Chapter 2. How to use AI in your business
This chapter covers
- What project leaders must know about AI
- Finding which business problems benefit from the use of AI
- Matching AI capabilities with the business problems you’re solving
- Finding the gap between the skills the data science team has and the ones your AI project needs
You can spend years learning about AI, but because of the fast evolution of this field, even fully proficient data scientists need to spend a significant portion of their time on continuous and ongoing learning. The market of AI books and papers is dominated by technical information about AI. With all that wealth of knowledge, it’s difficult to distinguish between what you need to know to manage AI and the knowledge necessary to have if you’re an engineer building an AI system.
This chapter talks about aspects of AI and ML that are necessary to understand to lead an AI project. It also teaches you how to find business problems that benefit from the application of AI. It provides examples of how to make AI insights actionable by linking AI capabilities with the business actions you already know you can take.
I’ve chosen the examples given in this and subsequent chapters from different business domains. It’s possible that some of the examples will come from a business domain unfamiliar to you. This is a good opportunity to practice one of the primary skills in successfully applying AI: adapting AI capabilities to business situations that you encounter for the first time.
2.1. What do you need to know about AI?
AI projects are very complex, combining business, computer science, mathematics, statistics, and machine learning. This section explains why technical knowledge about AI isn’t the primary knowledge needed for managing AI projects. If you’re an AI project leader who doesn’t have an analytical background, it’s understandable if you feel that you need to grasp all of those concepts to be able to make the best decisions.
The situation could be even worse: not only are the data scientists talking about concepts with which you’re unfamiliar, but those concepts might look like something you’re supposed to know but can’t fully recall. The jargon they use is often rooted in (or related to) statistical terminology. You might have taken a statistics class or two during your MBA program, and you might not have paid particular attention to all the topics covered. Don’t worry. The most important decisions for project success don’t require or even necessarily benefit from an extensive knowledge of statistics or the details of AI algorithms.
What you do need to know to manage AI projects is the same as with any other project: how to define metrics and processes that allow you to properly comprehend and monitor the direction and success of the project. Once you understand that, managing AI projects is similar to running those projects you’ve overseen before.
To use an analogy from a well-understood domain, if you were managing a factory, you wouldn’t think that you’d need to become as good a worker as your foreman to run it. For that matter, it’s a safe bet that quite a few executives who successfully manage factories aren’t remotely handy.
The same principles apply for IT projects. Do you really need to know your database as well as your database administrator (DBA)? Do you feel you need to become a DBA to manage a database project?[a] You manage database projects by separating the business and architectural aspects of such projects from the skills needed to maintain RDBMS systems.
In case someone wants to argue that AI isn’t a factory, no, it isn’t, but then neither is a database project. We’ve learned how to manage database projects without executives needing to become DBAs. Management as a profession is based on some universal principles for running organizations and projects, and that body of knowledge applies to AI too.
Just like a factory manager benefits from knowing how a factory works, having technical knowledge about AI doesn’t hurt the project leader. However, the factory manager can’t focus on the details of their foreman’s job as a substitute for knowing how to manage a factory and actively managing it. Likewise, an AI project leader must focus on management considerations.
Still, there’s often a feeling that managing AI projects requires significant focus on the details of how AI works internally, while comparable focus on details isn’t necessary when managing factories or database projects. To the extent that this is true, it isn’t so much because AI is different from other fields, but because AI is simply a much younger field.[b] In the case of factories, we’ve had enough time to develop management theory to understand that management knowledge isn’t the same as domain-level knowledge about manufacturing. Having more time has allowed us to build methods and systems that allow us to compartmentalize skillsets: those needed to run a factory versus those needed to build a product. The goal of this book is to help you to do the same with AI.
Yes, finding good data scientists is difficult, and, currently, they’re rare. Today, some of them might object to being compared to a foreman. But do you think that a shift foreman for a railroad was a common skillset when early railroads were built? Or that the DBA skillset was common when databases were introduced? That’s what I mean when I say AI is a young profession.
Most AI concepts that are relevant to making executive decisions could be explained to businesspeople in business terms. Ideally, your data scientists should be able to do that. If they can’t, you should supplement your project team with people who have expertise in both AI and business to help with communications.
Note
If you’re feeling that you need to better comprehend analytics to make business decisions, what you have isn’t a knowledge deficiency but rather a communications problem.
What you need to know to manage an AI project is how to relate AI concepts to business. Namely, you need to be able to answer the following questions:
- What can AI do, and how can I use that in my business?
- What type of AI project should I start with first?
- How will I measure how successful AI is in helping my business?
- How should I manage an AI project?
- What resources are scarce, and how should I best assign them?
The rest of this book shows you how to organize a data science project in such a way that you can apply the management skills you already possess with minimal modifications to run AI projects.
