Messenger Platform.

Deep Text, In June 2016, FB announced DeepText, a deep learning-based AI engine built to understand sentiment and context behind text on the social network. FB is using DeepText on its News Feed to understand the content in thousands of posts per second across more than 20 languages. DeepText is also being used in the Messenger app to detect intent of text as well as to power AI chat bots. DeepText has the potential to better understand posts and can extract intent, sentiment, and entities, using mixed content signals like text and images, and can automate the removal of objectionable content like spam. This deep learning process would allow FB to understand people's interests better and help it in recommending content.

AI on NewsFeed. FB created a Language Tech team as part of its FAIR group in 2013 to tackle problems related to speech recognition, natural language understanding and language translation. The team develops ML models to train the algos that power FB’s NewsFeed within hours, using trillions of data points.FB updates its learning models every 15 mins to 2 hrs so that it can react quickly to current events. Machine learning models rank ads, personalize content on News Feed, filter out offensive content, highlight trending topics, rank search results etc.

“M” suggestions on Messenger, In April 2017, FB rolled out “M”, an AI-based assistant embedded within the Messenger app to its US users and since has been released in half a dozen countries. M Suggestions uses deep learning to understand context and intent in a chat and comes up with suggestions including send stickers, pay/request money, share location, make plans, get a ride—option to book Lyft/Uber, save content, initiate voice/video calls, sharing GIFs, quick replies etc. FB aims to transform Messenger into a conversational platform to power Chatbots within the Messenger framework. These bots can handle customer-service requests queries and similar requests.

Moments, Moments, FB’s photo-sharing service, uses FB’s image recognition models to let users create private photo albums with a select group. At Moments' launch, FB revealed that its image recognition models could recognize human faces with 98% accuracy, adding that the algorithm could identify a person in one picture out of 800M in less than 5 secs.

Other efforts include Translations, Speech recognition, Photo Image Search, Real-time video classification, etc. FB has built an AI-based automatic translation system that helps people to see translated posts in their NewsFeed. In Aug 2017, FB announced that its AI models are processing 4.5B+ translation requests daily. FB also plans to incorporate speech recognition more deeply into its main app within the next five years and make it possible to give the app voice commands to conduct searches or complete activities like posting a group photo, removing red-eye or talking to a virtual assistant etc. FB introduced automatic image classifiers that allow users to search for photos based on the image content instead of relying on tags or surrounding texts. Similarly, FB is working on applying computer vision techniques to organize content and provide the ability to classify live videos in real time.

86



FireEye (FEYE US, N)

This is a good example of utilizing machine language approaches to security intelligence. The company’s Central Threat Intelligence product conducts over 14 million analyses per hour using machine language approaches. In addition, the company has adopted the machine learning approach inside of SmartVision, a module that can be utilized within the company’s network security (NX) solution to monitor and detect attacks that are looking to spread laterally through an organization using what is called east-west traffic.

Fortinet (FTNT US, OW)

Similar to others, Fortinet is utilizing machine learning capabilities to drive the intelligence component (FortiLabs) that supports its Fortinet fabric which aims to tie together the product portfolio into more of a platform approach.

Hexagon (HEXAB SS, N)

As building blocks for its analytics platform, Hexagon is investing it is own proprietary specific algorithms and machine learning frameworks that run on high performance dedicated hardware (such as NVIDIA). The following main application areas have been addressed by pilot projects/products:

Object/pattern detection in images (Agriculture, Mining, Construction, O&G, Metrology).

Video analytics (Public Safety, Digitalisation and Classification applications, etc).

Automated point cloud processing (various applications for object recognition, tracking and context interpretation)

Predictive maintenance (Mining, Agriculture, Construction and general machinery)

Productivity analysis & optimisation for many applications throughout Hexagon's vertical solutions.

In addition, Hexagon has developed its own IoT platform (Connectivity, Advanced Analytics, Mobility and Visualisation). Hexagon has application-tuned advanced analytics modules using ML and DL techniques that:

Run on the edge, cloud and/or on-premises (e.g. For Public Safety devices and Control Rooms, Autonomous trucks, Scanning and Metrology devices, etc).

Can be embedded into Hexagon’s middleware Edge Frontier (e.g. anomaly detection, predictive models, etc).

Enable cloud and edge analytics (e.g. for condition monitoring, predictive maintenance, fleet management, etc.).

HubSpot (HUBS US, OW)

HubSpot has begun to speak about AI/ML in a more meaningful way, since the introduction of GrowthBot (an AI chat bot) from its experimental HubSpot Labs. As the company devotes more resources into R&D, the conversation has broadened beyond chat bots, and into use-cases such as lead scoring for its Sales and CRM


(1-212) 622-6389

[email protected]



(1-212) 622-6389

[email protected]


Stacy PollardAC

(44-20) 7134-5420

[email protected]


Mark MurphyAC

(1-415) 315-6736

[email protected]


87



products. HubSpot acquired Kemvi (July 2017) for sales research, and Motion AI (Sept 2017) a visual chat bot builder.

IBM (IBM US, N)

IBM arguably kick-started the latest AI focus (aka Cognitive) with Watson, first introduced in 2011. The company focuses on offering Watson capabilities and solutions to enterprises only, unlike many other product companies (such as Apple, Amazon, Google) that offer solutions for consumers as well as enterprises.

IBM’s Watson solutions represent a set of comprehensive and adaptable intelligence systems with applications ranging from oncology to customer support using open APIs. Watson is configured to analyze both, clients’ in-house as well as public and IBM proprietary data, which we believe differentiates IBM from a lot of competitors. IBM, through its recent acquisitions of The Weather Channel, Merge etc. is quietly building its proprietary data that it believes can provide higher value to customers. IBM offers Cognitive solutions across healthcare, education, BFSI, media, and retail verticals, amongst others, based on the following products and services:

Conversational Chatbots. Virtual agents across a variety of channels, including mobile devices, messaging platforms. Chatbots built at scale for enterprises providing automated services.

Data discovery. Interpret data to monitor trends and surface patterns using cloud-native insight engine. Complex architectures can also be built using advanced AI functions such as natural language queries, passage retrieval, relevancy training, relationship graphs and anomaly detection. Data discovery can also be infused with news search engines.

Natural Language Understanding and Processing. Analyze text to extract meta-data from content such as concepts, entities, keywords, categories, relations and semantic roles across 9 languages and return both overall sentiment and emotion for a document.

Watson Knowledge Studio is a cloud-based application that enables developers and domain experts to collaborate and create custom annotator components for unique industries by analyzing mentions and relationships in unstructured data.

Visual Recognition. Tag, classify and search visual content using machine learning. Custom classifiers using own image collections can also be applied.

Speech. Convert both audio and voice from 7 languages into written text in real time. Can also convert written text into natural-sounding audio in a variety of languages and voices.

Language Translator and Classifier. Translate text from one language to another and also be able to classify it. Take news from across the globe and present it in desired language, communicate with customers in their own language.

Personality and Tone Analyzer. Predict personality characteristics like customers’ habits and preferences through written text. Also analyze emotions and tones in what people write online, like tweets or reviews (social listening). This service can also be integrated with chatbots.

In addition to vertical specific solutions, IBM also established AI solutions for public sector clients, notably including defense logistics solution and law enforcementsolutions (for public safety and policing).


(1-212) 622-6632

[email protected]


88



Separately, IBM Watson has specific horizontal applications solutions like Customer Engagement, Recruitment & Talent acquisition, Employee engagement etc.

Notable Industry Collaborations. 1) Integrate Salesforce CRM platform with Watson's AI engine 2) 10 year partnership with MIT to research on AI 2) With SAP on AI in procurement 4) with Twitter to analyze tweets relevant for enterprise clients. Similar to Alexa and other consumer focused platforms, IBM’s platform allows developers to build use cases/applications on Watson that they can offer to consumers/enterprises.

Notable Current R&D projects in AI. 1) Include Watson Beat - Machine Learning algorithm that learns to create music. IBM researchers are teaching a complex neural network to understand music theory, structure (pitch, time signature, and key signature), and emotional intent and co-create music with a human partner). 2) Gesture and Visual Recognition (for applications across Medical Image Analysis). 3) AI-based supercomputing system in collaboration with U.S. Air Force Research Laboratory (AFRL). 4) AI-based Cognitive Assistant for visually impaired 4) AI for Cancer detection - to analyze large amounts of imaging and text in electronic health records (EHR). 5) Block-chain applications.

