considerations for getting started with ai · considerations for getting started with ai ......

KEITH KIRKPATRICK Principal Analyst

CLINT WHEELOCK Managing Director

Considerations for Getting Started with AI Published 2Q 2018

WHITE PAPER

COMMISSIONED BY:

Considerations for Getting Started with AI

© 2018 Tractica LLC. All Rights Reserved.

2

SECTION 1 INTRODUCTION

Artificial intelligence (AI) is an umbrella term for multiple technologies that are designed to provide computers with human-like abilities of hearing, seeing, reasoning, and learning. These techniques, which include machine learning (ML), deep learning (DL), computer vision (CV), and natural language processing (NLP), unmask hidden patterns in large data sets, and then, using complex algorithms, can correlate findings between seemingly unrelated variables.

As AI gains traction, organizations are realizing that only larger-scale, enterprise-wide deployments are likely to provide full access to the operational and economic benefits of these new technologies. But enabling AI is not a plug-and-play proposition. Significant time, resources, and capital must be deployed, and in most cases, internal company teams are not experienced enough with AI, nor do they have the cutting-edge data science skills, software development expertise, or experience in selecting the right software platforms, hardware, and infrastructure to adequately embark upon a truly transformational AI implementation journey. Nevertheless, organizations can still take advantage of AI by tapping into internal operational knowledge and external expertise to bring AI solutions to market in a matter of months.

1.1 WHAT IS AI?

Figure 1.1 Artificial Intelligence Techniques

(Source: Cray, Inc.)

AI: Area of computer science that emphasizes the creation of intelligent machines that work and react like humans

Analytics: Search for “what, when, where and why” in data

Machine Learning: Predict future outcomes based on past observations from big data

Deep Learning: Train and use (infer) models that mimic the biology of the human brain, to interpret speech, images and text

ArtificialIntelligence

Big Data and Analytics

DeepLearning

MachineLearning



3

Before deploying an AI solution, organizations need to have a clear understanding of the underlying technologies and processes used to enable AI, as well as the benefits and limitations of potential solutions.

Machine Learning is a type of AI that involves using computerized mathematical algorithms that can learn from data and can depart from strictly following rule-based, pre-programmed logic. ML algorithms typically build a probabilistic model and then use it to make assumptions and predictions about similar sets of data.

Deep Learning is a form of ML that uses the model of human neural nets to make predictions about new data sets. Tractica believes this is currently the most promising of all AI technologies and is advancing other branches of the science, including cognitive computing, image recognition, and NLP.

Natural Language Processing enables computers to understand human language as it is spoken and written, and to produce human-like speech and writing. Machine translation of one human language into another language is also a form of NLP.

Computer Vision attempts to identify images of objects that can be seen. It can also include attempts to use the same technology to identify patterns in data, such as seismographic readings, which humans cannot see.

In each of the different methods of providing AI technology, the ability of the solution to provide “intelligence” is dependent upon the data’s quantity (more is better), granularity (greater segmentation is preferred), and quality (taken from reputable sources). Further, the complexity and structure of the algorithm itself will impact the results of the AI system, with the best solutions and algorithms generally being designed or tuned to identify hidden patterns, connections, and correlations in the data.

Each of these methods can be considered as steps on a continuum that moves from simply analyzing data to intelligently acting on data without human intervention. This is illustrated in Figure 1.2.



4

Figure 1.2 The Artificial Intelligence Continuum

(Source: Cray Inc.)

1.2 POTENTIAL BENEFITS OF USING MACHINE LEARNING AND DEEP LEARNING

ML and DL are likely to be the primary drivers of AI adoption, given that other AI technologies rely on ML and DL to make sense of the data captured. The key features of ML and DL are based on the ability to identify patterns in data, connect discrete data elements, and provide faster and more powerful analysis than humans or static analytics programs. As a result, enterprises are able to handle routine tasks more quickly and accurately, thereby increasing productivity and efficiency.

Furthermore, ML and DL can enable the development of systems that permit more intuitive, human-like processing of information, making it simpler and more intuitive for humans to interact with machines and technology.

The end benefit for enterprises is the ability to augment or replace functions that are time or resource intensive with automated, intelligent technology. This can lead to increased productivity and increased efficiency, and often can open up new technological, product, or service offerings that can directly improve a company’s bottom line.



