Applied Artificial Intelligence Handbook for Business Leaders PDF

“The authors of Applied AI have incredible depth of expertise and experience in AI, and they make the complex topic accessible to everyone. It’s rare to find technology experts as engaging and thought-provoking.” STEPHEN STRAUSS Head of Sales Enablement and Insights, PayPal “This book cuts the fluff and arms business leaders with exactly the right foundational knowledge to lead successful AI initiatives at their companies. It’s hands down the best playbook for executives starting on their automation journey.” JACK CHUA Director of Data Science, Expedia “As a deep learning researcher and educator, I’m alarmed by how much misinformation and misreporting occurs with AI. It’s refreshing to see a practical guide written by experienced technologists which explains AI so well for a business audience. In particular, I’m glad to see this book addresses critical issues of AI safety and ethics and advocates for diversity and inclusion in the industry.” RACHEL THOMAS Co-Founder, Fast.ai and Assistant Professor, USF Data Institute “Full of valuable information and incredibly readable. This book is the perfect mix of practical and technical. If you’re an entrepreneur or business leader, you need this guide.” JEFF PULVER Co-Founder, Vonage and MoNage “Applied AI is the perfect primer for anyone looking to understand the enterprise implications of emerging artificial intelligence technology—a must read for any business leader intending to stay ahead.” ALEX STEIN Sr. Director, Strategy & Business Development, Viacom International Media Networks “Recent progress in AI will dramatically impact all aspects of a business. This excellent book provides a practical examination of how to harness disruptive technologies to achieve scalable and sustainable business success.” STEVEN KUYAN Managing Director, NYU Tandon Future Labs Applied Artificial Intelligence A HANDBOOK FOR BUSINESS LEADERS Copyright © 2018 by TOPBOTS Inc. All rights reserved. No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the publisher, except in the case of brief quotations embodied in critical reviews and certain other noncommercial uses permitted by copyright law. For permission requests, contact the publisher “Attention: Permissions Coordinator,” at the website address below. Printed in the United States of America First Printing December 2017 ISBN 978-0-9982890-2-1 (Paperback) ISBN 978-0-9982890-5-2 (Kindle) Edited by Natalia Zhang Cover Illustration by Vanessa Maynard Interior Design by Vanessa Maynard Website: www.appliedaibook.com Email: [email protected] For all of you who build technology to make tomorrow better than today. TABLE OF CONTENTS APPLIED ARTIFICIAL INTELLIGENCE WHO THIS BOOK IS FOR How to Use This Book WHAT BUSINESS LEADERS NEED TO KNOW ABOUT ARTIFICIAL INTELLIGENCE 1. BASIC TERMINOLOGY IN ARTIFICIAL INTELLIGENCE AI vs. AGI Modern AI Techniques 2. THE MACHINE INTELLIGENCE CONTINUUM Systems That Act Systems That Predict Systems That Learn Systems That Create Systems That Relate Systems That Master Systems That Evolve 3. THE PROMISES OF ARTIFICIAL INTELLIGENCE Microfinance Social Justice Medical Diagnosis 4. THE CHALLENGES OF ARTIFICIAL INTELLIGENCE The Effects of Discrimination Malicious AI 5. DESIGNING SAFE AND ETHICAL AI Ethics and Governance Education as Remedy Collaborative Design HOW TO DEVELOP AN ENTERPRISE AI STRATEGY 6. BUILD AN AI-READY CULTURE Be Honest About Your Readiness Choose the Right Champions Build An Enterprise-Wide Case For AI Why You Need a Multi-Disciplinary “AI SWAT Team” Get Organizational Buy-In Educate Your Stakeholders 7. INVEST IN TECHNICAL TALENT Understand Different Job Titles Seek the Right Characteristics Optimize Recruiting Strategies Emphasize Your Company’s Unique Advantages 8. PLAN YOUR IMPLEMENTATION Rank Business Goals Perform Opportunity Analysis AI Strategy Framework Know Your Data and Analytics Technical Prerequisites Build vs. Buy Calculate ROI and Allocate Budget Pick the Right “True North” Metric 9. COLLECT AND PREPARE DATA Data Is Not Reality Common Mistakes With Data 10. BUILD MACHINE LEARNING MODELS AI Is Not a Silver Bullet Assessing the Performance of Your Models Common Mistakes With Machine Learning Models Machine Learning Workflow Maintain an Experimental Mindset 11. EXPERIMENT AND ITERATE Agile Development Technical Debt Deployment and Scaling Iteration and Improvement AI FOR ENTERPRISE FUNCTIONS 12. OBSTACLES AND OPPORTUNITIES Current Obstacles What AI Can Do for Enterprise Functions 13. GENERAL AND ADMINISTRATIVE Finance and Accounting Legal and Compliance Records Maintenance General Operations 14. HUMAN RESOURCES AND TALENT Matching Candidates to Positions Managing the Interview Process Intelligent Scheduling Career Planning and Retention Risk Analysis Administrative Functions 15. BUSINESS INTELLIGENCE AND ANALYTICS Data Wrangling Data Architecture Analytics 16. SOFTWARE DEVELOPMENT 17. MARKETING Digital Ad Optimization Recommendations and Personalization 18. SALES Customer Segmentation Lead Qualification and Scoring Sales Development Sales Analytics 19. CUSTOMER SUPPORT Conversational Agents Social Listening Customer Churn Lifetime Value 20. THE ETHICS OF ENTERPRISE AI SUMMARY AND ADDITIONAL RESOURCES ACKNOWLEDGEMENTS AUTHOR AND EDITOR BIOGRAPHIES Mariya Yao Adelyn Zhou Marlene Jia Natalia Zhang APPLIED ARTIFICIAL INTELLIGENCE WHO THIS BOOK IS FOR Applied Artificial Intelligence is a practical guide for business leaders who are passionate about leveraging machine intelligence to enhance the productivity of their organizations and the quality of life in their communities. If you love to drive innovation by combining data, technology, design, and people, and to solve real problems at an enterprise scale, this is your playbook. There are plenty of technical tomes on the market for engineers and researchers who want to master the nitty-gritty details of modern algorithms and toolsets. You can also find plenty of general interest content for the public about the impact of AI on our society and the future of work. This book is a balance between the two. We won’t overload you with details on how to debug your code, but we also won’t bore you with endless generalizations that don’t help you make concrete business decisions. Instead, we teach you how to lead successful AI initiatives by prioritizing the right opportunities, building a diverse team of experts, conducting strategic experiments, and consciously designing your solutions to benefit both your organization and society as a whole. How to Use This Book The first part of this book, “What Business Leaders Need to Know,” gives executives an essential education in the state of artificial intelligence today. We recommend reading this part in full before pursuing AI projects for your organization. Chapters 1 and 2 provide a non-technical introduction to AI, the techniques used to power modern AI systems, and the functional differences between different levels of machine intelligence. While you do not need to memorize every detail, a passing familiarity with technical definitions will help you separate hype from reality when evaluating a project proposal for your own organization. Chapters 3, 4, and 5 describe promising applications of AI in society as well as challenges that arise from biased or unethical algorithms. You’ll learn how collaborative design is essential to ensuring that we build benevolent AI systems. In the second part of our book, “How to Develop an Enterprise AI Strategy,” we walk you through the strategic and methodological steps required to implement successful AI projects for your company. These chapters act as a reference guide as you are building your initiatives. Read through them once to familiarize yourself with the content, and then refer back to specific sections as needed during your projects. Chapters 6 and 7 teach you how to prepare your organization to succeed in AI projects. You will learn strategies to manage important stakeholders and attract technical talent. In Chapter 8, we guide you through exercises that will help you to identify opportunities for AI adoption within your organization and develop a business plan for implementation and deployment. Chapters 9, 10, and 11 explain common technical challenges you will encounter in building AI and how to overcome them. The last section of our book, “AI For Enterprise Functions,” highlights popular AI applications for common business functions. Chapter 12 summarizes some of the challenges of adopting AI solutions for enterprises. Chapters 13 and 14 introduce common AI applications in essential administrative functions like finance, legal, and HR, while Chapters 15 and 16 describe how machine learning can dramatically improve business intelligence, analytics, and software development. Chapters 17, 18, and 19 focus on the revenue-generating functions of sales, marketing, and customer service. Finally, Chapter 20 emphasizes the ethical responsibility that you, as business and technology leaders, have towards your workforce as well as towards ensuring that any technologies that you build have a benevolent impact on your customers, employees, and society as a whole. Because AI technologies evolve very quickly, we created an educational website, appliedaibook.com, where we offer updated content and detailed case studies for specific industries. Supplemental content for this book can be found in our resources section at appliedaibook.com/resources. We also created social communities and discussion forums for our readers to connect with us and each other, which you can join by visiting appliedaibook.com/community. What Business Leaders Need to Know About Artificial Intelligence 1. BASIC TERMINOLOGY IN ARTIFICIAL INTELLIGENCE Think about the most intelligent person you know. What about this person leads you to describe him or her this way? Is she a quick thinker, able to internalize and apply new knowledge immediately? Is he highly creative, able to endlessly generate novel ideas that you’d never think of? Perhaps she’s highly perceptive and hones in on the tiniest details of the world around her. Or maybe he’s deeply empathetic and understands how you’re feeling even before you do. Human intelligence spans a wide spectrum of modalities, exhibiting abilities such as logical, spatial, and emotional cognition. Whether we are math geniuses or charismatic salesmen, we must utilize cognitive abilities like working memory, sustained attention, category formation, and pattern recognition to understand and succeed in the world every day. Though computers trounce humans at large-scale computational tasks, their expertise is narrow, and machine capability lags behind human intelligence in other areas. The rest of this chapter will help you to understand the state of artificial intelligence today. AI vs. AGI Artificial intelligence, also known as AI, has been misused in pop culture to describe almost any kind of computerized analysis or automation. To avoid confusion, technical experts in the field of AI prefer to use the term Artificial General Intelligence (AGI) to refer to machines with human-level or higher intelligence, capable of abstracting concepts from limited experience and transferring knowledge between domains. AGI is also called “Strong AI” to differentiate from “Weak AI” or “Narrow AI,” which refers to systems designed for one specific task and whose capabilities are not easily transferable to other systems. We go into more detail about the distinction between AI and AGI in our Machine Intelligence Continuum in Chapter 2. Though Deep Blue, which beat the world champion in chess in 1997, and AlphaGo, which did the same for the game of Go in 2016, have achieved impressive results, all of the AI systems we have today are “Weak AI.” Narrowly intelligent programs can defeat humans in specific tasks, but they can’t apply that expertise to other tasks, such as driving cars or creating art. Solving tasks outside of the program’s original parameters requires building additional programs that are similarly narrow. “We’re very far from having machines that can learn the most basic things about the world in the way humans and animals can,” said Yann LeCun, head of AI at Facebook, in an interview with The Verge. 