Published October 10, 2018

Abstract

Artificial Intelligence (AI) is one of the more heavily hyped and widely misunderstood technology topics of the past 30 years—it is right up there with blockchain. This report looks to demystify AI by defining, categorizing and describing the three distinct waves of AI including the capabilities, limitations, use cases and future state expectations of each category.

This is the first of a series of reports designed to provide TechVision Research clients with practical information to assist in developing plans and allocating investments in this complex and disruptive area. This report offers a high-level AI foundation and an enterprise reality check for Line-of-Business leaders and IT executives.

We believe that the starting point for any major enterprise program is for key stakeholders to get on the same page with consistent definitions and context. This can be a challenge in AI since there are so many widely divergent categories and so much marketing hype. We seek to address this by breaking AI into three distinct waves or categories. The first wave (symbolic reasoning/directed AI) works best in supporting well-defined, static applications, but has major limitations in broader use cases. The second wave (statistical/machine learning) provides a self-learning capability by applying real world data to statistical models to improve results over time. The third wave (contextual adaptation) blends the abilities of the previous two waves to understand and explain how a decision was made. Each wave is adept in solving distinct types of problems and are at a different maturity levels.

This foundational document will help enterprises understand, categorize, prioritize and ultimately, develop strategies leveraging the various types of AI. We provide high-level recommendations and an action plan for enterprises developing strategies, architectures, governance models and plans in AI and related areas.

Authors:

Gary Rowe                                                      Jeff Nichols

CEO / Principal Consulting Analyst              Principal Consulting Analyst

[email protected]                       [email protected]

Contributing Analyst:  Barbara Starr

Executive Summary

This report is a level set for enterprise technologists, architects, innovators, visionaries and business leaders developing plans and considering investments in AI technology. This document is intended as a level set; to get stakeholders on the same page in understanding the various types of AI and associated and use cases. While virtually every Global 1000 enterprise has some type of AI program underway, the range of AI capabilities and investments vary widely. To get a handle on these various categories, we start by describing three basic classes or waves of AI. We use category definitions defined by DARPA, the Defense Advanced Research Project Agency

The first wave is called symbolic reasoning or directed AI and this category focuses on leveraging expert knowledge to automate well-defined tasks. Areas like automating tax preparation and chess playing programs are examples of applications that fit within this wave. These AI services are effective when an enterprise is looking to automate a process with well-defined rules, but falls short in handling uncertainly.

The second wave introduces machine learning using statistical models and big data to “train” systems to achieve desired results – typically a correlation or classification of data. This provides flexibility for developers in that they don’t need to understand and program all contingencies in advance; the systems adapt with changing conditions and new information. This is the essence behind many of high-visibility applications that handle voice recognition, language processing, personalization and predictive analytics.

The third wave is called contextual adaption or cognitive computing in which AI systems are self-sufficient, in that they are both contextually aware and can discover and assess the logic behind their decision making process. This feedback loop and decision process refinement is the closest to mimicking real human behaviour, but is the most nascent area within AI.

So where should you start? There are many detailed recommendations throughout this report, but most enterprises will be well served by taking this path:

• Invest the time and money to make your technology and business leadership conversant with AI techniques and solutions. The age of AI-powered automation is starting now, and companies that ignore this technological shift will be left behind competitively.
• Understand that corporate data is the fuel for Wave 2 and 3 AI solutions. Make sure that you have an ongoing program to inventory, categorize and manage your data. Don’t let line of business or industry silos block this effort.
• Begin a Wave 2 or Wave 3 AI pilot. Select an area of the business that is strategic and data-intensive. Challenge an industry technology partner to help you derive value from the pilot AI solution from the beginning and learn from this experience.

Introduction

Artificial Intelligence (AI) and Machine Learning (ML) are not new. In fact, some of the current products and services have roots that extend back to the 1950s and earlier. Fully understanding AI is like understanding the Internet; it is so diverse we need to break it up smaller classes to make sense of it and to assess how different types of AI support appropriate use cases and business goals. Categorizing the types of AI and demystifying it is the foundation for developing an enterprise strategy, architecture and roadmap for your AI program. While AI systems and services can be considered as a separate category for budgeting, AI is increasingly integrated with many applications, security infrastructure and services throughout an enterprise. For example, security products and big data analytics engines have AI capabilities (or claim to have the capabilities) embedded within these services.