5

SECTION 2 IMPLEMENTATION CONSIDERATIONS

2.1 UNDERSTANDING THE JOURNEY

Enterprises need to understand AI in the context of their own business, industry, application area, and technology choices. What is AI? Am I already using AI within my organization in the form of business and data analytics? Once these fundamental questions are addressed, enterprises need to understand that this is not a one-time implementation but rather a continuous process that will need constant adjustment. Implementing an AI solution or strategy is perhaps best considered as a continuum or journey, rather than a finite project with a beginning, middle, and end. While basic timelines, such as project start dates, can be fixed and adhered to, maximizing the investment in AI requires a larger investment of time and resources to allow the solution to run, and then time to apply various tweaks or adjustments to the algorithm or output to ensure it meets the changing needs of a business or process.

In particular, ML and DL use a constant feedback or learning loop to refine the algorithm and its output, and therefore, cannot be considered to be a finalized product or solution. Support and maintenance of the algorithm is crucial to ensuring that the results remain valid and optimized, particularly if changes to the business or process are being made based on the ongoing output of the algorithm. This is especially true when using AI to discover new and more efficient processes, or when using AI to handle real-time or near real-time analysis of a specific task or process.

2.2 KEY BUSINESS DRIVERS TO JUSTIFY AI INVESTMENT

The use of AI is commonly tied to a desire to improve the efficiency of a process, product, or service; reduce the cost of performing a task or process; or to generate additional revenue or profit based on the improvement of the efficiency of a task or product. These process or efficiency improvements are largely due to the ability of AI to uncover new and more efficient ways of handling or doing a task, finding new patterns or correlations in the data, and leading to more efficient ways of connecting disparate data points; or simply improving on human or pre-programmed processes by setting up the algorithm to learn from past actions, making adjustments in a far faster and more data-driven way than possible by humans.

2.3 COMPETITIVE DIFFERENTIATION

Perhaps the biggest reason AI is projected to gain favor among enterprises is the ability to differentiate a product or service offering. Using ML or DL, it is possible to tweak or modify an algorithm to stress or weight a specific factor or factors, thereby changing or biasing the result. For example, an auto insurance company that wishes to target younger drivers may wish to tweak its underwriting algorithm to de-emphasize “years licensed” when it considers an underwriting decision.

AI could also be used to spawn new business ideas or business models, creating new lines of revenue and spearheading competitive and product differentiation in the marketplace. AI should be viewed as a cognitive engine rather than just an analytics engine, allowing organizations to start thinking on the lines of a living and breathing entity that can adapt to new competitors, prices, customer demand, business models and supply chain disruptions. AI can help organizations acquire language and vision capabilities, parsing and understanding documents, images or video, giving them the ability to react and adapt to



6

shifts in business environment like never before. Organizations that are willing to look beyond cost efficiencies should follow a parallel path of treating AI as a strategic business tool.

2.4 PROJECT METRICS AND GUIDELINES

From an organizational and revenue-generating standpoint, AI is still in its infancy. Thus, there is a wide range of fee and billing structures used by services firms, platform vendors, and other market participants. However, generally speaking, when an organization sells a platform or software package and then adds on a services component, there is usually a software license or service and support model put in place, in order to cover the development of the software or tools. Some integration and customization service work may be covered under these fees, but as integrations have become more complex, there is usually an additional services component that is charged to address the significant labor costs required to ensure a smooth integration.

While the exact structure and terms of an AI engagement will vary based on the provider, the client, and the use case, most AI engagements will follow a fairly standard structure. Generally, projects are structured in a staged approach (see Figure 2.1) , starting with (1) the research and selection of a use case. This stage of the process is critical, as a clear business case, along with stated goals metrics, must be decided upon to ensure the project remains on track.

Then, a proof-of-concept (POC) program is initiated, which requires (2) selecting data sources to feed into an algorithm, (3) building or customizing a pre-built algorithm to the customer’s requirements, and then (4) running and testing the solution to ensure it yields the expected benefits.

Once the POC has been completed, which can take anywhere from about 3 to 4 months, the client and any services providers will (5) evaluate the program, and decide whether to expand the program or to try a new use case.