Cognitive in GBS (IT Services business)IBM is also helping clients understand goals from AI deployment, train the “robots”, and help implement the platform. The company also provides services such as application automation, repetitive task automation (RPA, helpdesk/testing assistants), and Cognitive analytics services to its clients.

iFlytek (002230 CH, Not Covered)

iFlytek was established in 1999 by a group of speech scientists and engages in the development of algorithms and applications on NLP.

iFlytek’s core NLP technologies include 1) speech synthesis engine, to convert text to speech; 2) speech recognition engine, to recognize the speech of human beings (including dialects) and convert it into text; 3) voiceprint recognition engine, to identify the speaker based on the voice waveform. The algorithm compares the voice of the speaker with the voiceprint registered in the database and determines whether they match or not; 4) cloud platform, to provide SaaS-based NLP solutions.

Education is the key sector to monetize the AI technology, with current revenue contribution of 30%. iFlytek offers a comprehensive smart campus solution which could 1) customize study plans for each student based on the data analysis of their historical performance; 2) evaluate the test by machines through the recognition of handwriting; 3) improve the efficiency of English and Chinese teaching through interactions between students and machines.

Other verticals include healthcare (voice health record) and court (convert speech to text to replace court clerk).

Not Covered

89



Figure 59: iFlytek 1H17 revenue breakdown

Source: iFlytek

Figure 60: iFlytek’s smartphone applications

Source: iFlytek

Imperva (IMPV US, N)

The CounterBreach solution focuses on how users interact with enterprise data and utilizes machine learning to detect insider attacks one of the most difficult items to protect against, but one of the most prevalent sources of cybersecurity issues. The goal is to use the machine learning to identify inappropriate access to data that could lead to data breaches.

Indra (IDR SM, N)

Indra’s AI capability is incorporated into its platform iThink; a solution designed to manage the intelligence life cycle in a global way, with a particular emphasis on security. Indra's platform underpins a set of modules that enable data aggregation and capture from structured and unstructured data sources, as well as the subsequent analysis and representation of the data. Figure 61 below reflects the components of Indra’s iThink stack.

Customer service/interaction

software15%

Education30%

Internet applications12%

Public security and smart city

6%

Big data and services

9%

System integration28%


(1-212) 622-6389

[email protected]


Stacy PollardAC

(44-20) 7134-5420

[email protected]


90



Figure 61: iThink Technology Stack


Technological Platform – which is composed on a set of modules, and the integration of commercial tools.

Data Acquisition – the area of iThink that captures information. This is done through a web crawler which processes and categorises information from structured and unstructured sources. Manual data entry and also a content translation module support this process.

Information Structuring – locates entities in documents and structures their content, identifying relationships. Assisted analysis & visualisation, through graphical user representation, helps to structure the information before it can be interpreted and reflected in a visual way.

Knowledge Management – segmentation of the information, with the capacity to correlate data and create meaningful information. Geographic representation (e.g. spatial zones, influencer areas) is supported here.

Intelligence Production – identification of patterns and trends through the statistic module. The monitoring module supports the assessment of warning indications is able to assess the severity of an identified issue. The analysis module (which uses semantic technologies) also helps to make sense of the data.

Indra also engages in innovation projects. A recent example is its involvement in the European BeCamGreen project, which aims to develop a detection solution based on computer vision and big data. The solution will primarily be used for monitoring traffic – e.g. classifying vehicle types, identifying the number of occupants a vehicle contains and consequently helping to identify mobility patterns.

The solution will help policy makers and traffic monitors identify patterns and administer strategies (such as access restrictions dependent on number of passengers, vehicle type, license plate number etc.) aimed at reducing traffic congestion and incentivising the use of public transport where warranted. The solution leverages computer vision technology, deep learning & multispectral analysis (used in detecting human skin), as well as combining a big data engine to identify and predict traffic situations by utilising data in real-time. We view this solution as the type of application that could be used in digital cities of the future.

91



Micro Focus (MCRO LN, N)

Following the $8.8bn merger with HPE’s software business, we look at the combined entity’s AI capability, which stems mainly from assets previously part of HPE Software’s portfolio.

IDOL – acquired by HPE software, but originally from Autonomy, the IDOL platform is capable of supporting and analysing approximately over 1,000 formats of data. Built on machine learning & deep neural network algorithms,IDOL is aimed at searching through and rendering insights from audio, visual, video and text based data to deliver predictive information.

Figure 62: Typical use cases for IDOL


Vertica - Micro Focus has recently released Vertica 9 (Vertica was formerly HPE Software), the latest version of its enterprise level, big data analytics platform with embedded in-database machine learning capabilities. Vertica supports the entire predictive analytics process with massively parallel processing architecture and end-to-end machine learning management functions. Vertica 9 is able to analyse large volumes of unstructured data and generate insights. The platform is also intended to simplify the creation and deployment of machine learning models.

Stacy PollardAC

(44-20) 7134-5420

[email protected]


92



Figure 63: Vertica Architecture


Microsoft (MSFT US, OW)

Microsoft offers the Microsoft AI platform, which is an overarching frameworkenabling organizations to create their own intelligent tools. Microsoft AI platform consists of three main components: Microsoft Cognitive Services, Microsoft Cognitive Toolkit and Microsoft Bot Framework. The Cognitive Services component includes Vision (image processing algorithms that can identify the content of an image), Speech (Speech recognition, Speech-to-text), Natural Language Processing and more. The Bot Framework provides an integrated environment for bot development that allows customers to build bots quickly using templates. Microsoft also allows bots to interact with the cognitive services, to further boost its intelligence. The Cognitive Toolkit allows customers to harness the intelligence within large data sets through deep learning. Microsoft also provides a Machine Learning Studio which is a cloud based visual drag-and-drop authoring environment that allows users to build, deploy and share predictive analytics solutions.

NICE Ltd. (NICE US, N)

Nice Systems is a leading supplier of customer experience, workforce management, and cloud-hosted contact center solutions. The company’s Actimize division also positions Nice as a leader in software-based solutions that fight financial crime, money laundering and ensure regulatory compliance. With the acquisition of Nexidia in early 2016, the company has developed neural network technology designed to perform large-scale speech recognition and analysis, turning vast amounts of unstructured customer interactions (captured in the contact center) into data and actionable intelligence. In the case of contact centers, this technology (customer experience analytics) enables increased sales effectiveness, reduces customer churn, and ensures compliance with regulation and policy.

Mark MurphyAC

(1-415) 315-6736

[email protected]


Paul Coster, CFAAC

(1-212) 622-6425

[email protected]


93



In addition, NICE is also introducing robotic process automation (RPA) into low-value, high-volume (back office) workflows, typically found in the financial services industry. Attended robots are digital assistants that collaborate with analysts in completion of evidence gathering in case management. Unattended robots are used to gather data from legacy systems and update enterprise systems. We believe NICE’s differentiation will hinge upon combining analysis of unstructured speech data with traditional structured data to address more difficult process challenges, ultimately in support of cognitive decision management (using data mining, pattern recognition and cognitive computing).

Finally, in September, NICE introduced an “Open Framework Ecosystem for Cognitive Robotic Automation”. This platform combined third-party optical character recognition, chatbots, voicebots and machine learning technologies with NICE’s in-house text analytics and natural language processing to enable customers to expand and improve customer self-service and process automation. For example, NICE’s platform provides real-time customer data to chatbots enabling execution of bespoke requests.

Oracle (ORCL US, OW)

Apart from using AI and Machine learning across many parts of its cloud applications, Oracle is now using AI and machine learning to automate administration of its latest Oracle Database 18c making it a self-driving or autonomous database. This new database is expected to automatically provision, upgrade, patch and tune without any downtime, eliminating a lot of human labor as well as human error. Oracle recently also announced a Machine Learning and Advanced Analytics platform as a cloud service thus competing with the likes of Cloudera and Microsoft.

Palo Alto Networks (PANW US, OW)

The company has focused the use of machine learning into its endpoint protection solution called Traps. In fact, the use dates back prior to the acquisition of Traps andhas been enhanced through the acquisition of LightCyber earlier in 2017. There are several private companies that have focused on the use of machine language for endpoint protection to gain market share against entrenched legacy providers like Symantec and McAfee. Similar to FireEye, the company is also using the approach in its cloud-based intelligence and analytics area as a foundation for the log collection and analysis it is providing to customers.

Mark MurphyAC

(1-415) 315-6736

[email protected]



(1-212) 622-6389

[email protected]


94



Sage (SGE LN, OW)

Within its own products, Sage is looking to develop machine learning and AI based products, tailored for the purpose of reducing financial administration for business builders, from start-ups to enterprise.