AI has historically been heavily supported by DARPA, the Defense Advanced Research Projects Agency. DARPA has funded, defined, architected and built some of the core elements that comprise AI today.  Major commercial programs such as IBM Watson and Apple’s Siri have leveraged these DARPA-funded efforts. The investments made by DARPA over the past 40 plus years include core research, development, testing, education, evangelizing and building out AI in all the major categories we’ll describe in this report. To simplify and organize our understanding of AI, we’ll use DARPA’s wave model to characterize the various categories of AI.

This model breaks down AI into three basic waves; the first wave uses a set of rules to represent knowledge in a well-defined area. It follows a set of rules within a limited context. The second wave is focused on statistical learning or machine learning and leverages big datasets to train systems and improve problem solving. The third wave, called contextual adaption, adds explanations and an understanding of how decisions are made and uses that data to improve future decisions.

A starting point in understanding these three is to view them at the highest possible level by characterizing the waves as “Describe”, “Categorize” and “Explain”. The Describe phase (wave 1) is characterized by assembling expert knowledge in a well-defined domain; the categorize phase (wave 2) is characterized by statistical learning based on accumulated data; and the explain phase (wave 3) focuses understanding and optimizing how decisions are made.

The organizations investing in basic AI research and product development (e.g. IBM, Amazon, Google) will describe the three AI waves slightly differently than DARPA. You may hear AI waves or categories described as (1) basic automation, (2) machine learning, and (3) cognitive computing – or via other similar labels. In terms of the development of an enterprise AI roadmap and architecture, the choice of the exact nomenclature is not as important as understanding there are three basic categories of consisting of different AI capabilities. All are currently relevant to Global 1000 businesses, but have different use cases and are at different states of maturity.

The following table provides a succinct comparison of these waves in an enterprise context.

	Describe (Wave 1)	Categorize (Wave 2)	Explain (Wave 3)
Also Known As	Simple automation; symbolic reasoning	Machine learning	Cognitive computing; contextual adaptation
Maturity	High	Med-High	Low
Examples	Robotics Process Automation (RPA) Games Clustering algorithms Decision trees Rules-based systems	Neural networks Natural language processing Image recognition Event correlation	IBM Watson Google Deepmind HPE Haven Microsoft Cognitive Compute Services Apple Siri
Programming Model	Declarative code and rules	Application of statistical models to large datasets	Proprietary combinations of automation, machine learning and overt training with rich datasets
Failure Modes	Unexpected data or conditions lead to software crashes	Statistical models do not adequately represent reality and the software reaches false conclusions	Poor or sparse training data leads the software to false or low probability conclusions
Data and performance	Performance degrades with more data unless processing architecture is optimized	Performance generally improves with more data	Performance generally improves with more data, though proprietary subsystems or techniques may not scale well
Investment to implement in typical company	Low	Moderate	High
Hallmark features	Widely used and understood	Can detect hidden correlations of events Stronger with massive datasets	Can collaborate with humans via natural language Can explain how or why a conclusion was reached

Table 1 – Waves of AI

A key difference between Wave 1 AI and Wave 2 is that Wave 2 systems improve with larger and larger  data sets, where Wave 1 systems degrade with massive data. Another way to think about  the differences between the waves 1and 2 is consider them in terms of their system life cycles. Wave 1 is a classic software development model via programming for a defined data set. Wave 2 introduces the notion of training systems rather than rules or code-based programming.

Another important differentiator is that Waves 2 and 3 are scale-driven; they thrive with bigger and bigger data. The more data, the more scale, the more effective Wave 2 and 3 techniques are. Given the exponential growth in available data (internal and external), the market is moving rapidly toward Wave 2 and 3 techniques for corporate applications.