Figure 2.1 Proof-of-Concept Cycle

(Source: Tractica)



7

2.5 OWNERSHIP OF AI

The ownership of AI within an organization is key to how AI projects get implemented and managed on a day-to-day basis. Centralized management structures for AI require a Chief AI Officer in charge of AI projects, spearheading management and implementation of AI within an organization. Centralized AI management allows for a “paintbrush” approach to AI where AI can be applied to multiple silos within an organization, but risks central management not clearly understanding the key business metrics or needs within each of the silos. Centralized management is, however, a good approach to attracting the best talent, with AI engineers looking to join a company where the “Head of AI” is someone that has major standing in the AI community.

For the most part, decentralized AI is a better approach to follow as most companies don’t have a clear sense of how or where AI can help. Enterprises need to understand their enterprise and individual departmental goals first. A decentralized approach gives departments the freedom to define and drive their own AI strategy rather than have central management dictate terms. Also, in the long run as AI becomes the default way of running and managing a business, the CEO will be the default Chief AI Officer or Head of AI, and therefore having a separate business function for AI in a centralized approach separate from the CEO can turn out to be a short-sighted approach.

2.6 LEVEL OF INVESTMENT

While there are no hard-and-fast metrics on how much an initial AI deployment will cost, a pilot program can generally be started with an investment ranging from the low to mid-six figure range. The overall total investment can be impacted by the length of the pilot program, the complexity of the project, and the number of people working on the engagement.

It’s important to differentiate the cost of investing in AI, and the efficiency gain that is being targeted. While it is advisable to start with a reasonable investment, one should target a sizable business function with hundreds of millions or even billions of dollars in cost, where even a few percentage efficiency gains can lead to a meaningful return on investment.

Still, to take the successes of a pilot program and expand them across the organization to attain tangible benefits, additional investments in hardware, software, and support and maintenance will be required.

Increasingly, though, some vendors are tying service fees to specific milestones related to outcomes, which can include revenue increases, operational efficiency improvements, or other metrics that clearly demonstrate success. This model is particularly useful when a per-device licensing model would simply be too expensive for most organizations to handle.

The need for high-powered hardware that can handle the demands of DL is another key cost consideration for enterprises. Hardware vendors, of course, will price their offerings based on the power, size, and reliability of their solution, though the development of new AI computing hardware likely will help reduce the cost of hardware over time.

2.7 INFRASTRUCTURE CONSIDERATIONS

For some applications, AI solutions can be handled in the cloud, with data from an organization being transferred to a remote third party that handles the development, processing, and output for ML and DL processes. But for many applications and use cases where security and privacy are paramount, the organization will need to invest in its own AI infrastructure, which includes an AI platform, software, and hardware.



8

Moreover, training time for an individual AI model or a portfolio of models is extremely time-intensive. Cloud-based solutions generally charge enterprises by the compute hour, and paying for the amount of compute time required to train complex algorithms may wind up costing significantly more than simply purchasing the hardware and software outright.

Organizations that want to own and manage their own AI infrastructure will need to select an AI platform on which to build or install AI solutions. These platforms can range from well known, large frameworks such as TensorFlow from Google, to more specialized platforms, such as Ayasdi, which is a platform used to support the development of predictive models.

Then, the organization will need to acquire supercomputers powered by powerful, scalable processors and graphics processing unit (GPU) accelerators, to handle the large data sets required for training DL as well as the need for continuous, high-speed algorithmic processing power. Speed is paramount, as training a complex algorithm can take days or weeks, thereby consuming a significant amount of power. Moreover, time spent training the algorithm results in delaying the algorithm from being used for inference, or AI parlance for an algorithm that is actually being put to work. Of course, powerful GPUs offer significant benefits, as they are required for running billions of computations based on the trained network to identify known patterns or objects.

A key platform consideration needs to be ease of development and explanability of AI models. Most enterprise AI platforms today are moving away from code-based development, using drag and drop modules in graphical user interface (GUI) environments. Improved visualization helps to bring down development time and open AI development to a wider talent pool rather than just experienced data scientists. However, with drag and drop modules there is a risk that interpretability and explanability of AI models gets sidelined. As AI models start to take over key business decisions, and dictate how customers get serviced or targeted, companies need to pay attention to the problem of “black box” models. AI platforms need to provide interpretability functions in their models, like feature selection impact, which would help regulators or investors parse through the AI decision making process. The legal and regulatory requirements around AI model interpretability are expected to get tougher, and therefore AI platforms without model interpretability functions pose a risk.