Sage’s AI capability is currently best reflected through Pegg, an accounting Chatbot, which launched at the Sage Summit in July 2016. Users are able to log an expense or income, add receipts and Pegg will record (through messaging applications like Facebook messenger and Slack) the information for them. Pegg also uses AI to recognise nuances and patterns in the way its user operates their business, and so through more interactions with Pegg, the more the bot can learn. Pegg tailors itself to the user and can also provide recommendations.

salesforce.com (CRM US, OW)

Created in September 2016, Salesforce Einstein is a layer within the Salesforce platform that infuses Artificial Intelligence features and capabilities across all Salesforce Clouds. Einstein takes care of the data prep, modeling, and infrastructure needed to embed and scale predictive models throughout customers’ Salesforce workflows. Since its inception, salesforce has been expanding the application of AI across its various modules – Sales, Service, Marketing, Community, Analytics and Platform. In the Sales cloud, Einstein can automatically prioritize leads by scoring them based on history and past deals as well as showing the likelihood of a closure for a deal based on customer sentiment. In Service Cloud, Einstein can recommend responses to an issue, automate case classification or predict close case resolution time. Einstein enhanced the Marketing Cloud by optimizing email send times as well as predicting audience segment that share common traits, while it added Individual Product recommendations in the Commerce Cloud. Recently, salesforce also introduced myEinstein, which is a declarative platform to enable customers to develop custom AI applications. myEinstein includes two new services – 1) Einstein Prediction which allows customers to create custom AI tools that can predict outcomes for any field or object in salesforce; and 2) Einstein Bots which can be used to support customer service workflows by automating tasks such as answering basic questions and acting as the first point of contact.

SAP (SAP GR, OW)

SAP wants to be known as one of the big players in enterprise-level AI/ML solutions, competing with the likes of IBM Watson, Microsoft Azure Cognitive Services, and Salesforce Einstein. Given SAP’s unique expertise in ERP and business applications, the company is in a strong position to embed AI into numerous functional applications (e.g., CRM, HCM, SCM, and certainly into analytics). Likewise, since SAP's software creates and evaluates much of the data generated within an enterprise, SAP also has a unique contextual understanding of what data has been collected, what it is used for, and how automation, analysis and various AI tools could enhance enterprise efficiency.

Embed, Stand-alone, and PartnerWe believe SAP will both embed AI functionality into its core S/4HANA suite AND offer stand-alone AI tools via Leonardo. In addition, SAP has taken a partnership approach to developing machine learning capabilities, instead of doing it all in-

Stacy PollardAC

(44-20) 7134-5420

[email protected]


Mark MurphyAC

(1-415) 315-6736

[email protected]


Stacy PollardAC

(44-20) 7134-5420

[email protected]


Sage Pegg - accounting chatbot

Source: Sage.

95



house. For example, SAP uses Google open-source project TensorFlow for some ofits machine learning algorithms and NVIDIA for the hardware to train these algorithms.

SAP S/4HANA already contains numerous built-in analytics and AI-type functionalities, and in the most recent release (version 1709) SAP has added additional ML functions, for example:

SAP Cash Application software (introduced with version 1709) learns matching criteria from a company’s historical financial clearing data to automatically pair invoices with incoming payments.

SAP Fiori apps for contract consumption (also introduced with version 1709) predict contract expiration or goods consumption ahead of time, which enables more effective supplier negotiations and relationships.

Other AI examples include:

SAP BusinessObjects Predictive Analytics Integrator –interacts with application data to create Predictive Models and generate procedures to enable the consuming application to generate predictions at runtime. No data is created or save within the integrator; the data management and security of application data and associated predictions is managed by the consuming applications.

Smart resume matching in Fieldglass.

Kore.ai Customer Service Bot for SAP Hybris is a “smart assistant” that enterprises can roll out for their customers to have a more human-like conversational interface for creating, managing and escalating their own service tickets. Ultimately, it automates much of the service ticketing process.

Logistics Optimization, using AI with supply chain and IoT technology. For example, Citrosuco of Brazil (supplier of orange juice) is using SAP Leonardo in combination with other SAP tech to connect farms to manufacturing site and optimize transportation logistic worldwide.

SAP Brand Impact, powered by NVIDIA deep learning, is able to analyse brand attributes (e.g. logos) in almost real-time and with high accuracy, enhancing the ability to assess marketing campaign effectiveness.

Leonardo innovation stack leads with AISAP offers much of its AI & Machine Learning capabilities through the Leonardo platform (named after Leonardo DaVinci), a broad-based digital innovation suite which allows customers to take advantage of emerging technologies like AI, ML, advanced analytics, IoT and blockchain. Leonardo is built on top of SAP’s open platform-as-a-service (PaaS) offering, the SAP Cloud Platform. Below we outline the Leonardo technologies and some of the capabilities they have to offer:

Machine Learning – SAP Leonardo Machine Learning enables the intelligence in SAP Leonardo by learning from data (deep learning). It embeds intelligence into existing enterprise applications and enables entirely new applications. For example, SAP’s Service Ticket (part of SAP Hybris) Intelligence application which processes inbound social media posts, e-mails, and other channel interactions by automatically determining classifications, routing and responses. SAP’s Customer Retention application can anticipate customer behavior (e.g.

96



renewals), through the use of transactional data and digital interaction points. The Leonardo Conversational AI Foundation can help customers build conversational applications using natural language processing and machine learning algorithms.

Big Data – collecting, processing and storing a wide range of data from both structured and unstructured sources, at scale.

Data Intelligence – connecting, aggregating and anonymising data to prepare it for commercial consumption. Data can be used on its own; it can be supplemented with partner, customer or industry data, and then offered out as Data-as-a-Service (DaaS), offering a new revenue stream. Alternatively, the data can be used to develop data-driven apps, or algorithms can be applied to provide insights.

Analytics – this area includes business intelligence, predictive analytics and performance management capability. SAP’s BusinessObjects BI platform, which enables clients to tailor predictive models to their specific business requirements, characteristics and functionality, is able to enhance insights and decision making.

IoT – Connecting devices, people and processes with back end systems in real-time. For example, SAP’s Vehicle Insights application allows the user to monitor live vehicle conditions and run, connected car analytics.

1. IoT Foundation - this component will typically exist in the cloud and will encompass big data applications, integration capabilities, as well as libraries of machine learning and AI models.

2. IoT Applications - all applications related to IoT, for example, in the connected products space, there would be supply network applications, inventory management applications, logistics safety apps and so on.

Edge Computing - an area of Leonardo entailing data capture at the edge. Through sensors and other communication pathways, data will be collated, aggregated and analysed in real-time. SAP’s acquisition of edge computing system, Plat.One, has enhanced SAP’s capabilities in processing IoT data.

Blockchain – early stage blockchain technology is embedded in the SAP Cloud Platform, and SAP’s blockchain-as-a-service (BaaS) pilot provides registered customers with a way to pilot the technology. It is important to note this technology is still very much in its early stages, though offering a channel to interact with it is important, particularly as adoption begins to pick up. To give a sense of blockchain potential, the World Economic Forum is forecasting that 10% of global GDP will be stored in blockchain by 2027.

Leonardo will also act as a bridge, interlinking SAP’s six categories: Connected Products (insights, supply networks), Connected Assets (predictive maintenance, distributed manufacturing), Connected Fleet (logistics), Connected Infrastructure (networks and grids), Connected Markets (agriculture, smart traffic, cities, etc.), and Connected People (healthcare, sports, connected home), to create a truly holistic view of the enterprise.

97



Figure 64: SAP Leonardo Bridge


What’s special about Leonardo?There are many machine learning and AI offerings in the market (Google, IBM, Baidu, to name a few), so how will SAP compete? Leonardo sits within the system of record and does the analytics and AI (transactions and analytics in the same environment) in-context, in real time and without security risks. Advantages of Leonardo: 1) SAP’s IoT does not move data from system to system, but rather keeps it one place, safeguarded. 2) Analytics occurs within Leonardo; then it can automatically create a runtime solution which is immediately applied back into the system (the ERP system, the system of record). This means a single system is creating and running transactions, analysing itself, and then automatically acting upon the solution (by say ordering a replacement part and blocking off the engineer’s time to do the repair). 3) SAP deeply understands the enterprise data its customers create and collect. The goal is end-to-end automation that links through the enterprise from customer, to production, through to logistics, finance and employee.