1 “In particular areas machines have superhuman performance, but in terms of general intelligence we’re not even close to a rat.” In addition, the path towards AGI is also unclear. Approaches that work well for solving narrow problems do not generalize well to tasks such as abstract reasoning, concept formulation, and strategic planning—capabilities that even human toddlers possess but our computers do not. Modern AI Techniques We are often asked to explain the key differences between machine learning, data science, AI, deep learning, etc. All of these are examples of machine intelligence, but they vary in their usage and potential impact. While engineers and researchers must master the subtle differences between various technical approaches, business and product leaders should focus on the ultimate goal and real-world results of machine learning models. This section is a guide to today’s most popular techniques, but methodologies are constantly evolving. You don’t need to memorize the guide, but you should try to gain a passing familiarity with the basic characteristics of each technique. In general, most enterprise-scale technologies use a wide range of automation methodologies, but not all of them count as AI. Differentiating between methods that are AI and those that are not can be tricky, and there is often overlap. You will find that simpler approaches often outperform complex ones in the wild, even if they’re intellectually less “advanced.” Though AI refers to a larger umbrella of computational techniques, the most successful modern AI solutions are powered by machine learning algorithms. For simplicity, we use AI and machine learning as interchangeable terms in this book. STATISTICS AND DATA MINING Statistics is the discipline concerned with the collection, analysis, description, visualization, and drawing of inferences from data. Its focus is on describing the properties of a dataset and the relationships that exist between data points. Statistics is generally not considered part of AI, but many statistical techniques form the foundation for more advanced machine learning techniques or are used in conjunction with them. Descriptive statistics describes or visualizes the basic features of the data being studied. A simple application could be to find the best-selling retail item in a store in a specific period of time. Inferential statistics is used to draw conclusions that apply to more than just the data being studied. This is necessary when analysis must be conducted on a smaller, representative dataset when the true population is too large or difficult to study. Because the analysis is done on a subset of the total data, the conclusions that can be reached with inferential statistics are never 100 percent accurate and are instead only probabilistic bets. Election polling, for example, relies on surveying a small percentage of citizens to gauge the sentiments of the entire population. As we saw during the 2016 US election cycle, conclusions drawn from samples may not reflect reality!2 Data mining is the automation of exploratory statistical analysis on large- scale databases, though the term is often used to describe any kind of algorithmic data analysis and information processing, which may also include machine learning and deep learning techniques. The goal of data mining is to extract patterns and knowledge from large-scale datasets so that they can be reshaped into a more understandable structure for later analysis. SYMBOLIC AND EXPERT SYSTEMS Symbolic systems are programs that use human-understandable symbols to represent problems and reasoning.3 The most successful form of symbolic systems is the expert system, which mimic the decision-making process of human experts. Expert systems are comprised of a series of production rules, similar to if-then statements, that govern how the program accesses a knowledge base and makes inferences. Rule-based expert systems are most effective when applied to automated calculations and logical processes where rules and outcomes are relatively clear. As decision-making becomes more complex or nuanced, explicitly formalizing the full range of requisite knowledge and inference schemes required to make human-level decisions becomes impossible. The rules engine and knowledge base for any expert system must be hand- engineered by domain experts. This is a huge drawback due to the limited number of experts who can perform the task and the time needed to program such a complicated system. The “completeness” of the knowledge base is questionable and will require continued maintenance (another huge drawback that requires enormous expenditures), and the accuracy of the system is overly-dependent on expert opinions that could be wrong. While symbolic systems are historically not scalable or adaptable, recent research has investigated combining them with newer methods like machine learning and deep learning to improve performance. MACHINE LEARNING What happens if you want to teach a computer to do a task, but you’re not entirely sure how to do it yourself? What if the problem is so complex that it’s impossible for you to encode all of the rules and knowledge upfront? Machine learning enables computers to learn without being explicitly programmed. It is a field in computer science that builds on top of computational statistics and data mining. This book will focus primarily on discussing how machine learning is being applied in different industries across different functions, so you’ll want to understand the broad categories in this field and how they are applied to business problems. Supervised learning occurs when the computer is given labeled training data, which consists of paired inputs and outputs (e.g. an image of a cat correctly labeled as “cat”), and learns general rules that can map new inputs to the correct output. Supervised learning is commonly used for classification, where inputs are divided into discrete and unordered output categories, and for regression, where inputs are used to predict or estimate outputs that are numeric values. If you are trying to predict whether an image is of a cat or a dog, this is a classification problem with discrete classes. If you are trying to predict the numerical price of a stock or some other asset, this can be framed as a regression problem with continuous outputs. Unsupervised learning occurs when computers are given unstructured rather than labeled data, i.e. no input-output pairs, and asked to discover inherent structures and patterns that lie within the data. One common application of unsupervised learning is clustering, where input data is divided into different groups based on a measure of “similarity.” For example, you may want to cluster your LinkedIn or Facebook friends into social groups based on how connected they are to each other. Unlike supervised learning, the groups are not known in advance, and different measures of similarity will produce different results. Semi-supervised learning lies between supervised and unsupervised learning. Many real-world datasets have noisy, incorrect labels or are missing labels entirely, meaning that inputs and outputs are paired incorrectly with each other or are not paired at all. Active learning, a special case of semi-supervised learning, occurs when an algorithm actively queries a user to discover the right output or label for a new input. Active learning is used to optimize recommendation systems, like the ones used to recommend movies on Netflix or products on Amazon. Reinforcement learning is learning by trial-and-error, in which a computer program is instructed to achieve a stated goal in a dynamic environment. The program learns by repeatedly taking actions, measuring the feedback from those actions, and iteratively improving its behavioral policy. Reinforcement learning can be successfully applied to game-playing, robotic control, and other well-defined and contained problems. It is less effective with complex, ambiguous problems where rewards and environments are not well understood or quantified. Chapter 10 discusses the mechanics of building machine learning models in more detail. You can also find updated technical resources on our book website, appliedaibook.com. DEEP LEARNING Deep learning is a subfield of machine learning that builds algorithms by using multi-layered artificial neural networks, which are mathematical structures loosely inspired by how biological neurons fire. Neural networks were invented in the 1950s, but recent advances in computational power and algorithm design—as well as the growth of big data—have enabled deep learning algorithms to approach human-level performance in tasks such as speech recognition and image classification. Deep learning, in combination with reinforcement learning, enabled Google DeepMind’s AlphaGo to defeat human world champions of Go in 2016, a feat that many experts had considered to be computationally impossible. Much media attention has been focused on deep learning, and an increasing number of sophisticated technology companies have successfully implemented deep learning for enterprise-scale products. Google replaced previous statistical methods for machine translation with neural networks to achieve superior performance.4 Microsoft announced in 2017 that they had achieved human parity in conversational speech recognition.5 Promising computer vision startups like Clarifai employ deep learning to achieve state-of-the-art results in recognizing objects in images and video for Fortune 500 brands.6 While deep learning models outperform older machine learning approaches to many problems, they are more difficult to develop because they require robust training of data sets and specialized expertise in optimization techniques. Operationalizing and productizing models for enterprise-scale usage also requires different but equally difficult-to-acquire technical expertise. In practice, using simpler AI approaches like older, non-deep- learning machine learning techniques can produce faster and better results than fancy neural nets can. Rather than building custom deep learning solutions, many enterprises opt for Machine Learning as a Service (MLaaS) solutions from Google, Amazon, IBM, Microsoft, or leading AI startups. Deep learning also suffers from technical drawbacks. Successful models typically require a large volume of reliably-labeled data, which enterprises often lack. They also require significant and specialized computing power in the form of graphical processing units (GPUs) or GPU alternatives such as Google’s tensor processing units (TPUs). After deployment, they also require constant training and updating to maintain performance. Critics of deep learning point out that human toddlers only need to see a few examples of an object to form a mental concept, whereas deep learning algorithms need to see thousands of examples to achieve reasonable accuracy. Even then, they can still make laughable errors. Deep learning algorithms do not form abstractions or perform reasoning and planning in the same way that we humans do. PROBABILISTIC PROGRAMMING Probabilistic programming enables us to create learning systems that make decisions in the face of uncertainty by making inferences from prior knowledge. According to Avi Pfeffer in his book, Practical Probabilistic Programming, a model is first created to capture knowledge of a target domain in quantitative, probabilistic terms. Once trained, the model is then applied to specific evidence to generate an answer to a more specific query in a process called inference. While the research and applications are in its early days, many experts see probabilistic programming as an alternative approach in areas where deep learning performs poorly, such as concept formulation using sparse or medium-sized data. Probabilistic programs have been used successfully in applications such as medical imaging, machine perception, financial predictions, and econometric and atmospheric forecasting. Probabilistic