2.2. How is AI used?
You make money when you perform an appropriate business action. That leads us to where AI plays into any system it’s using—AI directs you in which action to take. This section explains how AI does that.
While AI, ML, and data science are perceived as new, the role that they play in making businesses successful isn’t new. There are quite a few professions that have historically used some form of data analysis to make money. Some examples of those professions include actuaries and quantitative analysts. Experts in the application of statistical methods to process engineering and quality improvement science also have a long history of using data analytics to improve business results. AI doesn’t change the way analytics and business relate, it just changes the methods for doing the analysis and the capabilities (and cost) of the analysis. There are significant parallels between how AI fits in with business today (and in the future) and traditional uses of data analysis in business.
To understand how to identify an opportunity to apply AI to a business problem, you first need to understand how a successful application of AI to a business problem has to look at a high level. In any problem in which data is used to inform further actions, there’s a common pattern describing that process. We collect data, analyze it, and then react to it. This is simply an age-old control loop, and it’s important to understand the elements of this loop. AI just adds new capabilities in the analytical part of that loop. Figure 2.1 shows how these elements interact.
Figure 2.1. The Sense/Analyze/React loop. Any successful analytical project must have all three elements of this loop.
Elements of figure 2.1 are as follows:
- Sense— The sensor part of the loop is where you get the data that the analysis looks for. For the majority of enterprise systems in the pre-big data age, data was in disparate databases. For big data systems, it’s common to store data in a data lake.
- Analyze— That’s the box in which you now apply AI to your dataset. Before AI, we used simpler algorithms (for example, a PID controller [34]) or human intervention (for example, a manual loan approval in the context of a bank). Although the introduction of AI to help with analysis is perceived as a recent development, it isn’t—AI research started in 1956 [35]. We’ve been using computerized systems to perform analysis for decades. What’s new is that today, with modern AI techniques, computerized analysis is much more powerful!
- React— Reactor/effector is the part responsible for the action in the real world. That reaction might be performed by a human or by a machine. Examples of manual reaction include many decision-support scenarios in which a management decision is made based on the results of the analysis. Examples of automatic reaction include robotics systems, smart thermometers [36,37], and automated vehicles [38].
The speed of closing the loop is the time between the moment some event occurs and the time when a reaction is performed. How much the speed of closing the loop matters depends on the domain. In a high-frequency trading system, there may be a strong requirement for completing a loop with the utmost speed. In other situations (for example, if you’re performing data analysis in the context of archeological research), timing requirements may be much more relaxed. Sometimes the ability to guarantee that you’ll meet time-critical deadlines matters too; an autonomous vehicle may require that your AI analytics never take longer than a specified time.
Note
The speed of closing the loop also depends on how often data is ingested into the system. Sometimes it’s acceptable that data is periodically ingested into your system. In other cases, data must be analyzed in real time as it’s arriving into the system. (This is called streaming analytics.)
An important consideration in the application of a Sense/Analyze/React loop is the question of who or what is reacting. It could be the system itself in some automated fashion. (That’s what a self-driving car [38] does.) Or, based on the results of the analysis, it could be a human. The latter case is much more common today within an enterprise use of data science.
The Sense/Analyze/React loop is applicable across many scales. It could be applied on the level of a single device (as in the case of smart thermometers like Nest [36] and ecobee [37]), a business process, multiple departments, the whole enterprise, a smart city, or an even larger geographical area. I believe that the Sense/Analyze/React pattern loop will, in the future, be applied to the level of whole societies, in systems such as disaster relief and the tracking and prevention of epidemics.
The Sense/Analyze/React pattern isn’t limited to the domain of big data and data science. That pattern applies to the domain of development and organizational processes too. You might be aware of various forms of the control loops that management sciences define and use. Examples of those loops are concepts like PDCA [39,40], OODA [41,42], and CRISP-DM [43], which have commonalities with and are further elaborations of this pattern. The Sense/Analyze/React pattern even applies to biology (for example, how octopuses and other animals behave [44]). In some domains, people might call the React part of the loop the Effector instead [45].
Automation of any business process is just an application of the Sense/Analyze/React loop. Using AI allows the application of that loop to some problem domain in which an automated reaction wasn’t previously possible.
Even uses of fully automated and rapid Sense/Analyze/React loops using complicated and computerized analysis are nothing new. Capital markets, especially combined with algorithmic trading, implement this pattern on a large scale. With the further advancement of the Internet of Things [46] and robotics, these large-scale, fully automated, closed control loops will become much more prevalent within the physical world.
2.3. What’s new with AI?
The advance of AI broadened the applicability of the Sense/Analyze/React loop, because AI brought to the table new analytical capabilities. This section explains those new capabilities.