First Wave of AI:

The first wave of Artificial Intelligence is based on narrowly defined systems that are focused on solving specific problems. This wave leverages expert knowledge in a target domain characterized by a set of specific rules and/or code implementing a flowchart. These systems don’t look to replicate human intelligence, but simply augment it by distilling large bodies of knowledge within elaborate decision trees built by teams of experts in these narrowly defined areas. This approach is also called symbolic reasoning. The following figure characterizes key elements of the first AI wave:

Figure 1 – First Wave characteristics

The first wave of  AI called symbolic reasoning perceives (collects information) based on a use-case appropriate taxonomy and structure within a manually defined set of rules. Experts define  the taxonomy, structure and rules, generally in a very narrowly defined space. Note that there isn’t self-learning or abstracting capability in this wave; the system perceives or collects/interprets data and then applies pre-defined rules to “reason” in support of solving a problem or interpreting a pattern.  Wave 1 is brittle and deterministic; it works best if it has a clear path and no variances, but typically can’t handle surprises.

Examples of First Wave Applications

Examples of Wave 1 AI applications include tax programs, gaming systems, expert systems, and  some medical diagnostics; basically anything in which experts can predefine a set of rules in a well-defined space. A game for example, has rules, specific actions that can be taken and responses to every action. Tax programs also have very specific rules and logic based on pre-defined factors such as income, taxes paid, dependents, deductions…within the context of very specific tax codes. In the case of TurboTax, it starts with experts such as tax lawyers and tax accountants that take the complexity of the tax laws and convert them to rules that the computer program can work through with different data sets. Note that early Wave 1 applications (like Tax programs) are improving their initial offerings by investing in Wave 2 technologies. The following graphic shows a few typical first wave applications.

Figure 2 – Examples of Wave 1 AI applications

All of these Wave 1 applications are in well-defined areas, they codify specific expert knowledge and apply this to address real world scenarios like optimally placing shipping containers, winning a chess match or filing code-compliant tax returns.

First Wave AI: Considerations for the Enterprise

The first thing that is important to realize about Wave 1 AI is that it isn’t made obsolete by subsequent Waves. Rule-based and decision tree-based systems will be with us for the foreseeable future, as either standalone applications or as subsystems of Wave 2 and 3 systems. The first wave of AI is really good at modeling real-world processes, but not good at learning or abstracting; which involves taking insights at one level and apply it to another level.

A cybersecurity example of how Wave 1 can be very powerful is evidenced by DARPA’s Cyber Security Grand challenge. This event, conducted over a 2 year period demonstrated that automatic defensive systems are capable of reasoning and assessing flaws, formulating patches and deploying them on a network in real time. This competition had teams analyze 131 different programs and find vulnerabilities as well as fix them automatically while maintaining performance and functionality. Collectively, all teams managed to identify vulnerabilities in 99 out of the 131 provided programs.

There’s another reason Wave 1 systems will remain relevant: Wave 1 systems are simpler and more reliable for well-defined problem spaces. Wave 1 systems do not require the massive datasets and computational capability (generally) of Waves 2 and 3, and provide reliable, deterministic results once any logical bugs are cleaned up after testing. So for simpler, bounded problem spaces, Wave 1 systems are the appropriate solution.

But Wave 1 has many use cases that are not a fit. Self-driving cars, for example, initially leveraged Wave 1 technologies, but were limited when scenarios that were not known in advance were introduced. Self-driving cars only became effective and could complete sophisticated demonstration courses when Wave 2 AI was introduced as the systems could learn from vast quantities of driving data and have fewer and fewer unrecognizable experiences.

It’s safe to say that every Global 1000 company has Wave 1 AI systems in use today. Decision trees (essentially, flowcharts with probabilities specified at branching points) are a common technique in traditional, legacy applications. And rule-based systems are also common ranging from firewall rules to expert system modules embedded in applications. But as we’ve described, Wave 1 isn’t sufficient for many of the more advanced use cases.

Second Wave of AI: Statistical Learning (Machine Learning)

Wave 2 provides better classification and prediction capabilities than Wave 1 based on the ability to learn and improve results over time. This wave, also known as Machine Learning (ML) addresses problems in specific domains by leveraging statistical models that are trained and improve themselves by leveraging big data.

Machine Learning is an application of AI that does what comes naturally to humans (well, except for teenagers); it offers the ability to automatically learn and improve from experience without being explicitly programmed. This is how Arthur Samuel defined Machine Learning in 1959 and is consistent with current definitions. The “learning” part of ML is generally looking to either minimize errors or maximize the probability of a prediction being accurate and uses data to support these goals.