programming is emerging as a hot area in technical research, but it has yet to be productized and operationalized for enterprise performance to the same degree that machine learning and deep learning have. We won’t cover probabilistic programming in detail in this book, but you can check out the MIT Probabilistic Computing Project7 for recommended readings and tutorials.8 OTHER AI APPROACHES There are many other approaches to AI that can be used alone or in combination with machine learning and deep learning to improve performance. Ensemble methods, for example, combine different machine learning models or blend deep learning models with rule-based models. Most successful applications of machine learning to enterprise problems utilize ensemble approaches to produce results superior to any single model. There are four broad categories of ensembling: bagging, boosting, stacking, and bucketing. Bagging entails training the same algorithm on different subsets of the data and includes popular algorithms like random forest. Boosting involves training a sequence of models, where each model prioritizes learning from the examples that the previous model failed on. In stacking, you pool the output of many models. In bucketing, you train multiple models for a given problem and dynamically choose the best one for each specific input. Other techniques, such as evolutionary and genetic algorithms, are used in practice for generative design and in combination with neural networks to improve learning. Approaches like Whole Brain Uploading (WBE), also known as “mind uploading,” seek to replicate human-level intelligence in machines by fully digitizing human brains. Yet other approaches seek to innovate at the hardware level by leveraging optical computing,9 quantum computing,10 or human-machine interfaces to accelerate or augment current methods. The AI industry moves very quickly, and algorithms and approaches are constantly under development or being invented. To get an updated overview of modern AI technologies, download our latest guide on our book website at appliedaibook.com/resources. 2. THE MACHINE INTELLIGENCE CONTINUUM If you’re not an AI researcher or engineer, understanding the subtle differences and applications of various machine learning approaches can be challenging. Business problems can usually be solved in multiple ways by different algorithms, and determining the comparative merits of different methodologies can be frustrating without technical experience or practical experimentation. To help business executives comprehend the functional differences between different AI approaches, we designed the Machine Intelligence Continuum (MIC) to present the different types of machine intelligence based on the complexity of their capabilities. While we’ve defined the continuum to contain seven levels, keep in mind that the distinction between levels is not a hard line and that many overlaps exist. Systems That Act The lowest level of the Machine Intelligence Continuum (MIC) contains Systems That Act, which we define as rule-based automata. These are systems that function according to some predefined script, often by following manually programmed if-then type of rules. Examples include the fire alarm in your house and the cruise control in your car. A fire alarm contains a sensor that detects smoke levels. When smoke levels reach a predefined level, the device will play an alarm sound until it is turned off manually. Similarly, the cruise control in your car uses a powered mechanism to control the throttle position in order to maintain a constant speed. You would never set your cruise control, take your hands off the wheel, and claim that you now have a self-driving car. Doing so would result in terrible outcomes. Yet most companies claiming to have AI are really just using Systems That Act, or rule-based mechanisms that are incapable of dynamic actions or decisions. Systems That Predict Systems That Predict are systems that are capable of analyzing data and using it to produce probabilistic predictions. Note that a “prediction” is a mapping of known information to unknown information and does not necessarily need to be a future event. Andrew Pole, a statistician for Target, explained to The New York Times that he was able to identify 25 products, including unscented lotion and calcium supplements, that can be used to predict the likelihood of a shopper being pregnant and even the stage of her pregnancy.11 Target uses this information to serve eerily well-timed advertisements and coupons that encourage such customers to spend more money at the store. Statistics power most Systems That Predict, but predictions are only as good as the data being used. If your data is flawed, or you choose sample data that does not sufficiently represent your target population, then you will get erroneous results. In business analysis, lack of data integrity and methodological mistakes are extremely common and often lead executives to the wrong conclusions. Systems That Learn While Systems That Learn also make predictions like statistical systems do, they require less hand-engineering and can learn to perform tasks without being explicitly programmed to do so. Machine learning and deep learning drive most of these systems, and they can function at human or better-than- human levels for many computational problems. Learning can be automated at different levels of abstraction and for different components of a task. Completing a task requires first acquiring data that can be used to generate a prediction about the world. This prediction is combined with higher-level judgment to execute an action. The outcome from that action provides measurable feedback that can be reused at earlier decision points to improve task performance. Many enterprise applications of statistics and machine learning focus on improving the prediction process. In sales, for example, machine learning approaches to lead scoring can perform better than rule-based or statistical methods. Once the machine has produced a prediction on the quality of a lead, the salesperson then applies human judgment to decide how to follow up. More complex systems, such as self-driving cars and industrial robotics, handle everything from gathering the initial data to executing the action resulting from its analysis. For example, an autonomous vehicle must turn video and sensor feeds into accurate predictions of the surrounding world and adjust its driving accordingly. Systems That Create We humans like to think we’re the only beings capable of creativity, but computers have been used for generative design and art for decades. Recent breakthroughs in neural network models have inspired a resurgence of computational creativity, with computers now capable of producing original writing, imagery, music, industrial designs, and even AI software!12 Image from “Generating Stories from Images” by Samim Winiger, reprinted with permission Engineer and creative storyteller Samim trained a neural network on 14 million lines of passages from romance novels and asked the model to generate original stories based on new images.13 Flow Machines, a division of Sony, used an AI system trained on Beatles songs to generate their own hit, “Daddy’s Car,” which eerily resembles the musical style of the hit British rock group. They did the same with Bach music and were able to fool human evaluators, who had trouble differentiating between real Bach compositions and AI-generated imitations. Autodesk, the leading producer of computer-aided design (CAD) software for industrial design, released Dreamcatcher, a program that generates thousands of possible design permutations based on initial constraints set by engineers. Dreamcatcher has produced bizarre yet highly effective designs that challenge traditional manufacturing assumptions and exceed what human designers can manually ideate. Image from Autodesk Dreamcatcher, reprinted with permission AI is even outperforming some artists economically! In 2016, Google hosted an exhibition of AI-generated art that collectively sold for $97,605.14 Systems That Relate Daniel Goleman, a psychologist and author of the book Emotional Intelligence, believes that our emotional intelligence quotient (EQ) is more important than our intelligence quotient (IQ) in determining our success and happiness.15 As human employees increasingly collaborate with AI tools at work and digital assistants like Apple’s Siri and Amazon Echo’s Alexa permeate our personal lives, machines will also need emotional intelligence to succeed in our society. Sentiment analysis, also known as opinion mining or emotion AI, extracts and quantifies emotional states from our text, voice, facial expressions, and body language.16 Knowing a user’s affective state enables computers to respond empathetically and dynamically, as our friends do. The applications to digital assistants are obvious, and companies like Amazon are already prioritizing emotional recognition for voice products like the Echo.17 Emotional awareness can also improve interpersonal business functions such as sales, marketing, and communications. Rana el Kaliouby, founder of Affectiva, a leading emotion AI company, helps advertisers improve the effectiveness of brand content by assessing and adapting to consumer reactions. Mental and behavioral health is also an area ripe for innovation. Affectiva originated from academic research at MIT that was designed to help autistic patients improve recognition of social and emotional cues.18 Systems That Master A human toddler only needs to see a single tiger before developing a mental construct that can recognize other tigers. If humans needed to see thousands of tigers before learning to run away, our species would have died out from predation long ago. By contrast, a deep learning algorithm must process thousands of tiger images before it can recognize them in images and video. Even then, neural networks trained on tiger photos do not reliably recognize abstractions or representations of tigers, such as cartoons or costumes. Because we are Systems That Master, humans have no trouble with this. A System That Masters is an intelligent agent capable of constructing abstract concepts and strategic plans from sparse data. By creating modular, conceptual representations of the world around us, we are able to transfer knowledge from one domain to another, a key feature of general intelligence. As we discussed earlier, no modern AI system is an AGI, or artificial general intelligence. While humans are Systems That Master, current AI programs are not. Systems That Evolve This final category refers to systems that exhibit superhuman intelligence and capabilities, such as the ability to dynamically change their own design and architecture to adapt to changing conditions in their environment. As humans, we’re limited in our intelligence by our biological brains, also known as “wetware.” Instead of re-architecting our own biological infrastructure during our lifetime, we evolve through genetic mutations across generations. We cannot simply insert new RAM to augment our memory capacity or install a new processor if we wish to think faster. While we continue to search for other intelligent life, we are not yet aware of any Systems That Evolve. Computers are currently constrained by both hardware and software availability, while humans and other biological organisms are constrained by wetware limitations. Some futurists hypothesize that we may be able to achieve superhuman intelligence by augmenting biological brains with synthesized technologies, but this research is currently more science fiction than science. Once an upgradable intelligent agent does emerge, we will reach what many experts call the technological “singularity,” when machine intelligence surpasses human intelligence.19 Self-evolving agents will be capable of ever-faster iterations of self-improvements, leading to the eventual emergence of superintelligence. To download a visual summary of the Machine Intelligence Continuum, visit the resources section of our book website at appliedaibook.com/resources. How we build today’s Systems That Learn, Systems That Create, and Systems That Relate will affect how we build tomorrow’s Systems That Master and Systems That Evolve. While no one can predict what superintelligence will look like, we can take measures today to increase the likelihood that the intelligent systems we design are effective, ethical, and elevate human goals and values. The next few chapters tell the story of how modern AI can be used for the good of humanity, the immediate challenges that may cause AI to go awry, and the collaborative design principles we can uphold to build the best AI systems. 