Figure 65: SAP technology stack


How can customers use Leonardo? Self-Service Express - customers are able to select implementation of working

code for predefined business use cases. SAP has a library of pre-built use cases, with the launch templates being for the popular predictive maintenance, or vehicle telematics use cases. Customers can be up and running within eight weeks.

98



Open Innovation Edition - aimed at finding new solutions for clients, moving from an initial concept to a workable prototype within nine weeks. Currently,SAP’s Build.me accelerator set can be used to create a closed loop circle across the application mockup and prototyping process.

Enterprise Level - designed for existing customers already utilizing premium SAP support packages. This is an extension of the open edition, and involves developing multiple solutions in parallel with the customer. Solutions will be tailored to each business problem, and customers can either host SAP on site or visit a global design centre for a day of brainstorming, resulting in a working prototype using SAP's build tool.

SenseTime (Private Company)

SenseTime is an AI start-up in China, primarily focusing on the development of computer-vision based AI applications. The company was established in 2014 by Professor Tang Xiaoou and Dr. Xu Li from Chinese University of Hong Kong. In 2017, SenseTime closed Series B funding totaling to US$4bn. By leveraging 5,500 NVIDIA GPUs and 10 GPU clusters, SenseTime has developed a deep learning platform, which enables the company to customize AI solutions for different verticals.

Video surveillance is the core sector of SenseTime, with comprehensive algorithm solutions that cover pedestrian and vehicle detection, characters recognition (e.g. gender, age and race) and people counting/density estimation. Thanks to its strong R&D capability, SenseTime has formed extensive collaborations with leading surveillance camera players in China, such as Uniview.

Smartphone is another vertical where SenseTime has a strong footprint. The company indicates that the fidelity of its existing face recognition solution is comparable to an 8-digit password, which has already been adopted by leading Chinse OEMs in the flagship models to unlock the phone. Qualcomm also plans to incorporate this algorithm into its high-end smartphone AP. Other smartphone applications include portrait beatification/makeup and face grouping for album.

In addition, SenseTime works closely with financial institutions to enable remote authentication, while it also engages in the development of autonomous driving system, which is still at early stage.

SenseTime has an R&D team of over 800 people. In terms of corporate-published academic papers on deep learning, SenseTime ranks No.2, behind only Microsoft and ahead of Google, Facebook and BAT. Through its diversified industrial partners, SenseTime has access to 10bn images across 18 sectors, which in turn sharpens its algorithms to create a virtuous cycle.

Not Covered

https://www.build.me/splashapp/

99



Figure 66: SenseTime's pedestrian and vehicle detection solution

Source: SenseTime

Figure 67: Portrait beautification/makeup solution for smartphone

Source: SenseTime

Siemens AG (SIE GY, N)

Artificial Intelligence and Machine Learning is a topic of increasing relevance across the Capital Goods sector. For most companies, the approach is to deploy bought-in AI base technologies and their internal developments into their own products and solutions (as embedded software). Siemens is the only company in our coverage that has a substantial stand-alone (sold independent of hardware) software business with €4.0bn of revenues in 2017 in addition to €1.2bn of digital services. With its MindSphere IoT operating system, it has a platform aimed at internal and external clients. The overall Siemens software and digital services business grew around 6% organically, based on our estimates. Siemens expects 8% market CAGR from 2018 to 2022. The core of the Siemens software offering is in its Digital Factory business with €2.6bn of revenues (2017, restated for IFRS15) with a full offering of CAD/CAM and PLM software with a focus on industrial markets (automotive, aerospace, process markets, semiconductor/chip design).

Artificial Intelligence is potentially a disruptive technology vs the traditional automation software. Traditional automation systems tell machines and infrastructure what do to through complex stacks of software, honed over years or decades with industry specific vertical knowledge that provides high entry barriers. If machines become self learning, the system changes fundamentally with the value of existing software stacks declining while the field opens up to new competitors. China has largely missed out on the automation opportunity over the past 20 years with no domestic company that has a meaningful size or international presence. Artificial Intelligence represents a potential opening for China to fulfill its ambitions to become a major player in Automation/smart manufacturing by 2025, as stated by its Made in China 2025 initiative.

For example, traditionally, a robot is "told" what to do through a set of preloaded software relevant to the customer's industry which can then be adapted through software provided by robot companies. With AI, a robot – after basic instructions are provided – would learn itself how to optimize the task.

Andreas WilliAC

(44-20) 7134-4569

[email protected]


100



Figure 68: Siemens Software and Digital Services Concept

Source: Siemens.

Siemens works with technology providers IBM Watson, Amazon, SAP, Microsoft Azure and Accenture for AI/Digitalization.

Siemens’s view on Artificial intelligence in Industry: “computing like a machine, deciding like a person”. Software is increasingly making automated decisions. In the machine learning sub-discipline, training data to is used to enable algorithms to learn the right outcome in line with human specifications. Artificial intelligence is based on the perception of information that can originate from sensors, images, language and text. From this information, the software draws its conclusions, learns, adjusts parameters accordingly and generates hypotheses. In the end, it reaches a decision on its own or makes a recommendation that human partners can use to underpin their own actions.

What does artificial intelligence mean for Siemens? Siemens has been active in this field for decades. Today, the company implements this technology in industrial applications:

Complex image recognition as used, above all, for interpreting the results of computed tomography (CT) and magnetic resonance imaging (MRI).

Autonomously learning, self-optimizing industrial systems such as those used in gas turbines and wind farms

Accurate forecasts of copper prices and expected power grid capacity utilization.

In addition, intensive work is being carried out on physical, autonomous systems for use in collaborative, adaptive, flexible manufacturing as part of Industrie 4.0.

Software AG (SOW GY, N)

Within the area of AI, Software AG offers Apama, a pattern matching technology,and the Cumulocity IoT Platform (subset of the Digital Business Platform), as well as surrounding technology and tools like Zementis and Terracotta.

Cumulocity IoT Platform is an open, application centric IoT platform offering analytics and pre-configured solution accelerators (such as predictive maintenance, condition monitoring and track & trace) on the customer’s choice of SaaS, PaaS or on-premises. The platform includes secure device connectivity and management, is device and network agnostic, and can be modified with new technology as its user matures.

Stacy PollardAC

(44-20) 7134-5420

[email protected]


101



Apama is the pattern matching technology, providing CEP (complex event processing). Apama Streaming Analytics offers rich analytics (aggregations, temporal, filtering and location) blending real-time and historic data, rich visualization, business dashboards, and high-performance messaging for any use case and vertical - across mobile, web and IoT. Apama is used for extreme scale and performance on top of in-memory architecture, and support predictive analytics and models.

Zementis is a connector between the data science coding world of Python, R, Knime, etc. and the entire Digital Business Platform. Zementis has a decision engine framework, uses open standard PMML and is focused on the operational deployment of predictive analytics and data mining solutions. Applications include fraud & risk scoring, sensor & device data processing, biometrics, etc.

Terracotta is Software AG’s in-memory data management technology – for speed and scale. It gives in-memory access to terabytes of data, enabling the operationalization of historical big data and its blending with real-time streaming data.

A few examples of customers using Cumulocity:

Deutsche Telekom offers the rebranded Telekom "Cloud of Things" IoT Platform (powered by Cumulocity) as the cornerstone of its M2M/IoT solution business for enterprise in Europe.

Lyreco fully automates the supply chain of its Nespresso premium coffee service using real-time usage data to improve the customers experience and save costs. Lyreco is using the vending machine (including telemetry devices and cellular connectivity), stock and operating management application. Plus its fully integrated in the company's SAP ERP system.

Definitiv (distributor of medical equipment, particularly permanent cooling of certain medicines, for example insulin) has rebranded Software AG's Cumulocity IoT Platform Tenant to track real-time monitoring of temperature, usage and batter level of each Definitiv mobile medicine cooler.

Industry specific platform: ADAMOSSoftware AG has partnered with DMG MORI, Dürr, Zeiss and ASM PT to form a joint venture called ADAMOS (ADAptive Manufacturing Open Solutions), which is dedicated to Industry 4.0 and Industrial Internet of Things (IIoT). These German partners also hope to attract other machine builders to become partners. ADAMOS is an open IIoT platform that offers data autonomy and access to leading software solutions. If provides the basis for employing digital services and new business models, and optimizing production with the aid of big data analyses - all specific to machine building as well as domain- and industry-specific applications.