Machine Learning is a sub-field within Artificial Intelligence and a major part of this second wave. In recent times the term deep learning is interchangeably used with Machine Learning, but it should be noted that deep learning is a sub-field within ML focusing on learning thru neural networks. The following graph characterizes the core capabilities and deficiencies of this Second Wave of AI.

Figure 3 – Second Wave characteristics

Whereas Wave 1 focuses on symbolic reasoning within well-defined areas, Wave 2 is all about learning and predicting. Wave 2 take large quantities of data and uses that data to teach the service how to better interpret and act upon required tasks. Increased emphasis on learning and perceiving (leveraging more data) differentiate this second wave from the first.

Voice recognition and facial recognition are good examples of how Wave 2 AI can get better results using collected data and comparing actual results to desired outcomes. A Wave 2 system doesn’t learn in the way we think of learning – they create statistical models to represent the problem domain they are trying to solve and then they train the statistical models on specific targeted data. This wave is very good at perceiving the natural world; doing things such as separating one face or sound from another. Wave 2 systems can also adapt to different situations over time by building and expanding its knowledge base.

Wave 2 is one of the higher visibility AI areas, has generated a lot of hype and it has played a major role in the dramatic growth and investment in the AI space over the past several years. There are several reasons why Wave 2 is gaining traction as described in the following diagram:

Figure 4 – Reasons for Wave 2 traction

As we said earlier, Wave 2 systems flourish with big data, cheap storage and cloud computing; all of which are major areas of growth and investment. The combination of more efficient compute power, better machine learning algorithms and a massive portfolio of patented innovation has been critical in driving growth and innovation in this space. A major strength in this Wave is its ability to classify data and predict the consequences of data.  Its major weakness is the lack of ability to understand context and minimal ability to reason. These are the more “human-like” functions we’ll begin to explore in Wave 3.

So how do machines learn? A basic model for Wave 2 is driven by what is called the manifold hypothesis; the grouping or clustering of data based on a natural commonality (people with red hair or 4 legged animals) to form a cluster or substructure. The different types of data can be clumped together with each manifold (or cluster) representing a different entity. AI will separate the different clumps and modify these groupings as more data is aggregated. We can see the progression between Wave 1 and Wave 2 as follows: Wave 1 can’t even describe the clumps (remember they need to know the exact state in advance), but the second Wave takes a new data space, look for commonality and leverages these on-going observations and the associated statistical models to better understand the spaces. This is what we mean by statistical/machine learning.

There are two basic categories of machine learning; supervised and unsupervised. Supervised ML trains the statistical models via known input and output data to predict outcomes based on evidence and the presence of some uncertainty as new data is presented. Unsupervised learning looks for hidden patterns or innate structures in data and uses that information to draw inferences from the collected data.

These are the foundational elements of Wave 2. We will also dig much deeper into these areas as well as multi-layered machine learning models called neural networks in future reports. TechVision will also focus on the development of enterprise strategies, vendor offerings and the development of enterprise programs in the ML space in future reports.

A major challenge in Wave 2 systems in that the results are not always accurate and can even be intentionally “spoofed”. We can’t fully depend on the results as they are a function of strength of the statistical models, the quality of the data and the integrity of the user. For example, Wave 2 AI determined in reviewing the following picture is of a boy with a baseball bat.

Figure 5 – Potential Wave 2 issues

No human would, of course, view this figure as a boy with a baseball bat, but a Wave 2 AI system made that determination. Based on this type of result we characterize Wave 2 AI as statistically impressive, but individually unreliable. Wave 2 will mostly get it right, but there are anomalies. There are many applications where “good enough” is fine, but cases where greater certainty is necessary.. Wave 2 may be sufficient to route a customer service call, but not to provide access to a bank account without further data. A caution with Wave 2 is that we have to be careful about the data we are using the train the ML systems. The old programming adage “garbage in garbage out” applies more than ever, as the machine-generated “code” based on learning is a direct result of data encountered. We need to be cautious in that skewed training data can create maladaptation (incorrect results on an on-going basis) and have dire business consequences.