3. THE PROMISES OF ARTIFICIAL INTELLIGENCE The promises of AI extend beyond the challenges of Silicon Valley and Wall Street. Emerging technologies like deep learning and conversational interfaces enable us to do far more than drive advertising clicks, streamline sales, and boost corporate profits. All around the world, entrepreneurs and executives leverage data combined with machine learning to fight social injustice and crime, address health and humanitarian crises, solve pressing community problems, and dramatically improve the quality of life for everyone. Micro nance When Sahil Singla joined FarmGuide, a social impact startup, he was shocked to discover that thousands of rural farmers in India commit suicide every year.20 When harvests fail, desperate farmers are forced to borrow from microfinance loan sharks at crippling rates. Unable to pay back these predatory loans, victims kill themselves—often by grisly methods like swallowing pesticides—to escape reprisal from their debt holders. Singla and his team are tackling this issue with deep learning. Recent growth of computational power and structured datasets has allowed deep learning algorithms to achieve better-than-human-level accuracy in a number of recognition and classification tasks. Computers can now recognize objects in images and video, transcribe speech to text, and translate languages nearly as well as humans can. Using deep learning, FarmGuide analyzes satellite imagery to predict crop yields for individual farms. In the US, Stanford University researchers have shown machine-driven methods for crop yield analysis to be comparably accurate to physical surveys conducted by the USDA.21 Armed with this previously unattainable information, Singla and his team can build better actuarial models for lending and insurance, thereby reducing the risk of loan sharks preying on at-risk farmers by providing them with lower and fairer interest rates for loans. Social Justice In Monrovia, the capital city of Liberia, fifteen-year-old Sarafina was being hounded by one of her teachers, who refused to give her a report card unless she had sex with him. Embarrassed, she kept the issue hidden from everyone, even her parents, until her father overheard a harassing phone call that the teacher made to their home. He confronted the teacher and successfully secured Sarafina’s report card, but his daughter was reprimanded and forced to move to another school.22 Sarafina’s experience is not unique. In Liberia, teachers enjoy high social status while children, especially young girls, are culturally trained not to speak out. While Sarafina’s story sounds extreme to Westerners, her experience is painfully common and largely ignored in many developing countries. Enter UNICEF’s U-Report, a social reporting bot that enables young people in developing countries to report social injustice in their communities via SMS and other messaging platforms. “U-Report is not just about getting questions answered, but getting answers back out,” explains Chris Fabian, Co-Lead of UNICEF’s Innovation Unit. “We get responses in real-time to use the data for policy change.”23 By using a natural language interface to capture insights and performing statistical analysis on the aggregated results, the team leverages their more than 4.2 million users worldwide to identify and tackle challenging social issues like violence against children, public health policy, and climate change. U-Report polled 13,000 users in Liberia to ask whether teachers at their schools were exchanging grades for sex. An astonishing 86 percent of reporters said yes.24 Within a week of the U-Report on the “Sex 4 Grades” epidemic, hotlines around the country were inundated with reports of child abuse. Simply exposing a pervasive taboo inspired many more victims to speak up and reach out for help. The outpouring of responses provoked a government response and led UNICEF and Liberia’s Minister of Education to collaborate on a plan to stop the abuse of authority. In many parts of the world, citizens can’t utilize the feature-rich but data- intensive mobile apps that many of us enjoy due to bandwidth limitations and limited access to phones with up-to-date features. Being limited to voice calls and SMS means that technologies like natural language processing (NLP), dialog systems, and conversational bots become critically important to delivering value. Medical Diagnosis AI can dramatically streamline and improve medical care and our overall health and wellbeing. The fields of pathology and radiology, both of which rely largely on trained human eyes to spot anomalies, are being revolutionized by advancements in computer vision. Pathology is especially subjective, with studies showing that two pathologists assessing the same slide of biopsied tissue will only agree about 60 percent of the time.25 Researchers at Houston Methodist Research Institute in Texas announced an AI system for diagnosing breast cancer that utilizes computer vision techniques optimized for medical image recognition,26 which interpreted patient records with a 99 percent accuracy rate.27 In radiology, 12.1 million mammograms are performed annually in the United States, but half yield false positive results, which means that one in two healthy women may be wrongly diagnosed with cancer. In these situations, the patients often undergo biopsies, an invasive procedure that removes tissue or fluid from a suspicious area for analysis. To reduce the number of unnecessary surgical interventions, researchers at MIT and Harvard Medical School have developed a diagnostic tool that uses machine learning to correctly identify 97 percent of malignant tumors. Since deployment, the technology has reduced the number of benign surgeries by 30 percent.28 Artificial intelligence technologies are already saving lives and transforming societies. If used wisely, AI can be used to tackle many of the world’s greatest challenges. Used unwisely, however, AI can unintentionally amplify many of humanity’s worst traits. We highlight the challenges that undermine benevolent AI in the next chapter. 4. THE CHALLENGES OF ARTIFICIAL INTELLIGENCE “The future is already here—it’s just not evenly distributed.” —William Gibson When Timnit Gebru attended a prestigious AI research conference in 2016, she counted six black people in the audience out of an estimated 8,500 attendees. There was only one black woman: herself. As a PhD from Stanford University who has published a number of notable papers in the field of artificial intelligence, Gebru finds the lack of diversity in the industry to be extremely alarming.29 Data and technology are human inventions, ideally designed to reflect and advance human values. As our creations grow exponentially more powerful and their footprint ever larger on our society, we need to be increasingly mindful of the need to build them to be robust against adverse and unintended consequences. We cannot blindly trust the output of automated systems without vetting the accuracy of both the input data and the decision-making process itself. Many machine learning algorithms already influence our daily decisions and actions, but bad data and methodological mistakes can easily lead to erroneous results. In California, a flight-risk algorithm in use by the San Francisco Superior Court mistakenly recommended a man for release before trial. Despite multiple previous probation violations and arrests for gun possession, the algorithm judged him to be a minimal flight risk because someone had mis-entered the number of days that he had already spent in jail. Five days after release, he and a partner shot a local photographer.30 More subtle and insidious is the danger that algorithms designed by an undiversified team of elites may overlook the needs and values of underrepresented groups and unintentionally amplify the discrimination against them. Amazon customers, for example, discovered that same-day delivery was unavailable in zip codes that contained predominantly black neighborhoods, while computer scientists at Carnegie Mellon found that women were less likely than men to be shown ads for high-paying jobs.31 Even if characteristics such as race, religion, gender, or ethnicity are eliminated from models, other features that are highly correlated with those characteristics may be included and introduce the same bias. The biases of technology creators trickle down to their creations. While AI researchers pride themselves on being rational and data-driven, they can be blind to issues such as racial or gender bias or ethical issues that aren’t easily captured by numbers. With AI now used in high-stakes systems to identify terrorists, predict criminal recidivism, and triage medical cases, homogenous thinking in the technology industry has dangerous implications. The E ects of Discrimination To Latanya Sweeney, the first black woman to receive a PhD in computer science from MIT, the shortcomings of AI come as no surprise. Currently a professor at Harvard and the director of their Data Privacy Lab, Sweeney’s research examines technological solutions to societal, political, and governance challenges. One of her important contributions illuminates discrimination in online advertising, where she discovered that online searches of names that are more associated with the black community are 25 percent more likely to be targeted by ads that implies the person being searched for has a criminal record.32 Sweeney also uncovered SAT test prep services that charge zip codes with high proportions of Asian residents nearly double the average rate, regardless of their actual income.33 While price discrimination based on race, religion, nationality, or gender is illegal in the United States, enforcement of existing law is challenging in e- commerce, where the evidence of differential pricing is obscured by opaque algorithms. In healthcare, AI systems are at risk of producing unreliable insights even when algorithms are perfectly implemented, because the availability of medical data is affected by social inequality. Poorer communities lack access to digital healthcare, which leaves a gaping hole in the medical information that is fed into AI algorithms. Randomized control trials often exclude groups such as pregnant women, the elderly, or those suffering from other medical complications.34 Such exclusions mean that the unique physical characteristics of these patients are not incorporated into studies, which in turn affects whether tested treatment will be effective on patients who don’t share the characteristics of the original clinical volunteers. In the worst-case scenario, the treatment may actively harm the patient. Advocacy for algorithmic fairness cannot solely be the responsibility of the disenfranchised. Lasting, fundamental changes can only happen when technology creators and the public at large awaken to the dangers of exclusion and make inclusion a true priority. Malicious AI We don’t have to wait for AI to gain sentience and go rogue, because the probability of bad people taking advantage of intelligent automation for evil purposes is 100 percent. As machine intelligence becomes more powerful, pervasive, and connected, embedding AI in all of our personal and industrial computing devices increases the risk of attacks that can compromise the security infrastructures that protect our resources and communities. Luminaries from the Future of Humanity Institute, OpenAI, Centre for the Study of Existential Risk, and leading universities in the US and UK issued a 100-page policy