Most importantly, the AI hardware solution needs to be robust enough to handle today’s applications, but also customizable, adaptable, and expandable to address future AI processing needs. Generally, organizations should select the latest generation of multi-core processors and GPUs in order to support the use of complex algorithms that may be integrating hundreds of discrete data inputs at once and ensure that any tweaks or modifications to the algorithm can be supported with adequate processing power.

As discussed previously, the performance of many algorithms will improve over time, as long as there is new data being fed into the model, which requires a large reserve of data storage and processing power. Most vendors offer a variety of storage options, but it is wise to select systems that can balance performance levels, scalability, and availability within the project’s budget level.



9

SECTION 3 CASE STUDY EXAMPLES

While AI utilization is far from ubiquitous, there is significant activity occurring within organizations at scale. By leveraging lessons learned during pilot programs, incorporating internal domain and process knowledge, and then integrating external hardware, software, and AI resources, organizations have been able to deploy commercially viable AI solutions. The following examples illustrate how ML and DL have been deployed successfully, providing organizations with tangible, real-world results.

3.1.1 OIL & GAS: LOCATING NEW RESERVES

The upstream oil & gas industry is focused on locating and producing crude oil and natural gas, and is often also referred to as the exploration and production (E&P) sector. Oil and gas can leverage AI in a number of ways, but in particular, planning and forecasting can be improved by using DL to help incorporate macroeconomic trends to drive investment decisions in exploration and production, taking into account economic, production, and weather patterns to drive investment decisions.

Applying AI in the operational planning and execution stages can significantly improve well planning, real-time drilling optimization, frictional drag estimation, and well cleaning predictions. AI techniques can also be applied in other activities, such as reservoir characterization, modeling, and field surveillance, to accurately characterize reservoirs in order to attain optimum production levels.

AI systems hold the key to pinpointing new drilling sites containing valuable crude oil, as ML and DL can transform basic data into valuable insights that can be used throughout the exploration and production process, including seismic, geology, drilling, petrophysics, reservoir, and production.

Geophysical feature detection is a critical part of the workflow in the oil & gas industry. Seismic surveys are carried out in the exploratory phase and during various other phases, from planning to field characterization before and during oil production. Once the data is gathered, the seismic traces are then processed and analyzed by human experts. Typically, this process can take several months.

Recently, Shell and the Massachusetts Institute of Technology (MIT) partnered to use AI techniques to automate this process and improve workflow efficiencies. Using DL, the raw seismic traces were analyzed to discover and locate subsurface faults in the underground structure, which are likely to contain hydrocarbons, before running migration and interpretation models. While there are still challenges in training and computational requirements, the study proved that geophysical feature detection could be automated.

Oil exploration capital expenditure is estimated to be around $100 billion per year, so any savings and efficiencies brought about by geophysical analysis is expected to be adopted widely across the oil & gas industry.

3.1.2 AUTOMOTIVE: OBJECT DETECTION AND RECOGNITION

Perhaps no other industry is as closely aligned with AI in the minds of consumers as the automotive industry, thanks to the media’s focus on self-driving cars. But AI is being used in other use cases that can support human or self-driving and address maintenance issues,



10

personalization services, and ride-sharing services.

The most well publicized use of AI in vehicles is the use of object detection and classification, which takes sensor data, often from cameras, and then uses complex algorithms to classify these objects so that the AI system can then “learn” their characteristics and recognize them in real time. The challenge is not in capturing images, as today’s high-definition (HD) cameras can present images in stunningly clear detail. However, in a moving environment, objects can appear to change size as a vehicle or camera approaches. The angle at which an object is viewed can also skew its appearance, and the presence of other factors (rain, bright sunlight, low lighting, glare, dirt, snow, or any other number of obstructions) can alter the appearance of an object, making it hard to accurately and consistently identify the object.

This is an area where machine vision and ML can provide invaluable support. By capturing a wide range of images of objects from a variety of vantage points, angles, and in different conditions, a repository of images that can be definitively classified as that object can be created, and used to “train” a ML system to identify and classify objects that resemble objects in the repository. By then assigning various other attributes to each object, such as whether the object is informational like a traffic sign, whether it is permanent or temporary like a road barrier, or whether or not it has the capability of motion and how it typically moves, the system can begin to develop logical rules on handling each object and the rules for dealing with them.