Sophos (SOPH LN, OW)

Sophos (network and endpoint security for SMEs) has recently accelerated its Artificial Intelligence footprint with the $100m acquisition (Feb 17) of the commercial software products business of US-based Invincea. The Invincea endpoint security portfolio prevents, detects, and remediates zero-day and sophisticated attacks, combining neural-network based machine learning and behavioural monitoring to enhance detection through artificial intelligence and stop evasive malware before damage occurs.

Alexander MeesAC

(44-20) 7742-3681

[email protected]


102



Intercept X V2, the latest version of Sophos’ anti-ransomware platform, incorporates and introduces onto the end-point, Invincea’s deep learning, predictive capabilities; helping to identify and block advanced malware. The introduction of deep learning and behavioural analytics into Intercept X should 1) Enhance quality of detection and 2) Greatly improve false positives, which in turn will make the security of the endpoint easier to manage.

Sophos is also using its deep learning capability in the development of malware protection within its next-generation firewall, and ultimately aims to facilitate a self-learning, interactive, autonomous two-way flow of information from both its network security and endpoint security platforms.

Splunk (SPLK US, N)

Splunk inherently stands to benefit from an explosion of data, given that its business model is leveraged to growth in machine data volumes. Splunk products including User Behavior Analytics, IT Service Intelligence, and its application for AWS, and other partner-developed SplunkBase solutions offer natively integrated ML. Splunk’s track record for acquisitions in the space include SignalSense (cloud advanced data collection and breach detection powered by ML) and select assets of Rocana (analytics) – just within the past quarter.

Tencent (700 HK, OW)

Tencent recently unveiled its Artificial Intelligence strategy and related product lines from four major aspects. In particular, the AI open platform will be the one key focus in 2018.

Establishing AI laboratories to focus on fundamental research

Tencent has started to invest in AI technologies since 2012 and established a number of laboratories to focus on the fundamental research of a few key areas including machine learning, computer vision, voice recognition and natural language processing. So far Tencent has built three AI-focused laboratories including AI Lab, Youtu Lab (primarily on computer vision) and WeChat AI team (on voice recognition, natural language processing, machine learning, etc).

Multi-dimensional application of AI in Tencent’s products

From a use case perspective, Tencent’s AI technology has been integrated into a number of key application use cases including gaming, social and content businesses. Some killer products (such as Honor of Kings, QQ and Kuaibao) have already started to leverage AI tech to enhance user experience. The use cases of AI technology not only applies to Tencent’s core products i.e. social, gaming and content, but also spreads to new initiatives e.g. finance and healthcare sectors.

Game AI: the most well-known application of AI in gaming is Fine Art, the Go-playing AI product of Tencent.

Content AI: AI is widely adopted in content generation, search, recommendation and distribution. For example, Dreamwriter, a machine writing platform of Tencent, is able to report news in sports, finance, securities and technology sectors, with average time to generate an article below 0.5 second. In addition, QQ Music has also adopted AI modules to personalize content push for different users.

Mark MurphyAC

(1-415) 315-6736

[email protected]


Alex YaoAC

(852) 2800-8535

[email protected]


103



Social AI, which is largely integrated into WeChat and QQ, killer social apps of Tencent. For example, Fan Yi Jun, a simultaneous translation app, is based on multiple technologies such as Neural Machine Translation (NMT), Optical Character Recognition (OCR), etc.

Finance AI. Typical examples in this area include Li Cai Tong, a wealth management product of Tencent, adopts chatbot which is based on natural language processing (NLP).

Healthcare AI. Tencent has recently launched Mi Ying, a medical imaging AI product which helps to identify potential disease at an early stage.

Figure 69: Use cases of Tencent AI technologies


Empowering all industries with AI capability: AIaaS (AI as a service)

Tencent Cloud: AI as a service. Tencent has deployed a few key AI technologies on its cloud computing platform, which can be requested by enterprises. As of June 2017, Tencent Cloud has provided 25 AI services across computer visions, voice recognition and NPL, to end users.

Meanwhile, Youtu Lab and WeChat AI have also opened up AI technologies through respective Youtu and WeChat open platforms.

Building an “open-source” ecosystem

Tencent values the importance to establish an open-source platform. As of today, Code, the open-source platform of Tenent, has opened up source codes of two major AI platforms, including ANGEL, a flexible and powerful parameter server for large-scale machine learning, and NCNN of Youtu Lab, a high-performance neural network inference framework optimized for the mobile platform.

Twilio (TWLO US, OW)

Twilio announced the GA (general availability) of its Speech Recognition capabilities in Oct 2017, which enables users to convert speech to text and analyze intent during any voice call. Twilio’s Automated Speech Recognition uses Google’s Cloud Speech API, which supports various languages.

TencentAI

Game AI

Content AI

Social AIFinance AI

Healthcare AI

Mark MurphyAC

(1-415) 315-6736

[email protected]


104



Ultimate Software (ULTI US, OW)

In October 2017, Ultimate Software announced the launch of its AI technology Xander, based on its acquisition of Kanjoya (Sept 2016). The technology can analyze both structured data elements along with employee sentiment (via NLP) in real time and deliver insight to help management make decisions. Xander is embedded within the core UltiPro solution.

Workday (WDAY US, OW)

AI is an integral part of the Workday system and surfaces in terms of intelligent recommendations to predictive analytics across HR as well as Finance. The move toward incorporating AI started in 2015 with the acquisition of Identified and was augmented more recently by the acquisition of Platfora. Workday uses AI for predictive analytics across its solution from predicting employee turnover to recommending learning modules on the HCM side to predicting customer collections and using AI to create financial forecasts on the Financials Management side.

Yitu (Private Company)

Yitu is an AI start-up focusing on computer vision. The company was founded by Dr. Zhu Long from UCLA and Mr. Lin Chenxi from Ali Cloud. In 2017, Yitu closed Round C financing of RMB3.8bn (US$580mn), led by Hillhouse Capital.

Yitu’s solution has been adopted as one of the verification methods for ATM by Agricultural Bank of China and China Merchant Bank. Under the this new solution, customers don't need to insert the debit card into the ATM for cash withdrawal, while the ATM uses a camera to capture the information of their faces, then analyzes the data and links to corresponding accounts. However, customers still have to input the registered phone number for verification to meet the security requirements.

Yitu is a pioneer in the application of computer vision in healthcare. The company has been working with some leading Chinese hospital to develop a diagnosis assistant platform. Current solutions include Chest CT review, ultrasonic diagnosis assistant and pediatrics diagnosis assistant.

Mark MurphyAC

(1-415) 315-6736

[email protected]


Mark MurphyAC

(1-415) 315-6736

[email protected]