recommendation paper, “The Malicious Use of Artificial Intelligence,”35 in which they described the fast-evolving threat landscape, identified key areas of security risk, and made high-level recommendations for preventative action that should be taken immediately. The report was alarming, pointing out that existing threats will get worse while new threats of an unknown nature will almost certainly emerge. AI will be used to multiply the effects of a malicious campaign—augmenting “human labor, intelligence, and expertise” to make the process of attacking easier and faster—and to broaden the types and number of possible targets. Advances in neural network algorithms that can produce hyperrealistic audiovisual input may be hijacked to produce fake news that looks like it came from a credible source, or to circumvent security systems that use voiceprints or other identifying features. In addition, AI may fundamentally alter the arena of cyber attacks by increasing the efficacy, precision, and untraceability of such attacks. These attacks may even target and hijack supposedly secure AI systems by exploiting their vulnerabilities. One grim possibility is the deployment of autonomous weapons systems, such as a drone, using facial recognition technology to identify and attack individuals in a crowd. Through wearables, standard computing devices, and the burgeoning Internet of Things (IoT), AI will inevitably permeate every corner of our existence. This means that our physical security, digital security, and even political security will be at risk of attack. While we spend much of our productive hours tethered to digital devices and roaming cyberspace, we still inhabit physical bodies and live in a material world. Nefarious AI can infect autonomous vehicles, connected appliances, and other devices to inflict bodily harm and property damage. Digital attacks may come as a coordinated and adversarial disruption of corporate data with the goal of compromising, devaluing, or altogether destroying an organization’s data architecture. Finally, the use of technology—including AI, predictive analytics, automation, and social media bots—can have far-ranging social impact. AI can be used for illegal surveillance, propaganda, deception, and social manipulation. 5. DESIGNING SAFE AND ETHICAL AI Ethics and Governance AI systems can’t simply be programmed to complete their core tasks. They must be designed to do so without unintentionally harming human society. As AI systems become more complex, the likelihood of facing ethical dilemmas also grows. Designing safe and ethical AI is a monumental challenge and a critical one to tackle now. To be effective, we must develop more sophisticated and nuanced policies that go far deeper and wider than simplistic, science fiction solutions like Asimov’s Three Laws of Robotics.36 In a joint study, Google DeepMind and the Future of Humanity Institute explored fail-safe mechanisms for shutting down rogue AI.37 In practical terms, these “big red buttons” will be signals that trick the machine to make an internal decision to stop, without registering the input as a shutdown signal by an external human operator. IEEE, the world’s largest association of technical professionals, published Ethically Aligned Design, a set of standards for the ethical design of artificial intelligence and autonomous systems.38 The publication lays out the chain of accountability for design and operation. It also emphasizes that to limit the possible extent of risks, such systems should not infringe on human rights, and their operations should be transparent to a wide range of stakeholders. Hypothetical fail-safe mechanisms and hopeful manifestos are important, but they are insufficient for addressing the myriad of ways in which AI systems can go awry. Homogeneous development teams, insular thinking, and lack of perspective lie at the root of many of the challenges already manifesting in AI development today. Luckily, as AI education and tools become more accessible, product designers and other domain experts are increasingly empowered to contribute to a field that was previously reserved for academics and a niche community of experts. Education as Remedy Tackling these challenges requires democratizing access to quality AI education and empowering collaborations between practitioners and multidisciplinary experts in order to gather missing data and build inclusive technology. Acquiring the requisite knowledge and resources to apply AI is a huge challenge for those who don’t live in Silicon Valley or other major research hubs. Many turn to massive open online courses (MOOCs) provided by companies such as Coursera, Udacity, and fast.ai as their only options. Rachel Thomas, a deep learning researcher with a doctorate in math from Duke University, started fast.ai with Jeremy Howard, the former president of Kaggle, to advance the mission of making deep learning accessible to all. As passionate champions of diversity and inclusion, the two have taught over 50,000 students globally, including Sahil Singla of FarmGuide. Fast.ai’s non-stop efforts to democratize AI education are paying off. Students of its MOOC are using techniques taught in the class to treat Parkinson’s disease, give visually impaired patients more independence, fight online hate speech, and end illegal logging and harmful human activity in endangered rainforests. The work is not done, however. Even with MOOCs, students in developing countries face an uphill battle compared to their counterparts in developed countries. Some struggle with the lack of structured datasets available in their language or culture, others with the lack of reliable internet infrastructure and access. Still others face a lack of career opportunities. Finally, the lack of affordable access to computational resources, such as graphic processing units (GPU) and reliable power sources, presents a major obstacle for students who want to build their own models. Even with the right hardware, complex neural network models can take days, if not weeks, to train. Even if computational resources were widely available, engineering education alone is insufficient to ensure that AI technologies are built safely and successfully. “Ethics training should be a mandatory part of engineering and computer science education,” emphasizes Rana el Kaliouby, founder of Affectiva, a company that makes machines more emotionally intelligent.39 El Kaliouby and her team regularly engage the public in open dialogue to uncover potential blind spots regarding transparency, privacy, security, and ethical concerns. Improving access to tools and education will bring in new expertise and viewpoints that can help evolve a field traditionally driven by an elite few. With AI’s exponential impact on all aspects of our lives, this collaboration will be essential to developing technology that works for everyone, every day. Collaborative Design As you embark on building your own AI technologies for your business or community, the following three principles of collaborative design will help you and your team approach AI development more holistically and successfully. Bringing in diverse expertise and thinking is critical to ensuring your technology is benevolent to all members of society and does not unconsciously reflect the biases of an elite minority. BUILD USER-FRIENDLY PRODUCTS TO COLLECT BETTER DATA FOR AI Data is a human construct, as are the tools that we design to gather it. Consumer-facing digital data is largely captured through the myriad of touch points that we have with our internet-connected devices and the complex ecosystem of apps, content, and networks that we access through them. If the products collecting requisite data to power AI systems do not encourage the right types of engagement, then the data generated from user interactions tend to be incomplete, incorrect, or compromised. In designing a product, you are building a specific journey for your customers to experience, and you will invariably influence user behavior and the resulting data trail. Manipulative products like clickbait headlines and aggressive calls-to-action (CTAs) optimize for short-term gains in lieu of long-term relationships, and the data they collect may not serve your ultimate business goals. Even if you are intentional in both your data collection and your product’s user experience (UX) design, remember that just because a user engaged with a button or clicked on an ad doesn’t mean you know their motives or intentions. The absence of experiential knowledge means that you cannot solely rely on data and algorithms to tell you which problems need solving. Machine learning and AI are not always the right solutions to a problem. Identifying the right problem and its solution requires tight integration and adaptation between your products and your users as well as a collaborative relationship between your team and your users. The UNICEF U-Report bot is a great example of this principle in action. Its key innovation was in designing the product to work over a single phone line for users who lacked smartphones and computers, not in its application of novel AI methodologies. PRIORITIZE DOMAIN EXPERTISE AND BUSINESS VALUE OVER ALGORITHMS When working with Fortune 500 companies looking to reinvent their workflows with automation and AI, we often hear this complaint about promising AI startups: “These guys seem really smart, and their product has a lot of bells and whistles. But they don’t understand my business.” In most cases, having and using a fantastic machine learning algorithm is less important than deploying a well-designed user experience (UX) for your products. Thoughtful UX design that delights users will drive up engagement, which in turn increases the interactions you can capture for future data and analysis. Thoughtful UX compensates for areas where AI capabilities may be lacking, such as in natural language processing (NLP) for open-domain conversation. In order to develop “thoughtful UX,” you’ll need both strong product development and engineering talent as well as partners who have domain expertise and business acumen. A common pattern observed in both academia and industry engineering teams is their propensity to optimize for tactical wins over strategic initiatives. While brilliant minds worry about achieving marginal improvements in competitive benchmarks, the nitty- gritty issues of productizing and operationalizing AI for real-world use cases are often ignored. Who cares if you can solve a problem with 99 percent accuracy if no one needs that problem solved? What’s the utility of a tool whose purpose is so arcane that no one is sure what problem it was trying to solve in the first place? EMPOWER HUMAN DESIGNERS WITH MACHINE INTELLIGENCE “Tools are not meant to make our lives easier,” says Patrick Hebron, author of Machine Learning For Designers, “[t]hey are meant to give us leverage so that we can push harder. Tools lift rocks. People build cathedrals.”40 Human designers can enhance their creations when they are supported by tools that use machine intelligence. The nascent field of AI design is one such area. While we are still figuring out which best practices should be preserved and which new ones need to be invented, many promising AI- driven creative tools are already in use. Hebron insists that machine learning can be used to simplify design tools without limiting creativity or removing control from human designers. Machine learning can transform how people interact with design tools through emergent feature sets, design through exploration, design by description, process organization, and conversational interfaces. Hebron believes that these approaches can streamline the design process and enable human designers to focus on the creative and imaginative side of the process instead of on technical mastery of the design software. This way, “designers will lead the tool, not the other way around.” How to Develop an Enterprise AI Strategy 6. BUILD AN AI-READY CULTURE You may have brilliant ideas for using artificial intelligence to improve your organization and community, but translating