The challenges we see in Wave 1 and Wave 2 indicate that we ultimately need to move beyond simple calculations based on big data sets and statistical models. We need to factor in some level of human-like contextual adaption and explanatory models. This is what we are moving to in Wave 3.

Figure 6 – Examples of Wave 2 applicaitions

Examples of Second Wave Applications

We experience the use of ML in personal applications every day. As shown in Figure 6 above, many of the applications within our mobile devices leverage these powerful data capabilities to enhance the user experience. For example,

• Google maps uses ML to optimize routes based on speed of users in the area and road conditions.
• Banks are using AI and ML to decipher and convert handwriting on checks into text via OCR to eliminate the need to drive to a branch to deposit a check.
• Uber uses AI and ML to determine the price of your ride, minimize the wait time once you hail a car, and optimally match you with other passengers to minimize detours.

In the enterprise space, there are many good, practical applications of ML at work as well. These include:

• Real-time classification of customers via multiple parameters, replacing traditional offline market segmentation approaches.
• Automated defect recognition on assembly lines and in the field. Machine vision approaches can compare image sensor data to a trained model and indicate whether or not the component in view is within defined QA tolerance levels.
• Facial recognition and facial expression recognition via deep learning networks. Currently used in airports, law enforcement, and social media networks.
• Authentication of people to devices by facial recognition, used by Apple in iOS 10 and the Apple X iPhone.
• Computational biology supporting drug discovery and DNA sequencing.
• Predictive maintenance in manufacturing (automotive, aerospace).
• Predictive models in health care

As Machine Learning tends to use the same inputs as humans (images, sounds, language), its application in industrial and consumer processes are vast. Fast forward 5-10 years, it will be hard to find a modern human endeavor that is not at least partially automated via ML.

Wave 3:

There has been tremendous interest in Artificial Intelligence (AI) and Machine Learning (ML) in recent times. As we discussed earlier, the AI resurgence can be attributed to many factors like easy availability of electronic data, cheaper and faster computing, the open source nature of software and hardware in this space, investment by large players in anticipation of a break thru and advances in algorithms that at least try to replicate human nature.

With all these advancements and escalating investments, we are at the tip of the iceberg when it comes to Wave 3 cognitive computing. To many of the practitioners in the field it is clear that we are more artificial and less intelligent at present and want to be less artificial and more intelligent. The main reason lies in the quality of processing versus the quantity of processing.

Machines and humans are different. At present computers are extremely powerful in quantity of processing where as they are weak in quality of processing. For example, humans are extremely good at separating noise and signal whereas computers find it very difficult. Today they (AI systems) outperform humans in narrow tasks to a great extent by the sheer power of brute force computing as opposed to the clever nature of solving problems.

The AI Wave that is in an embryonic state right now is Wave 3; contextual adaption or cognitive computing. This is where the models can “explain” their decisions and use that information to drive improved decisions in the future. Wave 3 takes on some of the best attributes from Wave 1 and Wave 2 and adds a contextual layer to make it more useful, relevant and adaptive for a wide variety of use cases. Key elements of Wave 3 AI include base-line explanatory models for classes of real world events, driving improved ease of use and information sharing via natural communication amongst both people and machines as well as an enhanced level of learning, reasoning and adaptation to new tasks and events.

Figure 6 – Third Wave characteristics

From a user’s point of view, the most important quality of Wave 3 systems is their accessibility – non-technical people can and will interact deeply with Wave 3 systems because they are built to be collaborative. Wave 3 systems can typically interact via natural language rather than technical code or models. Also, Wave 3 systems typically use and interpret a broad array of data types, including text, images, sounds, and device telemetry.

Wave 3: Future State and Planning Considerations

From a technology professional’s point of view, Wave 3 systems are the pinnacle of current AI technology. Combining the cutting edges of neural networks, computer and network architecture, machine learning algorithms, sensor technology and data management, Wave 3 systems have tremendous potential for revolutionizing corporate computing as well as everyday life.

There is however a dark side to Wave 3 systems. Complex, adaptive systems are already showing us their technical and societal problems. A great example is Tesla’s autopilot software, a complex machine learning system with multi-spectral real time inputs and direct control of a moving vehicle. It works well in almost all settings and situations, but for autonomous driving “almost” is problematic.