Not Covered

105



Appendix 3: AI acquisitions 2012-2017

Table 8: AI Acquisitions 2012-2017

Acquirer Company YOA Description

Ford Argo AI 2017 Automated drivingFord Princeton Lightwave 2017 Lidar startupFord SAIPS 2017 Image and video processing algorithms, as well as deep learning tech for classifying input signals.Facebook Ozlo 2017 AI chatbot startup with an advanced knowledge layerGoogle Kaggle 2017 Platform that hosts data science and machine learning competitions. Large data scientist community.Google Halli Labs 2017 Recommendation/personalizationGoogle AIMatter 2017 Image processingMicrosoft Maluuba 2017 Natural language processingApple Realface 2017 Cyber security and machine learning, specialising in facial recognition techApple Lattice 2017 Turns dark data into structured data. Human level quality, high scale. Learning via distant supervision.Amazon Harvest.ai 2017 Cyber security, uses machine learning to analyse user behaviour, identify threats and prevent attacksSpotify Niland 2017 Machine learning startup specialising in music search and recommendationsSplunk SignalSense 2017 Cloud-based advanced data collection and breach detection solutions, leveraging machine learningMoneyFarm Ernest 2017 Fintech chatbotJones Media Verve.ai 2017 Marketing Machine LearningElement Data Behavior Matrix 2017 Emotional analytics platformAmazon Body Labs 2017 3D Body ModelHubSpot Motion AI 2017 Marketing bot platformCloudera Fast Forward Labs 2017 AI research, specialises in consulting larger enterprises on emerging ML trendsM-files Apperento 2017 Natural languageQualcomm Scyfer 2017 Builds machine learning solutions for companies in different verticalsWorkday Pattern 2017 Pattern recognition, customer relationshipsLyft DataScore 2017 Specialises in customer acquisition and retention via a data driven approachHubSpot Kemvi 2017 Startup applying AI and ML to help sales teams. E.g. best time to reach out to potential customers.Amazon Graphiq 2017 Provides visualisations on complex data. Bought to help improve Alexa.Amazon GameSparks 2017 "Backend-as-a-service" for game developersBaidu KITT.AI 2017 Machine learning startup - Natural language processing technologyBaidu Raven Tech 2017 Develops AI-based voice assistant that supports plugins (enabling it to work with other web services)Sophos Invincea 2017 Malware threat detection, network breach prevention, and pre-breach forensic intelligence. Leverages ML.Sophos Barricade 2017 Behavioural-based analytics engine built around machine learning techniquesGE BITSTEW SYSTEMS 2017 Industrial internet application for machine intelligence - data integration, analysis & predictive automationGE Wise.io 2017 Machine learning powered solution used to identify patterns and trends in dataFacebook Zurich Eye 2016 Computer visionFacebook Masquerade 2016 Image filtering, with a focus on videoAmazon Angel.ai 2016 Acqui-hire. Builds chat botsApple Emotient 2016 Emotion-detection technology to improve understanding of customer sentimentApple Turi 2016 AI tech analyses facial expressions to detect emotionsApple tuplejump 2016 Applies machine learning concepts and analytics to large complex dataeBay Expertmaker 2016 Applies machine learning to extract insights from large amounts of noisy dataeBay SalesPredict 2016 Predictive analytics used to predict customer buying behaviour and customer conversionGoogle Moodstocks 2016 Visual Search StartupGoogle Api.ai 2016 Provides tools to developers to help them build conversational botsIntel Itseez 2016 Computer vision and pattern recognitionIntel Nervana 2016 Deep learning startup developing software and hardwareIntel Movidius 2016 Computer vision chipmaker, for use in drones and virtual reality products, among others.Microsoft Genee 2016 AI powered scheduling service. Uses NLP and optimised decision-making algorithms.NICE Nexidia 2016 Advanced customer analyticsOracle Crosswise 2016 Provider of machine learning based cross-device data, supports marketers.Oracle Palerra 2016 Data security, user behaviour analytics, with automated incident responsesSalesforce PredictionIO 2016 Open-source machine learning serverSalesforce MetaMind 2016 AI-based personalization and customer support solutions for companiesSamsung Viv 2016 AI Virtual AssistantTwitter Magic Pony 2016 Machine learning and visual processing technologyMicrosoft SwiftKey 2016 Creates keyboard apps for Android and iOS devicesAmazon Orbeus 2015 Image recognition, based on neural networksAOL Sociocast 2015 Predictive analyticsApple Perceptio 2015 Developing advanced AI for smartphonesApple Vocal IQ 2015 Speech-processing for improved human-machine interactionFacebook Wit.ai 2015 Speech recognition and voice interfacesGoogle Timeful 2015 Smart scheduling appGoogle Granata Decision Systems 2015 Prescriptive analytics initially focused on marketing resource management

106



Table 9: (Cont.)

Acquirer Company YOA DescriptionIBM Explorys 2015 Predictive healthcare data analyticsIBM AlchemyAPI 2015 Natural language capabilities including keyword extraction and categorizationIntel Saffron 2015 Cognitive computing platformMicrosoft Equivio 2015 Machine learning powered compliance solutionsSalesforce Tempo AI 2015 Smart calendarTwitter TellApart 2015 Predictive advertising for e-commerce and retailTwitter Whetlab 2015 The company claims to have developed a technology to make machine learning better and fasterIBM COGNEA 2015 Cognitive computing and conversational artificial intelligence platformApple Faceshift 2015 Technology used in animation, capturing facial expressions in real-timeAOL Gravity 2014 Personalized advertisementsAOL Convertro 2014 Marketing intelligenceGoogle DeepMind 2014 Develops self-learning algorithmsGoogle Emu 2014 AI-based instant messagingGoogle Jetpac 2014 Aggregates social media pictures and analyzes their locations to provide a travel guideGoogle DM Dark Blue Labs 2014 Deep learning-based technology for understanding natural languageGoogle DM Vision Factory 2014 Object and text-recognition using deep learningIBM Cogenea 2014 AI-based virtual assistantNokia Desti 2014 Travel planning application using AI and NLP to build knowledge on destinationsNokia Medio Systems 2014 Location based predictive analyticsTwitter Madbits 2014 Deep-learning-based visual intelligence platform to identify contents of imagesApple Novauris 2014 Automatic speech recognitionFacebook JIBBIGO 2014 Speech recognition and machine translation startupGoogle DNNResearch 2013 Use of deep learning and neural networks for image searchMicrosoft Netbreeze 2013 Social-monitoring analyticsNICE Causata 2013 Provider of real-time Big Data analyticsYahoo IQ Engines 2013 Image-recognition softwareYahoo LookFlow 2013 API for image recognition and categorizationYahoo SkyPhrase 2013 Natural-language processingIntel Indisys 2013 Natural-language processingFacebook Face.com 2012 Facial recognition, with specialty in mobile

Source: Company reports and J.P. Morgan estimates.

Appendix 4: Other J.P. Morgan reportscovering the topic of AI

More of Moore: ASML reinstating Moore’s Law – a key AI enabler, by Sandeep Deshpande.

Processor Event Takeaways: Multiple Architectures to Address Embedded, AI/Deep Learning Applications, by Harlan Sur.

NVIDIA Corporation: Investor Day Takeaways, by Harlan Sur.

TSMC: Computing – Next growth engine or a bridge too far? by Gokul Hariharan.

Big Data and AI Strategies: Machine Learning and Alternative Data Approach to Investing, by Marko Kolanovic.

European Capital Goods: Automation and IoT, by Andreas Willi

China Industrial Automation, by Karen Li.










109



Disclosures

Analyst Certification: The research analyst(s) denoted by an “AC” on the cover of this report certifies (or, where multiple research analysts are primarily responsible for this report, the research analyst denoted by an “AC” on the cover or within the document individually certifies, with respect to each security or issuer that the research analyst covers in this research) that: (1) all of the views expressed in this report accurately reflect his or her personal views about any and all of the subject securities or issuers; and (2) no part of any of the research analyst's compensation was, is, or will be directly or indirectly related to the specific recommendations or views expressed by the research analyst(s) in this report. For all Korea-based research analysts listed on the front cover, they also certify, as per KOFIA requirements, that their analysis was made in good faith and that the views reflect their own opinion, without undue influence or intervention.

Important Disclosures

Gartner: All statements in this report attributable to Gartner represent J.P. Morgan's interpretation of data opinion or viewpoints published as part of a syndicated subscription service by Gartner, Inc., and have not been reviewed by Gartner. Each Gartner publication speaks as of its original publication date (and not as of the date of this report). The opinions expressed in Gartner publications are not representations of fact, and are subject to change without notice.

Company-Specific Disclosures: Important disclosures, including price charts and credit opinion history tables, are available for compendium reports and all J.P. Morgan–covered companies by visiting https://www.jpmm.com/research/disclosures, calling 1-800-477-0406, or e-mailing [email protected] with your request. J.P. Morgan’s Strategy, Technical, and Quantitative Research teams may screen companies not covered by J.P. Morgan. For important disclosures for these companies, please call 1-800-477-0406 or e-mail [email protected].

Explanation of Equity Research Ratings, Designations and Analyst(s) Coverage Universe: J.P. Morgan uses the following rating system: Overweight [Over the next six to twelve months, we expect this stock will outperform the average total return of the stocks in the analyst’s (or the analyst’s team’s) coverage universe.] Neutral [Over the next six to twelve months, we expect this stock will perform in line with the average total return of the stocks in the analyst’s (or the analyst’s team’s) coverage universe.] Underweight [Over the next six to twelve months, we expect this stock will underperform the average total return of the stocks in the analyst’s (or the analyst’s team’s) coverage universe.] Not Rated (NR): J.P. Morgan has removed the rating and, if applicable, the price target, for this stock because of either a lack of a sufficient fundamental basis or for legal, regulatory or policy reasons. The previous rating and, if applicable, the price target, no longer should be relied upon. An NR designation is not a recommendation or a rating. In our Asia (ex-Australia) and U.K. small- and mid-cap equity research, each stock’s expected total return is compared to the expected total return of a benchmark country market index, not to those analysts’ coverage universe. If it does not appear in the Important Disclosures section of this report, the certifying analyst’s coverage universe can be found on J.P. Morgan’s research website, www.jpmorganmarkets.com.