Luxury automaker Mercedes-Benz has introduced AI technology into its 2018 S-class sedan, via its Drive Pilot driver-assistance features. These semi-autonomous driver-assist technologies include adaptive cruise control, a lane-keeping setup that can handle limited steering duties and autonomous lane changes, and a 360° array of radar and ultrasonic sensors for keeping track of lane markings, other cars, and road signs.

The system uses DL technology to train the system to distinguish between various elements that may be in the car’s operational envelope (such as lane markings, signs, roadside barriers, other vehicles, pedestrians, and animals) to provide input into the system. After ingesting images of these objects during training, the DL technology can then extrapolate and learn what these images may look like from different angles, when lighting conditions change, or when the car is moving at different speeds.

Thanks to DL, driver-assistance systems can be brought to market more quickly than if an object-detection system needed to be fed every object that could be encountered, from all angles and in all lighting conditions.

3.1.3 LIFE SCIENCES: DRUG COMPOUND DISCOVERY

In healthcare, AI is largely being implemented as a tool to more efficiently and accurately review data, and uncover patterns in the data that can be used to improve analyses, uncover inefficiencies, and streamline care, from both a clinical and an operational perspective. The overarching driver is to provide better care for humans, while reducing costs and administrative headaches and bottlenecks. Some of the use cases that have shown promise and results include the use of ML to analyze huge clinical and genomic databases and identify relevant predictive biomarkers for specific types of diseases.

Researchers are using a range of data sets (e.g., genomic data, gene expressions, proteomics, clinical data, etc.), and integrating signals and timeframes from this data to develop molecular profiles. In some cases, humans provide canonical disease or drug maps to cover various therapeutic areas and disease types. Hundreds of canonical pathways are analyzed and enriched to infer a disease or drug’s mechanisms of actions (MOA). From



11

here, molecular profiling data is fed into algorithms that use ML to identify biomarkers and/or drug sensitivity to specific biomarkers.

AI can also be used to increase the speed and efficiency of drug discovery and testing. AI offers new ways for researchers to leverage existing databases, develop new databases involving bigger and more diverse data, and to predict how molecules will behave and how likely they are to make a useful drug, thereby saving time and money on unnecessary tests. DL could help with drug development by finding patterns in sparse pathology data combined with large genomic data sets.

The ability of AI to find, identify, and analyze patterns is also yielding benefits in analyzing medical images. Historically, analyzing medical images has been difficult, highly prone to human error or oversight, and time-consuming and costly. Medical images like magnetic resonance imaging (MRIs), X-rays, computed tomography (CT) scans, and other diagnostic images are essential to better understanding and diagnosing a wide range of conditions. When it comes to diagnosing critical conditions, including cancer, neurodegeneration, and heart disease, the faster and smarter the speed, precision, and predictive capabilities, the better. Analyzing images is a strong application for DL and CV within the realm of patient data processing. In particular, DL is now being applied to automate the analysis and increase accuracy, precision, and understanding of images down to the pixel.

The methods for drug discovery, the process by which new medications are discovered, has largely centered around identifying the active ingredient from traditional remedies simply by serendipity. Upon sequencing the human genome (which enabled rapid cloning and synthesis of large quantities of purified proteins), it has become common to use high-throughput screening of large compounds’ libraries against isolated biological targets. New drug development costs still run about $2.6 billion per year and take as long as 14 years, and less than 10% of potential medications make it to market, according to research from Tufts University and the U.S. Food and Drug Administration (FDA).

AI offers new ways for researchers to leverage existing databases, develop new databases involving bigger and more diverse data, and predict how molecules will behave and how likely they are to make a useful drug, thereby saving time and money on unnecessary tests. Many large pharmaceutical companies are partnering with AI drug discovery startups in a bid to reduce costs and time to market.

GlaxoSmithKlein (GSK) recently announced a $43 million partnership with Exscientia to search for drug candidates for up to 10 disease-related targets. Atomwise recently partnered with drug giant Merck and published first findings on Ebola treatment drugs last year. BenevolentAI is a British company focused on developing better drugs to target diseases of inflammation and neurodegeneration, and rare cancers. The idea is to use much of the dark data within pharma research and development (R&D) organizations and apply vast data sets available on human health and biological systems to DL systems that learn and reason from interaction between human judgement and data. Numerous other companies are emerging in this space, such as Calico, Numerate, Globavir, NuMedi, twoXAR, and Cloud Pharmaceuticals.