The complexity of adaptive systems is also a concern. Once instantiated, these systems do what they are designed to do – adapt their algorithms to the data they are presented with and according to feedback loops. This adaptation makes troubleshooting by their human creators difficult or impossible – we already see systems in production in which the original creators cannot say precisely why or how the system is getting results (or making mistakes).

Aside from the obvious technical challenge this poses, ethical and legal problems arise from adaptive systems taking independent action in the real world. Even with the problems currently seen in autonomous driving systems, is it better and safer to have everyone on the road guided by vastly capable networked systems. And who is to blame if/when the system breaks down and people or property are harmed?

On another front, machine learning-based facial recognition is becoming a commodity, and with the proliferation of networked cameras across the world we are all subject to having our presence, location, identity and activities logged constantly, without our consent. What should the owners of the data from this surveillance do with it, ethically?

To deal with some of these issues, Google, for example has published specific principles for their AI research (see https://ai.google/principles/ ). This is an excellent first step in establishing foundations for safety, privacy and integrity in AI-based systems.

Examples of Third Wave Applications

The most well-known Wave 3 AI application is probably IBM’s Watson. Watson is not necessarily the pinnacle of current AI, but it is the most heavily marketed. Watson has many faces (use cases), but a great example is Watson trained for health care diagnoses. In particular oncology, IBM researchers and leading health care organizations (e.g. Mayo Clinic and Memorial Sloan-Kettering) are using Watson to diagnose cancer from symptoms and cell images. Progress has been slow but steady, and Watson is functioning more as a physician’s assistant rather than an independent diagnostician at this time.

Amazon’s Alexa is another recognizable application of Wave 3 AI. Alexa is a complex system with edge devices (the in-home Alexa speaker), a back-end natural language processing system, a link to hundreds of information sources, and a link to Amazon’s own back-end transactional system, the Amazon online store. Alexa can be used by non-technical people to converse casually, to retrieve information, and to purchase items from Amazon.

State of the Industry

The state of commercial AI research is strong and evolving rapidly. The world’s largest companies are all making investments in basic or applied AI research. This is in addition to the government-sponsored research efforts of DARPA and others, plus extensive university-led primary research.

Some of the best-known AI-based systems come from IBM, Google, Amazon and Microsoft – giants of the technology industry. The landscape is by no means restricted to these four, but we will focus on them to provide an overview of what’s happening in commercial AI and where it may be going.

IBM Watson

IBM has invested heavily in becoming a pre-eminent AI solution provider, using the tradename WatsonÔ to brand a collection of software systems using Wave 2 and 3 technology. IBM’s Watson is, of course, famous for competing on the television game show Jeopardy in 2011 and for its predecessor Deep Blue winning a chess match with Garry Kasparov in 1997. The platform has advanced substantially since then, and custom-trained versions of Watson are now providing insights in industries include medical/healthcare, tax, education and weather forecasting. Watson itself is a proprietary IBM system, available to corporate users as a cloud service. Other software components also carry the Watson label and should not be confused with the core AI platform. One of Watson’s claims to fame is its advanced ability to collaborate with people using natural language input and output. Overall, Watson is a key part of IBM’s enterprise solutions strategy and is expected to yield more than $10B in annual revenues by 2024.

Google

Google has a large commitment to and investment in AI as a strategic technology. They created a consolidated AI division in 2017, tying together all the company’s AI-related products and services. Google AI has several distinct foundations:

• A likely-unmatched reservoir of data to power and train machine learning algorithms, including indexes of most Internet websites and the extensive images collected by Google Maps Street View and Satellite View
• Their 2014 $500M acquisition of the cognitive engine Deepmind
• A proprietary chipset for machine learning called a “tensor processor”
• An extensive research team dedicated to advancing the AI state of art
• An AI-focused venture fund, Gradient Ventures

Google’s core offerings for business are its Tensorflow open sourced machine learning library, Cloud AutoML services (currently focused on vision and image recognition), and several open-sourced massive datasets. Much like Amazon, it’s becoming increasingly hard to separate Google’s AI initiatives from its core business – AI is strategically and completely embedded in their overall approach to services.