Coverage Universe: Pollard, Stacy E: AVEVA Plc (AVV.L), Amadeus IT Group SA (AMA.MC), Atos (ATOS.PA), Capgemini (CAPP.PA), Dassault Systèmes (DAST.PA), Hexagon (HEXAb.ST), Indra (IDR.MC), Micro Focus (MCRO.L), SAP (SAPG.DE), SOFTWARE AG (SOWG.DE), Sage Group (SGE.L), Worldpay (WPG.L)

J.P. Morgan Equity Research Ratings Distribution, as of October 02, 2017

Overweight(buy)

Neutral(hold)

Underweight(sell)

J.P. Morgan Global Equity Research Coverage 45% 45% 11%IB clients* 52% 47% 33%

JPMS Equity Research Coverage 45% 49% 6%IB clients* 68% 62% 53%

*Percentage of investment banking clients in each rating category.For purposes only of FINRA/NYSE ratings distribution rules, our Overweight rating falls into a buy rating category; our Neutral rating falls into a hold rating category; and our Underweight rating falls into a sell rating category. Please note that stocks with an NR designation are not included in the table above.

Equity Valuation and Risks: For valuation methodology and risks associated with covered companies or price targets for covered companies, please see the most recent company-specific research report at http://www.jpmorganmarkets.com, contact the primary analyst or your J.P. Morgan representative, or email [email protected]. For material information about the proprietary models used, please see the Summary of Financials in company-specific research reports and the Company Tearsheets, which are available to download on the company pages of our client website, http://www.jpmorganmarkets.com. This report also sets out within it the material underlying assumptions used.

http://www.jpmorganmarkets.com/

mailto:[email protected]

http://www.jpmorganmarkets.com/



https://www.jpmm.com/research/disclosures

110



Equity Analysts' Compensation: The equity research analysts responsible for the preparation of this report receive compensation based upon various factors, including the quality and accuracy of research, client feedback, competitive factors, and overall firm revenues.

Registration of non-US Analysts: Unless otherwise noted, the non-US analysts listed on the front of this report are employees of non-US affiliates of JPMS, are not registered/qualified as research analysts under NASD/NYSE rules, may not be associated persons of JPMS, and may not be subject to FINRA Rule 2241 restrictions on communications with covered companies, public appearances, and trading securities held by a research analyst account.

Company-Specific Disclosures: Important disclosures, including price charts and credit opinion history tables, are available for compendium reports and all J.P. Morgan–covered companies by visiting https://www.jpmm.com/research/disclosures, calling 1-800-477-0406, or e-mailing [email protected] with your request. J.P. Morgan’s Strategy, Technical, and Quantitative Research teams may screen companies not covered by J.P. Morgan. For important disclosures for these companies, please call 1-800-477-0406 or e-mail [email protected].

Analysts' Compensation: The research analysts responsible for the preparation of this report receive compensation based upon various factors, including the quality and accuracy of research, client feedback, competitive factors, and overall firm revenues.

Other Disclosures

J.P. Morgan ("JPM") is the global brand name for J.P. Morgan Securities LLC ("JPMS") and its affiliates worldwide. J.P. Morgan Cazenove is a marketing name for the U.K. investment banking businesses and EMEA cash equities and equity research businesses of JPMorgan Chase & Co. and its subsidiaries.

All research reports made available to clients are simultaneously available on our client website, J.P. Morgan Markets. Not all research content is redistributed, e-mailed or made available to third-party aggregators. For all research reports available on a particular stock, please contact your sales representative.

Options related research: If the information contained herein regards options related research, such information is available only to persons who have received the proper option risk disclosure documents. For a copy of the Option Clearing Corporation's Characteristics and Risks of Standardized Options, please contact your J.P. Morgan Representative or visit the OCC's website at https://www.theocc.com/components/docs/riskstoc.pdf

Legal Entities Disclosures U.S.: JPMS is a member of NYSE, FINRA, SIPC and the NFA. JPMorgan Chase Bank, N.A. is a member of FDIC. U.K.: JPMorgan Chase N.A., London Branch, is authorised by the Prudential Regulation Authority and is subject to regulation by the Financial Conduct Authority and to limited regulation by the Prudential Regulation Authority. Details about the extent of our regulation by the Prudential Regulation Authority are available from J.P. Morgan on request. J.P. Morgan Securities plc (JPMS plc) is a member of the London Stock Exchange and is authorised by the Prudential Regulation Authority and regulated by the Financial Conduct Authority and the Prudential Regulation Authority. Registered in England & Wales No. 2711006. Registered Office 25 Bank Street, London, E14 5JP. South Africa: J.P. Morgan Equities South Africa Proprietary Limited is a member of the Johannesburg Securities Exchange and is regulated by the Financial Services Board. Hong Kong: J.P. Morgan Securities (Asia Pacific) Limited (CE number AAJ321) is regulated by the Hong Kong Monetary Authority and the Securities and Futures Commission in Hong Kong and/or J.P. Morgan Broking (Hong Kong) Limited (CE number AAB027) is regulated by the Securities and Futures Commission in Hong Kong. Korea: This material is issued and distributed in Korea by or through J.P. Morgan Securities (Far East) Limited, Seoul Branch, which is a member of the Korea Exchange(KRX) and is regulated by the Financial Services Commission (FSC) and the Financial Supervisory Service (FSS). Australia: J.P. Morgan Australia Limited (JPMAL) (ABN 52 002 888 011/AFS Licence No: 238188) is regulated by ASIC and J.P. Morgan Securities Australia Limited (JPMSAL) (ABN 61 003 245 234/AFS Licence No: 238066) is regulated by ASIC and is a Market, Clearing and Settlement Participant of ASX Limited and CHI-X. Taiwan: J.P.Morgan Securities (Taiwan) Limited is a participant of the Taiwan Stock Exchange (company-type) and regulated by the Taiwan Securities and Futures Bureau. India: J.P. Morgan India Private Limited (Corporate Identity Number - U67120MH1992FTC068724), having its registered office at J.P. Morgan Tower, Off. C.S.T. Road, Kalina, Santacruz - East, Mumbai – 400098, is registered with Securities and Exchange Board of India (SEBI) as a ‘Research Analyst’ having registration number INH000001873. J.P. Morgan India Private Limited is also registered with SEBI as a member of the National Stock Exchange of India Limited (SEBI Registration Number - INB 230675231/INF 230675231/INE 230675231), the Bombay Stock Exchange Limited (SEBI Registration Number - INB 010675237/INF 010675237) and as a Merchant Banker (SEBI Registration Number - MB/INM000002970). Telephone: 91-22-6157 3000, Facsimile: 91-22-6157 3990 and Website: www.jpmipl.com. For non local research reports, this material is not distributed in India by J.P. Morgan India Private Limited. Thailand: This material is issued and distributed in Thailand by JPMorgan Securities (Thailand) Ltd., which is a member of the Stock Exchange of Thailand and is regulated by the Ministry of Finance and the Securities and Exchange Commission and its registered address is 3rd Floor, 20 North Sathorn Road, Silom, Bangrak, Bangkok 10500. Indonesia: PT J.P. Morgan Securities Indonesia is a member of the Indonesia Stock Exchange and is regulated by the OJK a.k.a. BAPEPAM LK. Philippines: J.P. Morgan Securities Philippines Inc. is a Trading Participant of the Philippine Stock Exchange and a member of the Securities Clearing Corporation of the Philippines and the Securities Investor Protection Fund. It is regulated by the Securities and Exchange Commission. Brazil: Banco J.P. Morgan S.A. is regulated by the Comissao de Valores Mobiliarios (CVM) and by the Central Bank of Brazil. Mexico: J.P. Morgan Casa de Bolsa, S.A. de C.V., J.P. Morgan Grupo Financiero is a member of the Mexican Stock Exchange and authorized to act as a broker dealer by the National Banking and Securities Exchange Commission. Singapore: This material is issued and distributed in Singapore by or through J.P. Morgan Securities Singapore Private Limited (JPMSS) [MCI (P) 202/03/2017 and Co. Reg. No.: 199405335R], which is a member of the Singapore Exchange Securities Trading Limited and/or JPMorgan Chase Bank, N.A., Singapore branch (JPMCB Singapore) [MCI (P) 059/09/2017], both of which are regulated by the Monetary Authority of Singapore. This material is issued and distributed in Singapore only to accredited investors, expert investors and institutional investors, as defined in Section 4A of the Securities and Futures Act, Cap. 289 (SFA). This material is not intended to be issued or distributed to any retail investors or any other investors that do not fall into the classes of “accredited investors,” “expert investors” or “institutional investors,” as defined under Section 4A of the SFA. Recipients of this document are to contact JPMSS or JPMCB Singapore in respect of any matters arising from, or in connection with, the document. Japan: JPMorgan Securities Japan Co., Ltd. and JPMorgan Chase Bank, N.A., Tokyo Branch are regulated by the Financial Services Agency in Japan. Malaysia: This material is issued and distributed in Malaysia by JPMorgan Securities (Malaysia) Sdn