3.1.4 FINANCIAL SERVICES: INSURANCE

The financial services market is awash in data; transaction data used for inter-institution and market-making activities, product sales, customer data, and operational data used for managing the day-to-day operations of an institution, managing security, and marketing operations. The ability to harness this data, identify patterns, and create new efficiencies is a prime driver of AI technology in the financial services industry. While it is a highly regulated



12

industry, financial services firms have made the case that efficiency and safety can, in many cases, be better handled by machines than humans, because they are not able to be swayed by emotion, stress, or other outside factors.

For example, credit rating agencies are now beginning to explore AI, ML, and DL to aid in credit scoring, primarily to assess creditworthiness more precisely through more nuanced evaluations of data. Instead of looking at one or a few separate variables, AI engines help consider mitigating interactions between multiple variables. For instance, even if a consumer skipped payments on two debts within 24 months, but paid consistently for 12 months straight, and obtained new lines of credit, that may be weighted to mitigate the risk of the past missed payments. The other potential benefit is to consider people who might not have been able to get a score in the past, via traditional logistic regression-based scoring (which looks at credit history).

Using NLP, ML, and DL, in some cases, companies are using AI to automatically generate reports that can handle the identification and extraction of data from relevant internal and external data sources. Using ML, the system can then apply predictive modeling and data enrichment, and then create hundreds of “what if” scenarios and perform trend analysis.

An insurance company’s primary objective is to use and process customer data to model risk factors, improve insurance products, prevent losses and fraud, and reduce the amount of money it pays out. In the near term, insurance companies are seeing potential for reducing fraud by detecting anomalies or patterns associated with fraudulent activity. Algorithms can also help speed up claims processing by automatically assessing the severity of a claim and predicting costs from historical data, sensors, images, or other data sources.

FitSense is a company focused on personalizing insurance products by using app and device data. It has built a data aggregation platform that integrates, processes, and securely stores data across various channels (e.g., wearables, biometrics, health apps, demographic data, etc.). It uses ML and NLP to model and interpret raw data into specific customer and risk profiles, and then leverages that data to help insurance companies design and substantiate new insurance products and services.

In particular, the app enables insurance companies to offer their own white-labeled self-quantification, health management, and incentive programs. Other companies developing in this space include DreamQuark, Big Cloud Analytics, and CogniCor, which offers a chatbot assistant for complaints and claims resolution and then uses interactions to improve insurance products and services.



13

3.1.5 CONSUMER: CUSTOMER RECOMMENDATIONS

AI technology is also being used with consumer products and in the retail environment to make life more convenient or to improve the efficiency of operations. By harnessing the power of ML or DL, products, attributes, and shopping-related advertisements and marketing information can be tailored to granular actions and triggers, creating a truly personalized product experience. Furthermore, because DL algorithms can be used to uncover new connections between data points, seemingly unrelated consumer behavior or data can be more efficiently tracked and harnessed by retailers, marketers, and product manufacturers on an ongoing basis.

Sentiment analysis, which involves understanding the emotional context of buyers can be very useful for gaining an overview of public option, ideation, or feedback on a given topic. Common approaches for measuring brand sentiment include the net promoter score (NPS), up/down votes, emojis, basic Likert scales, or other measures.

However, AI and NLP are now enhancing sentiment analysis by capturing and understanding the unstructured, more nuanced, and qualitative feedback, not just the best fitting response in a multiple-choice scenario. This data is combined with structured data sets for advanced analytics to surface trends. For example, retailers can track social media sentiment analysis, and then use NLP to dig deeply into the rich nuances of comments and feedback.

The ability to see beyond simple happy-neutral-angry or like-dislike then allows retailers to plan and act according to far more nuanced categories, personas, product lines, or campaigns. The majority of data available to most organizations is “dark,” unstructured, and unused, but potentially full of valuable insights, so AI can be used as a tool to shed light on sentiments found in call logs, emails, transcripts, videos, rating applications, and audio data.

A large pharmaceutical company interested in optimizing C Space, its online community of caregivers for people with schizophrenia, recently partnered with AI software company Luminoso to better understand the major issues these caregivers face and how to provide them with better resources and communications. Together, they used NLP and deep analytics on vast amounts of rich, but disparate and unstructured data, pulling together content from online communities, online discussion boards, multiple research projects, photo collages, and open-ended responses from surveys.