those ideas into viable software requires having the right mindset, dedicated leadership, and a diverse support team. In this chapter, we highlight many of the organizational and political issues that routinely block technical innovation and give you strategies for overcoming them. Be Honest About Your Readiness Despite many public claims to innovation, many corporations are still playing catch up on existing technologies such as big data, mobile, and the Internet of Things (IoT). Many brands have built up their social media presence and now offer mobile-friendly apps and websites, but these are merely digital consumer endpoints, not the basis for an enterprise-wide technological transformation. Other companies have accumulated big piles of data, but aren’t actively transforming their information assets into improved business practices. DO YOU HAVE A CENTRAL TECHNOLOGY INFRASTRUCTURE AND TEAM? A key milestone in the corporate digital transformation is the development of a centralized data and technology infrastructure. These two elements connect consumer applications, enterprise systems, and third-party partners and provide access to a single source of truth that contains relevant, up-to- date, and accurate information for all parties. Designing and implementing the infrastructure needed for enterprise-scale AI requires a strong and dedicated technology team that can develop internal application programming interfaces (APIs) to standardize access to both data and your company’s internal business technology. Doing so will enable your company to streamline enterprise-wide data analysis, accelerate product development, and respond more quickly in evolving markets. Internal APIs will also reduce the communication overhead needed to hunt down specific data, negotiate access, and interpret variations. You will also avoid duplicating software development work across different departments that have overlapping needs and goals. Non-technical companies typically see technology as a secondary priority and leave software projects to siloed business units. This leads to technical sprawl, which manifests when different business units implement their own initiatives without consulting each other, build conflicting or incompatible solutions, compromise security due to inconsistent standards and access, and overload IT departments that struggle to monitor and manage everything. If your company has not yet succeeded in managing technical sprawl (or if you have not yet begun to tackle the problem), we recommend that you tackle that problem before trying to launch a complex AI initiative. If unaddressed, technical sprawl will lead to your company investing in fits and starts and buying third-party AI products for narrow purposes, which will only exacerbate your existing problem. Building and maintaining a strategic, centralized, and secure architecture also requires strong executive commitment led by the C-Suite, plus ongoing operational collaboration from all departments and business units. DOES YOUR CORPORATE CULTURE VALUE DATA AND ANALYTICS? There is no point in laboriously gathering data and running sophisticated machine learning models if the analysis will be ignored. Many of the world’s largest enterprises have historically grown through gut decisions from influential executives, not from collaborative, data-driven decision- making. Due to past successes, some leaders prioritize their own beliefs and methods and are openly hostile to analytical approaches and centralized technology. Almost all of us have worked with colleagues with dogmatic qualities in our professional careers. They have a special name: HiPPO, which stands for “highest paid person’s opinion.” HiPPOs insist that their strategy is the right direction for the company, based largely on the fact that they came up with the idea. They often emerge with little warning to ram a new “vision” through the company or to shoot down initiatives that they perceive to be competing with their agendas. Executives who exhibit such behaviors rarely mean to be malicious and do not recognize themselves to be HiPPOs, preferring to style themselves as being “experienced” or “visionary.” Intuition-driven approaches may have worked in a bygone business era when no one had access to data or computing. However, now that software has eaten the world,41 fortune favors the nerds. In the 15 years between 2002 and 2017, the top five publicly-traded companies by market cap shifted from GE, Microsoft, Exxon, Citi, and Walmart to technology companies like Apple, Alphabet, Microsoft, Amazon, and Facebook. While data alone cannot make decisions for you, combining the right information with experience, creativity, and an unbiased perspective will enable executives to make better decisions. Nearly every company has a few executive HiPPOs. While you can probably manage a handful of naysayers, your company is unlikely to be competitive in AI if you’re up against a HiPPO army or an extremely powerful C-Suite HiPPO. We have seen data and analytics initiatives at major companies severely hindered or even cancelled by antagonistic executives. Choose the Right Champions Who should own AI initiatives at your company? One pattern stands out clearly: in every single tech firm that currently leads in AI, the CEO has come out strongly in favor of prioritizing AI company-wide. Microsoft CEO Satya Nadella describes AI as being “at the intersection of our ambitions. We want to democratize AI just like we brought information to your fingertips.”42 Sundar Pichai, CEO at Google, boldly stated that “we will move from mobile first to an AI first world.”43 Amazon’s Jeff Bezos calls our modern times the “golden age” of AI, stating that “we are now solving problems with machine learning and artificial intelligence that were in the realm of science fiction for the last several decades.”44 CEO, CTO, CIO, CDO, OR CAO? Finding the right stakeholder to champion a high-risk, high-reward technology initiative is half the battle. In a company that is traditionally conservative towards technology and digital investments, you may have a hard time convincing your CEO to champion AI initiatives. If that’s the case, try to find executive buy-in as high up as possible, ideally within the C-Suite or even at the board level. Successful enterprise AI applications can be led by many different executive roles, but whoever leads can’t simply rely on aspirational press releases. True leadership has to be demonstrated through vision, action, and budget. The executive should also possess high levels of technical sophistication, including the ability to understand—or the willingness to learn—the nuances and challenges of developing data, analytics, and machine learning products. “We’ve had a few people tell us that the biggest predictor of whether a company will successfully adopt machine intelligence is whether they have a C-Suite executive with an advanced math degree,” says Shivon Zilis, an experienced investor in AI and a partner at Bloomberg Beta. “These executives understand it isn’t magic—it is just (hard) math.”45 The ideal characteristics of an executive AI champion include: C-Suite executive level or higher Business and domain expert Credible and influential Technically knowledgeable Analytical and data-driven Controls sufficient budget Encourages experimentation Understands and accepts risks Collaborates well with decision-makers across multiple business units CEO In an ideal world, the CEO and the Board of Directors recognize the rising importance of AI and automation everywhere. As a result, they have empowered your executives with the decision-making capability, financial budget, and organizational resources to succeed. More importantly, they are technically savvy enough to understand the risks involved and are committed to driving progress. Leading technology CEOs have virtually all committed to the importance of AI for their businesses in public, but we’ve found that CEOs of non- technology companies can get caught up with existing strategy initiatives and lose traction on AI efforts. While you should always strive to have your CEO’s blessings, you may want to concentrate on finding a C-Suite champion who can dedicate substantial time to shepherding AI investments to fruition. CTO Chief Technology Officers create technology for an enterprise’s external business or individual customers. The CTO defines the technology architecture, runs engineering teams, and continuously improves the technology behind the company’s product offerings. Creativity, technical skill, and ability to innovate are essential to a CTO’s success. With technology products increasingly dependent on machine learning approaches to improve performance, a company that primarily produces software will need its CTO to prioritize investments in AI. For years now, Google, Facebook, Amazon, Microsoft, and other large technology corporations have prioritized integrating machine learning into their customer-facing products, as have leading companies in virtually every sector. Companies that don’t traditionally build end-user technology often don’t have CTOs, which can make transitions challenging if they want to create digital experiences for customers without relying on external agencies. CIO Chief Information Officers manage technology and infrastructure that underpin their company’s business operations. The CIO runs an organization’s IT and Operations to streamline and support business processes. Unlike the CTO, the CIO’s customers are internal users, functional departments, and business units. CIOs typically adapt and integrate third-party infrastructure solutions to meet their unique business needs and do less custom development than CTOs do. For non-technology companies, the CIO can be the right stakeholder if the primary benefits of adopting AI lie in improving analytics and business operations rather than in functions that affect external customers, such as in sales and marketing. However, in companies that view the CIO as “the IT guy” who has to report to another executive, the better stakeholder may be the higher-level executive who owns the final business decisions. Regardless of whether they are your primary AI champion, CIOs will likely play a vital role in implementing AI in an organization due to the need to develop and integrate infrastructure to support AI. ML systems and data mining systems require complex storage, networking, and computing systems that will require the CIO’s input to implement in many enterprises. CDO Since data touches all aspects of enterprises, Chief Data Officers (CDOs) are becoming increasingly common,46 but their mandate is more often the security, regulation, and governance of enterprise data. Depending on their focus, they typically report to CIOs, CFOs, Chief Risk Officers (CRO), or Chief Security Officers (CSO). Companies that have the CDO report directly to the CEO tend to value data and analytics more highly than those that don’t. Many enterprises started investing in centralized data infrastructure and capabilities less than five years ago, which means many new CDOs are still occupied with the monumental task of laying out their company-wide data initiatives. Consequently, they may not be able to focus on championing new AI investments. For some businesses, it may make more sense to appoint a Chief Data Supply Officer (CDSO) or comparable role to support the CDO. This person can direct a company’s data towards the end goal of adopting machine learning. They consider questions such as how best to manage competing sources of data, the cost of access and of data churn, where to store data, and how to simplify access to data.47 CAO Along with the CDO, the Chief Analytics Officer (CAO) is a relatively new role that has emerged to manage enterprise investment in big data and analytics. Companies that are early in the maturity cycle for big data may still be working to integrate, clean, organize, prepare, and transform data into an institutional asset. Once a CDO or comparable leader has organized high-quality data, a CAO can then apply meaningful analytics to solve business problems. The