http://www.jpmipl.com/

https://www.theocc.com/components/docs/riskstoc.pdf



https://www.jpmm.com/research/disclosures

111



Bhd (18146-X) which is a Participating Organization of Bursa Malaysia Berhad and a holder of Capital Markets Services License issued by the Securities Commission in Malaysia. Pakistan: J. P. Morgan Pakistan Broking (Pvt.) Ltd is a member of the Karachi Stock Exchange and regulated by the Securities and Exchange Commission of Pakistan. Saudi Arabia: J.P. Morgan Saudi Arabia Ltd. is authorized by the Capital Market Authority of the Kingdom of Saudi Arabia (CMA) to carry out dealing as an agent, arranging, advising and custody, with respect to securities business under licence number 35-07079 and its registered address is at 8th Floor, Al-Faisaliyah Tower, King Fahad Road, P.O. Box 51907, Riyadh 11553, Kingdom of Saudi Arabia. Dubai: JPMorgan Chase Bank, N.A., Dubai Branch is regulated by the Dubai Financial Services Authority (DFSA) and its registered address is Dubai International Financial Centre - Building 3, Level 7, PO Box 506551, Dubai, UAE.

Country and Region Specific Disclosures U.K. and European Economic Area (EEA): Unless specified to the contrary, issued and approved for distribution in the U.K. and the EEA by JPMS plc. Investment research issued by JPMS plc has been prepared in accordance with JPMS plc's policies for managing conflicts of interest arising as a result of publication and distribution of investment research. Many European regulators require a firm to establish, implement and maintain such a policy. Further information about J.P. Morgan's conflict of interest policy and a description of the effective internal organisations and administrative arrangements set up for the prevention and avoidance of conflicts of interest is set out at the following link https://www.jpmorgan.com/jpmpdf/1320742677360.pdf. This report has been issued in the U.K. only to persons of a kind described in Article 19 (5), 38, 47 and 49 of the Financial Services and Markets Act 2000 (Financial Promotion) Order 2005 (all such persons being referred to as "relevant persons"). This document must not be acted on or relied on by persons who are not relevant persons. Any investment or investment activity to which this document relates is only available to relevant persons and will be engaged in only with relevant persons. In other EEA countries, the report has been issued to persons regarded as professional investors (or equivalent) in their home jurisdiction. Australia: This material is issued and distributed by JPMSAL in Australia to "wholesale clients" only. This material does not take into account the specific investment objectives, financial situation or particular needs of the recipient. The recipient of this material must not distribute it to any third party or outside Australia without the prior written consent of JPMSAL. For the purposes of this paragraph the term "wholesale client" has the meaning given in section 761G of the Corporations Act 2001. Germany: This material is distributed in Germany by J.P. Morgan Securities plc, Frankfurt Branch which is regulated by the Bundesanstalt für Finanzdienstleistungsaufsicht. Hong Kong: The 1% ownership disclosure as of the previous month end satisfies the requirements under Paragraph 16.5(a) of the Hong Kong Code of Conduct for Persons Licensed by or Registered with the Securities and Futures Commission. (For research published within the first ten days of the month, the disclosure may be based on the month end data from two months prior.) J.P. Morgan Broking (Hong Kong) Limited is the liquidity provider/market maker for derivative warrants, callable bull bear contracts and stock options listed on the Stock Exchange of Hong Kong Limited. An updated list can be found on HKEx website: http://www.hkex.com.hk. Korea: This report may have been edited or contributed to from time to time by affiliates of J.P. Morgan Securities (Far East) Limited, Seoul Branch. Singapore: As at the date of this report, JPMSS is a designated market maker for certain structured warrants listed on the Singapore Exchange where the underlying securities may be the securities discussed in this report. Arising from its role as designated market maker for such structured warrants, JPMSS may conduct hedging activities in respect of such underlying securities and hold or have an interest in such underlying securities as a result. The updated list of structured warrants for which JPMSS acts as designated market maker may be found on the website of the Singapore Exchange Limited: http://www.sgx.com. In addition, JPMSS and/or its affiliates may also have an interest or holding in any of the securities discussed in this report – please see the Important Disclosures section above. For securities where the holding is 1% or greater, the holding may be found in the Important Disclosures section above. For all other securities mentioned in this report, JPMSS and/or its affiliates may have a holding of less than 1% in such securities and may trade them in ways different from those discussed in this report. Employees of JPMSS and/or its affiliates not involved in the preparation of this report may have investments in the securities (or derivatives of such securities) mentioned in this report and may trade them in ways different from those discussed in this report. Taiwan: This material is issued and distributed in Taiwan by J.P. Morgan Securities (Taiwan) Limited. According to Paragraph 2, Article 7-1 of Operational Regulations Governing Securities Firms Recommending Trades in Securities to Customers (as amended or supplemented) and/or other applicable laws or regulations, please note that the recipient of this material is not permitted to engage in any activities in connection with the material which may give rise to conflicts of interests, unless otherwise disclosed in the “Important Disclosures” in this material. India: For private circulation only, not for sale. Pakistan: For private circulation only, not for sale. New Zealand: This material is issued and distributed by JPMSAL in New Zealand only to persons whose principal business is the investment of money or who, in the course of and for the purposes of their business, habitually invest money. JPMSAL does not issue or distribute this material to members of "the public" as determined in accordance with section 3 of the Securities Act 1978. The recipient of this material must not distribute it to any third party or outside New Zealand without the prior written consent of JPMSAL. Canada: The information contained herein is not, and under no circumstances is to be construed as, a prospectus, an advertisement, a public offering, an offer to sell securities described herein, or solicitation of an offer to buy securities described herein, in Canada or any province or territory thereof. Any offer or sale of the securities described herein in Canada will be made only under an exemption from the requirements to file a prospectus with the relevant Canadian securities regulators and only by a dealer properly registered under applicable securities laws or, alternatively, pursuant to an exemption from the dealer registration requirement in the relevant province or territory of Canada in which such offer or sale is made. The information contained herein is under no circumstances to be construed as investment advice in any province or territory of Canada and is not tailored to the needs of the recipient. To the extent that the information contained herein references securities of an issuer incorporated, formed or created under the laws of Canada or a province or territory of Canada, any trades in such securities must be conducted through a dealer registered in Canada. No securities commission or similar regulatory authority in Canada has reviewed or in any way passed judgment upon these materials, the information contained herein or the merits of the securities described herein, and any representation to the contrary is an offence. Dubai: This report has been issued to persons regarded as professional clients as defined under the DFSA rules. Brazil: Ombudsman J.P. Morgan: 0800-7700847 / [email protected].

General: Additional information is available upon request. Information has been obtained from sources believed to be reliable but JPMorgan Chase & Co. or its affiliates and/or subsidiaries (collectively J.P. Morgan) do not warrant its completeness or accuracy except with respect to any disclosures relative to JPMS and/or its affiliates and the analyst's involvement with the issuer that is the subject of the research. All pricing is indicative as of the close of market for the securities discussed, unless otherwise stated. Opinions and estimates constitute our judgment as of the date of this material and are subject to change without notice. Past performance is not indicative of future results. This material is not intended as an offer or solicitation for the purchase or sale of any financial instrument. The opinions and recommendations herein do not take into account individual client circumstances, objectives, or needs and are not intended as recommendations of particular securities, financial instruments or strategies to particular clients. The recipient of this report must make its own independent decisions regarding any securities or financial instruments mentioned herein. JPMS distributes in the U.S. research published by non-U.S. affiliates and accepts responsibility for its contents. Periodic updates may be provided on companies/industries based on company specific developments or announcements, market conditions or any other publicly available information. Clients should contact analysts and execute transactions through a J.P. Morgan subsidiary or affiliate in their home jurisdiction unless governing law permits otherwise.

http://www.sgx.com/

https://www.jpmorgan.com/jpmpdf/1320742677360.pdf

112



"Other Disclosures" last revised November 11, 2017.

Copyright 2017 JPMorgan Chase & Co. All rights reserved. This report or any portion hereof may not be reprinted, sold or redistributed without the written consent of J.P. Morgan. #$J&098$#*P

Documents

An Investors' Guide to Artificial Intelligence - Flamingo AI · PDF fileGlobal Equity Research 27 November 2017 An Investors' Guide to Artificial Intelligence AI adoption at an inflection