What’s new with AI and big data is that automated analysis has become cheaper, faster, better, and (using big data systems) capable of operating on much larger datasets. Analysis that used to require human involvement is now possible to do with computers in areas like image and speech recognition. Thanks to this new AI-powered capability, whole Sense/Analyze/React loops became viable in these contexts, when it wasn’t economical to apply them before.
The following are some examples where an introduction of AI makes it possible to automate tasks that previously required a human to perform them:
- Automated translations from one language to another— Language translation is nothing new and is something that humans have done since the beginning of time. What’s new is that
AI has reached a level at which automated translations are now viable and, as such, a translation web service becomes practical.[a]Note that the actual control loop in a real translation system typically requires that at least two control loops are present. One loop translates from one language to another, while another loop makes money for this service. That second loop may also work by collecting information about what translations you need, analyzing them, and performing some action that makes money for the provider of the translation service.
- Autonomous cars— We’ve had some form of the automobile for the last 250ish years, always requiring a human operator.[b] What’s new with AI is that we may be on the verge of being able to construct a car that doesn’t need a human driver.
The first self-propelled vehicle with wheels was invented in 1769 [47], with the first gasoline-powered cars appearing in 1870 [48].
- Ability to diagnose eye diseases— We’ve all read letters at a distance for ophthalmologists and optometrists and stared into the bright light on command. What’s new is the ability of AI to detect diabetic retinopathy from simple retinal images [49].
- Ability to read comments posted on the web— If you read enough material in the comments section of a website, you can surmise whether people are enthusiastic or skeptical
with regards to some topic. Now AI can do it too. AI can read a much larger number of comments faster and cheaper than a human
ever could, and then tell you whether the audience is predominantly enthusiastic or skeptical. We call this capability sentiment analysis.[c]At the time of this writing, AI isn’t nearly as good a reader of web content as a human is, and it’s struggling with cynicism and subtle messages in text—it often misses even the basic thesis of the message. However, for the purpose of answering the question, “Has sentiment about the product improved in the last three months?” AI is good enough and can provide an answer much more cheaply than you or I can.
- Product recommendations— Each one of us has friends that recommend books, movies, and products we might like. When AI does that (for example, on the Amazon website), it’s called a recommendation engine. Historically, when the size of datasets was small, humans were able to perform the same analysis that was done by AI. In some cases, what AI does is worse than what humans can do looking at the same dataset. But AI is more economical in the long run, and it can operate on datasets that are too big for humans to look at.
What’s not new or different with AI is that analysis still can’t make money all by itself. Note that in none of the use cases given in the previous examples did I talk about how to make money. While some of those use cases are clearly straightforward to monetize (for example, an autonomous vehicle that drives better than humans), in others it may not be clear how to monetize AI.
Sometimes you won’t be able to make a profit regardless of how good your AI-powered analysis is. Suppose that an ill-advised manufacturer of traffic signs decides to do sentiment analysis of the public opinion of those signs. The manufacturer would likely lose money on this analysis. It’s not clear that drivers’ feelings have a significant practical influence on the selection of traffic sign vendors (or for that matter, that sentiment about the sign would be determined by the choice of vendor as opposed to where the sign is placed).
When you perform an analysis, you incur the cost of that analysis. Profit may happen when you react based on the results of analysis. If there’s no business action you can take after getting the results of analysis, such an analysis is always a loss.
2.4. Making money with AI
If AI allows for improved analytics with the Sense/Analyze/React loop, how do you make money with AI? By finding a situation in which AI allows you to apply Sense/Analyze/React loops so that one of the business actions available to you could be automated using that loop. This section shows you how. Figure 2.2 presents the general process of making money with AI.
Figure 2.2. Making money with AI is based on finding a business problem in which you can apply the Sense/Analyze/React loop to one of the actions you can take.
You can apply this control loop in a new context because of the capabilities of AI. But to successfully apply the Sense/Analyze/React loop, you must make sure that all components of the loop are technically possible:
- On the Sense side, you must have the ability to collect the data that AI-supported analysis will need. Chapter 3 addresses how to ensure you’ve collected the appropriate data for your chosen AI method.
- On the Analysis side, you must make sure that you stay within the boundaries of what’s possible with the available AI technology.
- On the React side, you must link the results of the analysis with one of the actions that you can actually implement in your business. You’ll make a list of the possible business actions that you can take, and ask, “Is there an AI analysis that I can perform that will better inform this business action?”
Once you know that the Sense/Analyze/React loop is applicable to your business problem, you know that you have the ability to solve that business problem using AI. Let’s start with an example.
2.4.1. AI applied to medical diagnosis
Suppose you’re part of a software development team in a large hospital. Your team’s goal is to apply AI to clinical and diagnostic procedures in the hospital. This section shows you how to find a use case in which AI can help.