Luminoso’s software then vectorized the data, meaning it effectively turned the text into mathematical vectors, and then mapped unstructured data based on relationships between topics and ideas. It uncovered key themes and associations about the emotional composition of caretakers, their struggles, concerns, resource needs, and how they change over time. The pharmaceutical company also used the findings to improve community management, messaging, and support services.



14

3.2 CONCLUSIONS AND RECOMMENDATIONS

As organizations face an expanding number of AI projects, the demands on systems supporting ML and DL are growing in lockstep. Therefore, it is important to select platforms, software, hardware, and service provides that can scale with the growth of an AI deployment.

Partnering with vendors that can provide their particular platform or product, and provide honest, technology- and vendor-agnostic consulting, support, and maintenance services is a recipe for success. Assuming the vendor has been involved in the AI space, it likely has significant data-science expertise, as well as real-world experience with handling deployments, including how to overcome the inevitable hurdles and challenges.



15

SECTION 4 ACRONYM AND ABBREVIATION LIST

Computed Tomography ............................................................................................................................. CT

Computer Vision ......................................................................................................................................... CV

Deep Learning ............................................................................................................................................. DL

Exploration & Production .......................................................................................................................... E&P

Food and Drug Administration (U.S.) ....................................................................................................... FDA

Graphical User Interface ........................................................................................................................... GUI

Graphics Processing Unit ........................................................................................................................ GPU

High Definition ............................................................................................................................................ HD

Machine Learning ....................................................................................................................................... ML

Machine Reasoning ................................................................................................................................... MR

Magnetic Resonance Imaging .................................................................................................................. MRI

Massachusetts Institute of Technology ..................................................................................................... MIT

Mechanisms of Actions ........................................................................................................................... MOA

National Language Processing ................................................................................................................ NLP

Net Promoter Score ................................................................................................................................. NPS

Proof-of-Concept ..................................................................................................................................... POC

Research and Development.....................................................................................................................R&D

Software-as-a-Service ............................................................................................................................ SaaS



16

SECTION 5 TABLE OF CONTENTS

SECTION 1 ...................................................................................................................................................... 2 Introduction ................................................................................................................................................. 2

1.1 What Is AI? .................................................................................................................................... 2 1.2 Potential Benefits of Using Machine Learning and Deep Learning .............................................. 4

SECTION 2 ...................................................................................................................................................... 5 Implementation Considerations ................................................................................................................ 5

2.1 Understanding the Journey ........................................................................................................... 5 2.2 Key Business Drivers to Justify AI Investment .............................................................................. 5 2.3 Competitive Differentiation ............................................................................................................ 5 2.4 Project Metrics and Guidelines ..................................................................................................... 6 2.5 Ownership of AI ............................................................................................................................. 7 2.6 Level of Investment ....................................................................................................................... 7 2.7 Infrastructure Considerations ........................................................................................................ 7

SECTION 3 ...................................................................................................................................................... 9 Case Study Examples ................................................................................................................................. 9

3.1.1 Oil & Gas: Locating New Reserves ......................................................................................... 9 3.1.2 Automotive: Object Detection and Recognition ....................................................................... 9 3.1.3 Life Sciences: Drug Compound Discovery ............................................................................ 10 3.1.4 Financial Services: Insurance ............................................................................................... 11 3.1.5 Consumer: Customer Recommendations ............................................................................. 13

3.2 Conclusions and Recommendations .......................................................................................... 14 SECTION 4 .................................................................................................................................................... 15 Acronym and Abbreviation List ............................................................................................................... 15 SECTION 5 .................................................................................................................................................... 16 Table of Contents ...................................................................................................................................... 16 SECTION 6 .................................................................................................................................................... 17 Table of Charts and Figures..................................................................................................................... 17 SECTION 7 .................................................................................................................................................... 18 Scope of Study .......................................................................................................................................... 18 Sources and Methodology ....................................................................................................................... 18 Notes .......................................................................................................................................................... 19



17

SECTION 6 TABLE OF CHARTS AND FIGURES

Chart 7.1 Tractica Research Methodology ............................................................................................ 19

Figure 1.1 Artificial Intelligence Techniques ............................................................................................. 2 Figure 1.2 The Artificial Intelligence Continuum ....................................................................................... 4 Figure 2.1 Proof-of-Concept Cycle ........................................................................................................... 6



18

SECTION 7 SCOPE OF STUDY

This white paper examines the hardware, software, and services used to deploy AI technology within commercial enterprises, government entities, and other large organizations. The technologies covered include ML, DL, NLP, CV, MR, and CV. The paper uses insights from relevant Tractica reports, including Artificial Intelligence for Enterprise Applications and Artificial Intelligence Services, while the case examples discuss real-world implementations of AI technology.