Amazon

Amazon was a bit later to the AI party than pioneers like IBM, but their progress since 2010 has more than made up for its delayed start. Amazon’s AI thrust has at least three components; (1) A consumer-oriented AI-driven Q&A system fronted by Amazon’s Alexa and Echo devices, (2) extensive machine learning libraries and services offered by AWS (Amazon Web Services), and (3) the AI-driven Amazon recommendation system, powering its consumer website. The analyst team at TechVision originally believed that only the AWS AI services would be of interest to our corporate clients, but Alexa as a voice I/O platform is gaining substantial traction in more innovative enterprise solutions. For example, multiple companies are using a custom-trained Alexa subsystem to interact with customers via telephone connections..

AWS Machine Learning is used by clients as diverse as Zillow (real estate), Pinterest, NASA (aerospace), the National Football League and Netflix. A relatively recent addition to Amazon’s offerings is AWS Sagemaker, a machine learning toolkit for the masses.

Amazon has created a virtuous cycle for itself with its Wave 2 AI offerings. Machine learning thrives on data and feedback, and both AWS and Alexa platforms generate massive amounts of data from customer interactions. The ever-growing big data repository helps designers and refines the algorithms to improve performance. Both Amazon’s customers and Amazon itself benefit from these data-driven improvements, which may (or may not) ultimately outweigh any privacy concerns held by customers.

Microsoft

Unlike Google and IBM, Microsoft has no overarching AI brand or product. Microsoft’s AI offerings are focused around machine learning solutions hosted in their Azure cloud. Microsoft’s approach is very pragmatic and friendly to the corporate user. Their Azure-based offering includes infrastructure, services and tools for building AI solutions. In the context of this report, we would classify Microsoft’s solutions as Wave 2 AI.

A set of APIs and libraries for “Vision, Speech, Language, Knowledge and Search” are the building blocks for what Microsoft calls “Cognitive Compute Services”. For companies that are very Microsoft-centric in their IT architecture, these services provide a great starting point for AI prototyping and generally useful data input and manipulation. Microsoft also offers ML-based templates for image classification, predictive maintenance, information discovery and chatbot applications as part of their AI services.

In summary, it’s clear that the landscape for AI products and services is rich and rapidly growing. TechVision’s clients are strongly recommended to stay current with that landscape and the new capabilities that will be served up over the coming years. Stay tuned for in-depth coverage including Wave 2 and Wave 3 as well as Robotic Process Automation (RPA) which is primarily Wave 1 AI.

Recommendations

As of late 2018, it is clear that all Global 1000 companies will find themselves using the AI technologies outlined in this report, either directly or indirectly. The capabilities and advantages offered by Wave 2 and 3 AI techniques are too compelling to ignore. Technology and business leadership should become acquainted with at least the basic concepts and tradeoffs of AI-based solutions. For IT leadership, a more in-depth understanding of tradeoffs and use cases will be required.

TechVision Research summary recommendations and next steps follow:

• Understand that the products, services and solutions offered by AI suppliers are different and somewhat confusing to traditional IT professionals. Clients should take the time to educate enterprise, solution and software architects on the capabilities, categories, the timing and tradeoffs offered by AI solutions.
• Modify your system development or system acquisition life cycle to evaluate AI-based solutions explicitly. Too often, only traditional approaches are considered.
• Consider modifying your move-to-production life cycle to test Wave 2 and 3 systems, which can exhibit completely unexpected behavior when interacting with other systems. Companies that move aggressively into AI solutions will need more robust and dedicated test environments to be sure that adaptive systems do not cause harm.
• Advanced or predictive analytics practitioners are often already conversant with Wave 2 techniques. Consider extending the charter of those practitioners in your company to include the AI charter.
• As data is the fuel for AI-based systems, get a handle on your enterprise data inventory and your business data architecture. If not already done so, commission a team to design your business data architecture, inventory and keep track of data stores, understand the growth of this data, and provide appropriate access and critical security controls.
• Understand that there is an opportunity to engage and learn (mutually) via Wave 2 and Wave 3 AI vendors that include many smaller companies and startups with strong ML and algorithmic expertise, often specializing in a vertical markets. What these vendors often don’t have are vast stores of real-world data and what they do have is some leading edge technology that may be lacking in many enterprises. As a result, Wave2 startups and Global 1000 organizations have some mutual synergy worth exploring.
• Launch a pilot Wave 2/3 project in a strategic area – Customer-facing systems, industrial control systems, or cybersecurity systems, for example. Become acquainted with the industry offerings and how to use them to your advantage. Look to solve a specific problem and iterate based on what you discover.
• As you consider which major cloud service(s) to integrate with your legacy systems (e.g. Microsoft Azure, Amazon AWS, SAP), make sure you take the provider’s AI and machine learning services into account. If AI becomes a strategic advantage for you, partnering with the right cloud service provider is crucial. Few companies can afford to have their data scientists, analysts and application solution developers supporting multiple complex architectures.