roles overlap, and the titles are often interchangeable. Many mature companies also combine the two roles so that a single executive is responsible for both the enterprise data management and the ensuing analytical functions. Corporations that see analytics as a critical asset will often have the CAO reporting directly to the CEO rather than to the CTO or CIO. Other Important Roles The roles that we highlighted tend to be executives with sufficient technical expertise, organizational resources, and enterprise clout to lead major AI initiatives. However, successful investments can be led by a myriad of roles including Chief Digital Officers, Chief Security Officers / Chief Information Security Officers, Chief Risk Officers, Chief Innovation Officers, Chief Science Officers, Chief Strategy Officers, etc. The exact scope and role of these positions in the C-Suite hierarchy can vary widely across organizations, so you’ll need to clarify their responsibilities within your own organization before pitching them to be your champion. AI initiatives can also be led by CMOs, COOs, CFOs, or other corporate leaders who own business decisions and maintain significant political influence within their organizations. DO YOU NEED A CHIEF AI OFFICER? “A hundred years ago electricity transformed countless industries; 20 years ago the internet did, too. Artificial intelligence is about to do the same,” writes Andrew Ng, former Stanford University Professor of Computer Science and a widely respected technical expert on machine learning.48 Just as CIOs became needed with the rise of the Internet and CDOs became needed with the increasing importance of data, Ng proposes that organizations establish a new role, the Chief AI Officer (CAIO), to govern and champion the role of AI in enterprises. Is a CAIO necessary? At the time of this writing, less than three dozen professionals on LinkedIn report having this title. Ng states the primary benefit of having CAIOs is that they can centralize a powerful AI team that can build and use AI technology to accelerate and streamline business functions across an organization, not just in siloes. “Let’s say your company has a gift card division,” Ng uses as an example. “Because AI talent is extremely scarce right now, it is unlikely that they will attract top talent to work on gift cards at the division level. A dedicated AI team has a higher chance of attracting AI talent and maintaining standards.” Unfortunately, these benefits only accrue if you hire the right person, which is a non-trivial task. Not only do successful CAIOs need to be masters of AI and data infrastructure technologies, they must be able to collaborate effectively with different departments and different roles, understand their priorities when formulating solutions to business problems, be charismatic enough to win support for new initiatives, and have enough industry clout to attract highly sought-after talent to their teams. AI is not a magical solution that instantaneously solves all challenges. Different, often simpler, approaches can also drive many improvements and advancements. Unless your technology initiatives are driven by clear business goals and viability, you run the risk of using AI aimlessly, like a hammer looking for nails. Deploying AI successfully also requires that your organization be “AI-ready,” i.e. have a strong culture of data-driven decision-making and technical experimentation. Otherwise, even the most brilliant CAIO in the world won’t do your company any good. GET BOARD LEVEL BUY-IN Many times, major investment projects will require board-level buy-in. A shift in company culture is often difficult when public companies are slaves to investors who expect quarterly results. Quarterly performance goals create pressure to drive toward short-term initiatives. For example, retail brands may recognize the need to completely evolve the way they sell to customers to compete with technology entrants like Amazon, but they will likely fumble when these strategic changes require major, longer-term R&D investments. In the face of exponential change ushered in by AI, companies need to prioritize longer-term investments for both growth and survival. These sweeping changes and resource allocations typically require board-level buy-in. While board members are not part of the executive team, they exert significant sway in shaping strategic corporate initiatives, inspiring or forcing the company to make technology investments. They can also provide support to public company leadership who are especially subject to the whims of quarterly earnings reports. Keeping your board educated and updated is essential if you aspire to larger projects. Build An Enterprise-Wide Case For AI Your case for investing in AI and in automation will depend on your champions and stakeholders since they possess different business priorities, performance metrics, technical aptitude, propensity for risk, and political relationships. Presenting a clear ROI on AI initiatives is the best way to persuade executive stakeholders, but this can be challenging when enterprise AI adoption is early and still being proven in many sectors. Many corporations are still completing their big data investments and have yet to broach analytics. We emphasized earlier the importance of being honest about whether or not your organization is ready for AI. Enterprises early in the maturity cycle for big data and analytics may need to wait until a basic data and analytics infrastructure is in place before chasing AI. If you’re unable to secure executive champions right away, you may still be able to pull off a limited pilot or prototype within your own department or team. Pilot costs vary widely depending on scope, scale, application, and timeline. A short, proof-of-concept marketing project for an e-commerce company that uses an off-the-shelf solution may cost tens of thousands of dollars and be implemented in days. A more fleshed-out project, such as the codification of regulatory contracts for an investment bank, may cost millions and require years. A repository of clean, accessible data will help drive down project costs and time. Read part three, “AI for Enterprise Functions,” to develop ideas for how AI can be beneficial for your business. In this section, we highlight popular applications for internal operations like finance, legal, and HR, as well as customer-facing functions like sales, marketing, and customer service. Why You Need a Multi-Disciplinary “AI SWAT Team” Executives alone cannot bring about organizational change, especially of the magnitude that AI can potentially make across an enterprise and industry. Some of your most important stakeholders are your front-line employees and middle managers who will be integrating, using, and overseeing AI tools every day. Many of your employees likely have a strong fear that AI and automation will take away their jobs. Unlike you, they may not initially understand how these powerful technologies can be used to eliminate their lower-value tasks, free them to perform creative and strategic work, and thus augment their output. We have seen firsthand how failure to educate, include, and adapt to these important voices within an organization can lead to resistance, political infighting, and internal sabotage. You will need to put together a multi-disciplinary, cross-departmental “AI SWAT Team” composed of stakeholders in different departments, different functional roles, and different hierarchical levels. The responsibility of this team will be to support your champion and to identify, prioritize, execute, and evangelize your highest-ROI opportunities for automation across the whole company. They will also be critical for identifying potential pitfalls in your organizational design, technical capabilities, and strategic and tactical plans. These diverse views will be critical to help you and your champions form a clearer and fuller picture of the true impact of machine intelligence in your company. Nearly every job function is now touched by technology that can benefit from machine learning, and new roles will emerge due to new computing capabilities. Therefore, while your “AI SWAT Team” will be comprised of a small group of programmers, planners, and sponsors, the team should be also surrounded by advisors from non-technical departments, such as HR or Finance, who can inform you of the areas in most urgent need of automation and the areas with the highest ROI. Coordination across departments and functions will also enable you to identify commonalities that can be addressed by machine learning and tackle them with centralized rather than siloed solutions. The size of your SWAT team may vary depending on the size of your organization but typically ranges from five to fifteen members. The membership may rotate as you mature as a digital organization and face different challenges in your evolution. Thus far, the discussion assumes that your organization is already willing to take a stab at investing in AI. But what if it’s resistant to the idea? How can you convince your organization that AI is a good idea in the first place? Get Organizational Buy-In Most AI implementations are cross-functional and require input from multiple departments. A retooling of your accounts payable system will need inputs from finance, legal, security, and technology. A project may also pull in human resources, if employees need to be reassigned, and operations, if processes require adjustments. To succeed, you will need support from other executives, their front-line managers, and their staff. Here are some things that you can do to win support from your organization. Focus on Revenue Potential A key strategy is to appeal to your business leaders about the potential of increasing the bottom line. AI can save time and energy, reduce costs, and increase profits, which then provide executives an opportunity to grow their business lines and advance their careers. Stay Ahead of the Competition The fear of missing out (FOMO) is also a very strong motivating force. If business unit leaders fail to take action, emphasize that they are setting themselves up to fall behind competitors who are jumping on new technology. Not investing in the organizational and technical requirements to adopt AI may mean falling so far behind that you’re unable to compete in the future. Start Small and Show Early Wins Pick a smaller, sure-win project to demonstrate possibilities. While returns may be limited, an early success will give you confidence when you request that the project scope be expanded. Aim for something with a short time horizon that can be completed with a small task force. For example, in customer care, Nuance recommends routing a tiny portion of customer support queries to an AI system at the onset. Initially, an automated support system can answer 20 to 30 percent of frequently asked questions, but accuracy can increase to over 80 percent and expand to more topics as the system learns over time.49 Don’t Call It Artificial Intelligence When pitching your project, emphasize the value that new technology can deliver instead of the technical details of implementation. Karl Bunch, former SVP of Xaxis, a subsidiary of WPP, built a prototype of an algorithmic adtech trading platform with a small team of engineers working in their spare time. To keep expectations low, he refused to describe the technology as “machine learning” or “artificial intelligence” until after the system started showing great promise after a few quarters. What started out as a skunkworks project has now become a core part of Xaxis’s ad tech trading platform. Allay Fears of Sudden Job Loss Given the negative media hype surrounding AI, your employees understandably have concerns over their job security. You can allay these fears and promote a healthy work environment in which both humans and machines cooperate and thrive. Research finds that while 45 percent of tasks are automatable, only five percent of overall jobs have been supplanted by automation.50 AI systems largely handle individual tasks, not whole jobs. High costs, legal regulations, and social