SOURCES AND METHODOLOGY

Tractica is an independent market research firm that provides industry participants and stakeholders with an objective, unbiased view of market dynamics and business opportunities within its coverage areas. The firm’s industry analysts are dedicated to presenting clear and actionable analysis to support business planning initiatives and go-to-market strategies, utilizing rigorous market research methodologies and without regard for technology hype or special interests including Tractica’s own client relationships. Within its market analysis, Tractica strives to offer conclusions and recommendations that reflect the most likely path of industry development, even when those views may be contrarian.

The basis of Tractica’s analysis is primary research collected from a variety of sources including industry interviews, vendor briefings, product demonstrations, and quantitative and qualitative market research focused on consumer and business end-users. Industry analysts conduct interviews with representative groups of executives, technology practitioners, sales and marketing professionals, industry association personnel, government representatives, investors, consultants, and other industry stakeholders. Analysts are diligent in pursuing interviews with representatives from every part of the value chain in an effort to gain a comprehensive view of current market activity and future plans. Within the firm’s surveys and focus groups, respondent samples are carefully selected to ensure that they provide the most accurate possible view of demand dynamics within consumer and business markets, utilizing balanced and representative samples where appropriate and careful screening and qualification criteria in cases where the research topic requires a more targeted group of respondents.

Tractica’s primary research is supplemented by the review and analysis of all secondary information available on the topic being studied, including company news and financial information, technology specifications, product attributes, government and economic data, industry reports and databases from third-party sources, case studies, and reference customers. As applicable, all secondary research sources are appropriately cited within the firm’s publications.

All of Tractica’s research reports and other publications are carefully reviewed and scrutinized by the firm’s senior management team in an effort to ensure that research methodology is sound, all information provided is accurate, analyst assumptions are carefully documented, and conclusions are well-supported by facts. Tractica is highly responsive to feedback from industry participants and, in the event errors in the firm’s research are identified and verified, such errors are corrected promptly.



19

Chart 7.1 Tractica Research Methodology

(Source: Tractica)

NOTES

CAGR refers to compound average annual growth rate, using the formula:

CAGR = (End Year Value ÷ Start Year Value)(1/steps) – 1.

CAGRs presented in the tables are for the entire timeframe in the title. Where data for fewer years are given, the CAGR is for the range presented. Where relevant, CAGRs for shorter timeframes may be given as well.

Figures are based on the best estimates available at the time of calculation. Annual revenues, shipments, and sales are based on end-of-year figures unless otherwise noted. All values are expressed in year 2018 U.S. dollars unless otherwise noted. Percentages may not add up to 100 due to rounding.

PRIMARYRESEARCH

SECONDARYRESEARCH

SUPPLY SIDE DEMAND SIDE

Industry Interviews

Vendor Briefings

Product Evaluations

End‐User Surveys

End‐User Focus Groups

Company News & Financials

Technology & Product Specs

Government & Economic Data

Case Studies

Reference Customers

QUALITATIVEANALYSIS

QUANTITATIVEANALYSIS

Company Analysis

Business Models

Competitive Landscape

Technology Assessment

Applications & Use Cases

MarketSizing

Market Segmentation

Market Forecasts

Market Share Analysis

Scenario Analysis

MARKET RESEARCH

MARKET ANALYSIS



20

Published 2Q 2018

© 2018 Tractica LLC

1650 38th Street, Suite 101E

Boulder, CO 80301

Tel: +1.303.248.3000

Email: [email protected]

www.tractica.com

This publication is provided by Tractica LLC (“Tractica”). This publication may be used only as expressly permitted by license from Tractica and may not otherwise be reproduced, recorded, photocopied, distributed, displayed, modified, extracted, accessed or used without the express written permission of Tractica. Notwithstanding the foregoing, Tractica makes no claim to any Government data and other data obtained from public sources found in this publication (whether or not the owners of such data are noted in this publication). If you do not have a license from Tractica covering this publication, please refrain from accessing or using this publication. Please contact Tractica to obtain a license to this publication.

considerations for getting started with ai · considerations for getting started with ai ......

Documents