Conclusions

Artificial intelligence, once the province of science fiction, is now a practical reality that should be in use in virtually all Global 1000 corporations. All waves of AI are candidates for various enterprise use cases and key to developing sustainable competitive advantages in most industries.

AI-based systems include capabilities that traditional code-based systems cannot replicate. IT leaders must gain expertise in understanding how AI-based systems will augment (or in some cases replace) traditional enterprise solutions, and must form strong partnerships with business leadership to make optimal use of corporate data. It is important to educate internal teams, gain practical experience and iterate in building out your AI program. The above recommendations give you a starting point in this challenging but critical area.

In subsequent reports TechVision Research will detail the Wave 2 and Wave 3 AI landscapes and describe how leading companies are already taking advantage of their distinct capabilities.

About TechVision

World-class research requires world-class consulting analysts and our team is just that. Gaining value from research also means having access to research. All TechVision Research licenses are enterprise licenses; this means everyone that needs access to content can have it. We know major technology initiatives involve many different skill sets across an organization and limiting content to a few can compromise the effectiveness of the team and the success of the initiative. Our research leverages our team’s in-depth knowledge as well as their real-world consulting experience. We combine great analyst skills with real world client experiences to provide a deep and balanced perspective.

TechVision Consulting builds off our research with specific projects to help organizations better understand, architect, select, build, and deploy infrastructure technologies. Our well-rounded experience and strong analytical skills help us separate the hype from the reality. This provides organizations with a deeper understanding of the full scope of vendor capabilities, product life cycles, and a basis for making more informed decisions. We also support vendors when they carry out a product and strategy review and assessment, a requirement analysis, a target market assessment, a technology trend analysis, a go-to-market plan assessment, or a gap analysis.

TechVision Updates will provide regular updates on the latest developments with respect to the issues addressed in this report.

About the Authors

Gary Rowe is a seasoned technology analyst, consultant, advisor, executive and entrepreneur. Mr. Rowe helped architect, build and sell two companies and has been on the forefront the standardization and business application of core infrastructure technologies over the past 35 years. Core areas of focus include identity and access management, blockchain, Internet of Things, cloud computing, security/risk management, privacy, innovation, AI, new IT/business models and organizational strategies.

He was President of Burton Group from 1999 to 2010, the leading technology infrastructure research and consulting firm. Mr. Rowe grew Burton to over $30+ million in revenue on a self-funded basis, sold Burton to Gartner in 2010 and supported the acquisition as Burton President at Gartner.

Jeff Nichols has spent over 30 years as an executive, technologist, consultant and thought leader including 10 years in the health care space and 10+ years in the energy industry. He was most recently Director of IT at Sempra Energy leading the company’s efforts in networks, “big data”, technology infrastructure, compliance, information management, & enterprise architecture. Recent efforts he led include starting an RPA program, a wide-ranging technology transformation program and an application rationalization program.  He was formerly Executive Director at Kaiser Permanente after leading a consulting firm focused on IT in the health care industry.

Key areas of focus include AI, RPA, Machine Learning, innovation, IoT solutions, technology strategy & planning, ITIL operations, information security, business analysis, program management and system architecture.

Appendix 1: DARPA’s historic support for AI

DARPA has a history of supporting advanced efforts over a sustained period of time to drive innovation. The following graphic describes some of the major efforts taken on by DARPA over the past 4 decades in the AI area.