resistance to AI all hinder the progress of technology adoption. With the rise of autonomous vehicles, many believe that the jobs of America’s 1.7 million truck drivers are in imminent danger. The reality is that trucking jobs will likely require many years to replace. Michael Chui, a McKinsey partner, told The New York Times that the replacement and retrofitting of America’s truck fleet with autonomous navigation will require a trillion-dollar investment that few, if any, companies will immediately undertake.51 Even if financing can be secured, autonomous vehicle technology is not yet approved for industrial or for individual use. If humans can outsource repetitive and mundane tasks to AI, then they can devote more attention to tasks requiring strategic skills such as judgment, communication, and creative thinking. Eliminating boring jobs that employees dislike can also improve morale and interest as they take on increasingly more meaningful work. Accenture’s Operations group, which has more than 100,000 employees, initially calculated that automation would replace 17,000 jobs in their accounts payable and marketing operations. However, headcount actually grew as employees moved to more strategic advisory services that expanded their business lines.52 The internet and mobile technology revolutions have created far more jobs than they have destroyed. AI will likely have the same effect, but new opportunities in the digital economy will require superior technical skills and knowledge. Demonstrating your commitment to retrain your employees for changing roles and responsibilities will go a long way toward gaining their consent and trust. Educate Your Stakeholders How would you react if you found out that your CEO doesn’t know how to use a mobile phone, insists on handwritten correspondence, and has never heard of the internet? You can’t expect your C-Suite to be experts in AI, but you can ensure that leading executives have a baseline education about machine intelligence. When we train executives in the intricacies of AI, our curriculum is divided into an introductory module on our Machine Intelligence Continuum (MIC) framework and AI applications within a specific industry, a training module on how to evaluate an organization for AI-readiness, and a hands-on project to design and implement a pilot. Keep in mind that your executives will not need to know about the finer technical differences between algorithms. Your overall goal should be to teach your executives to gain a practical appreciation of what AI can do for your company. You’ll want to include a theoretical introduction to help them separate hype from reality and also hands-on experience to help them understand both the limitations and potential of AI implementations in a corporate setting. To learn more about our executive education offerings, visit appliedaibook.com/education. 7. INVEST IN TECHNICAL TALENT Now that your company leaders and key stakeholders are both on board, the next challenge is to find people who have the necessary technical skills to staff your initiative. Finding the right people can be no less of a challenge than wrangling internal political support. Jean-François Gagné, founder of leading AI company Element.AI, calculated that there are fewer than 10,000 people in the world currently qualified to do state-of-the-art AI research and engineering.53 Most of them are gainfully employed and hard to poach. If you’re looking to recruit fresh graduates, the head of a prominent Silicon Valley AI lab recently confided to us that American universities only graduate about 100 competent researchers and engineers in this field each year! The high demand for specialized AI talent, coupled with the painfully low supply, means that companies need to adopt new strategies when recruiting for a new AI initiative. Wealthier firms can afford to throw money at the problem by acqui-hiring AI startups at one to five million dollars per engineer.54 Based on a study of public job listings among US employers, Forbes found that the top 20 AI recruiters, led by Amazon, Google, and Microsoft, spend more than $650 million annually to woo elusive researchers and engineers.55 What should you do if you don’t have the deep pockets to go head-to-head against the Googles and Amazons of the world? Whether you’re a new startup or an established enterprise looking to expand an AI project team, the following tips may help your company stand out in the noisy and competitive AI job market. Understand Di erent Job Titles Many companies struggle just to understand what “artificial intelligence” is, much less the myriad of titles, roles, skills, and technologies used to describe a prospective hire. Titles and descriptions vary from company to company, and terms are not well-standardized in the industry. However, most of the roles you encounter will resemble the following: Data Science Team Manager A data science team manager understands how best to deploy the expertise of his team in order to maximize their productivity on a project. This manager should have sufficient technical knowledge to understand what his team members are doing and how best to support them; at the same time, this manager must also have good communications skills in order to liaise with the leadership or non-technical units. Though a person originally hired for a more junior role might organically fill this leadership position, teams that do not have designated and experienced managers are generally less productive. Machine Learning (ML) Engineers As their title indicates, ML engineers build machine learning solutions to solve business and customer problems. These specialized engineers deploy models, manage infrastructure, and run operations related to machine learning projects. They are assisted by data scientists and data engineers to manage databases and build the data infrastructure necessary to support the products and services used by their customers. Data Scientists Data scientists typically work in an offline setting and do not deal directly with the production experience, which is what the end user would see. Data scientists collect data, spend most of their time cleaning it, and the rest of their time looking for patterns in the data and building predictive models. They often have degrees in statistics, data science, or a related discipline. Alternately, many have programming backgrounds and hold degrees in computer science, math, or physics. Researchers, Research Scientists Researchers are more focused on driving scientific discovery and less concerned with pursuing industrial applications of their findings. They often build on promising leads uncovered by data scientists and experiment with novel approaches, much of which originates from or is inspired by work done in academic or industry research facilities. Applied Research Scientists, Applied Research Engineers Applied researchers straddle research and engineering. Unlike pure researchers, they are more concerned with practical research, such as identifying and implementing workable solutions to a specific problem or formulating industrial applications for scientific discoveries. Data Engineers, Distributed Systems Engineers Given the vast amounts of data and computation power required, most ML models face scalability issues. A talented infrastructure engineer can resolve challenges associated with large datasets, allowing researchers and data scientists to focus on their models rather than on data issues. Though this role is not explicitly focused on machine learning, it’s a vital component of a complete ML team. The composition of a machine learning team will change in response to the nature and timing of the project. Projects in fundamental research require more data and research scientists, whereas projects closer to production will require more applied researchers and infrastructure engineers. Due to the limited talent pool, many enterprises may not be able to hire applicants to fill all of these roles in-house. In this case, these companies could potentially fill in the gaps with enterprise solutions like Machine Learning as a Service (MLaaS) or AutoML technologies. Seek the Right Characteristics The skills required for successful careers in machine learning are different from those in traditional software development. Software development often has clearly structured tasks with well-defined deadlines for delivery and release. Once a key feature is done, engineers typically disengage and move on to another development project. While bug fixes and operational maintenance are required after completion, successful software development projects start with relatively clear specifications and product design and are launched when they meet release requirements. By contrast, machine learning is highly exploratory and experimental, with less clear timelines and success metrics. Ideal performance targets may not be knowable in advance and may shift during a project. Algorithms require ongoing support, training, and feedback in order to perform optimally. Mathematical Aptitude A background in mathematics and statistics is far more valued in machine learning than in traditional software engineering. Training ML models requires a sufficient background to understand which algorithms to apply and how to interpret and improve upon the results. For cutting-edge AI research positions, advanced mathematical intuition is a prerequisite in order to design and develop novel methodologies. Curiosity Training ML algorithms requires a constant sense of curiosity. The model builder needs to take in abstract information and make sense of it through continuous experimentation. This person will need to enjoy constantly learning new information and taking on new challenges. Creativity As ML tools and methodologies are still relatively new, the ability to think through ideas and to come up with novel ways to tackle a problem is highly valued. There will inevitably be many challenges that require new perspectives and solutions. Perseverance Artificial intelligence research is an ever-evolving pursuit. There are few simple answers, and it may easily take months to successfully train a viable algorithm. A successful individual will continue to try new techniques in the face of repeated failures until a solution can be found. Rapid Learning AI is evolving rapidly and keeping up-to-date with the accomplishments in the field is critical. Successful candidates should be able to stay on top of the latest technical developments and to quickly and intelligently apply what they have learned. Passion for Your Problem “We get plenty of resumes from people with talented machine learning and data science backgrounds,” says Zhen Jiang, Lead Analytics Supervisor at Ford Motor Company. “What I am much more concerned about is whether they have a passion for cars and mobility.”56 Talented engineers and researchers can go to any company in any industry that they want. Focus on finding applicants who are particularly excited by the unique problems that you face and the datasets that you own. Check whether they have done past research or projects related to your industry, and seek out talent at topical events that attract both enthusiasts and a more focused audience. Knowing When to Stop Perfect is the enemy of the good. Look for pragmatic applicants who recognize that a “good enough” model that meets product deadlines is better than a model that sits in development awaiting “just a few more tweaks.” This quality can be hard to find, as most scientists and machine learning experts, especially those with PhDs, are often trained to seek perfection. Optimize Recruiting Strategies There is no one-size-fits-all hiring strategy for technical talent. When recruiting, be sure to tailor your approach to the level, background, and career goals of your prospects in order to maximize your chance of success. JUNIOR ENGINEERS Given the high interest in AI, machine learning, and data science, many universities of

Applied Artificial Intelligence Handbook for Business Leaders PDF

Document Details

Tags

Related

Summary

Full Transcript

Upgrade to continue