To keep this example small and manageable, I’ll concentrate on a single diagnostic workflow: a patient getting an eye exam. An image of a patient’s retina is taken to check if there are any diseases. Assume that this procedure consists of the steps shown in figure 2.3.[1]
An actual optometry/ophthalmology exam is more complicated and is simplified here for the sake of illustration.
Figure 2.3. A workflow of a routine optometry exam. We’ll apply AI to automate part of this workflow.
The workflow shown in figure 2.3 consists of the following steps:
- Patient is briefed about the procedure. This step could be performed by a technician with minimal involvement of the optometrist.
- Patient is positioned in the imaging device, and an image is taken. This step is also performed by the technician.
- Image is developed in a workflow that’s already automated.
- Image is read by the optometrist, looking for abnormal conditions. If necessary, additional doctors would be consulted.
Now you’d find a place where AI can help. In this workflow, you have three steps to which you could apply AI: two interactions with the patient and the final reading of the eye image.
To illustrate why you don’t start by asking, “What analysis can I perform?” let’s construct a scenario in which you start with an analysis based on knowing that AI can do something for you.
You might be aware of voice assistants like Apple Siri [50] and know that their voice recognition is getting better. What if you combine a voice assistant/voice recognition with the chatbot so the patient can be briefed by a machine? You’re lucky to have a good data science team that’s happy to work with this cool technology. This looks like a good application of AI, doesn’t it? Let’s build a quick prototype!
Unfortunately, any time you spend on such a prototype will be wasted. Replacing a technician has limited value—the cost of the time a technician spends on briefing the patient is relatively small. More importantly, you’re serving a diverse patient population, including multiple languages, disabilities, ages, and comfort with interacting with the machine. Human technicians are good at dealing with this population; today’s AI isn’t, if for no other reason than some segments of the population aren’t used to talking with a voice assistant.
The idea you had was based on an interesting technology. The use case is inherently interesting and straight from sci-fi—many sci-fi stories feature patients talking with an AI doctor. The problem is that you’ve tried to apply AI to a situation in which a business action isn’t profitable from the start, due to factors beyond your control.
This is a common trap. Everyone who works with AI in a business setting claims that they have a good business case, but often the business case was an afterthought, and the team’s initial excitement about the project was caused by the opportunity to work with interesting AI technology.
In the worst case scenario, it might be impossible to monetize the project from the start, even if its technical portion was successful. Good AI implementation can’t bail out a poor business case.
Let’s use a systematic approach to better apply AI to this optometric exam. You start by enumerating the domain actions that you can perform and then see if you can apply the Sense/Analyze/React loop to those actions.
Tip
Start by asking, “What are my viable domain actions?” As the number of domain actions that you can take is limited, you need to consider only a small number of use cases.
In this workflow, you have the two interactions of a technician with the patient, and you have the ophthalmologist/optometrist reading eye images to check for the presence of eye diseases. The interactions with the patient consist of an initial briefing and then positioning the patient in the imaging device so that a good image can be taken. You saw a moment ago why you can’t automate the briefing. What about positioning the patient? That requires robotics expertise, and your executives are adamant that you’re a software development company and not a robotics company. In your case, no action in the domain of direct patient interaction is viable.
What about interpreting the images? It turns out that interpreting images for certain eye diseases is complicated and that, in some cases, optometrists may miss important conditions. Professional interpretation is also costly and something that your hospital would save money on if you could make an alternate system that’s helpful when diagnosing eye diseases. This use case is worth further investigation.
Further research from your data science team shows that there has been significant progress in the application of computer vision to medical diagnosis. You find that Google’s team created an AI capable of diagnosing cases of moderate to severe diabetic retinopathy [49]. You have enough data from past optometry exams that you can train AI on that data. To make sure the Sense/Analyze/React loop is applicable in this use case, you need to cover only the Sense part. That proves to be easy; you already have an image of the retina of the patient, and you can send that image to your AI system.
2.4.2. General principles for monetizing AI
The previous example showed you how to find an opportunity for using AI in one business scenario. This section shows you what general principles you can extract from this example. Figure 2.4 shows those principles for applying AI.
Figure 2.4. General principles for applying AI to a business problem. The basic idea is to make sure that you’ll be able to implement all parts of the Sense/Analyze/React loop.
The approach shown in figure 2.4 covers each part of the Sense/Analyze/React loop:
- Sense— Can you collect the data you need? How much does it cost to collect that data?
- Analyze— Can AI do that analysis under ideal circumstances, or has anyone ever succeeded in doing something similar with AI? Is it well known that AI has such a capability? Does your team have expertise in applying those AI methods? How difficult is it to apply them?
- React— Find a domain action that would be of value and possible for AI. What is the economic value of that action? This information lets you judge if automating that action with AI is economically viable.
Chapters 3 and 4 will discuss how to use business metrics to cover the economic aspects and the application of the Sense/Analyze/React loop. For the time being, lets concentrate on how to cover the React and Analyze parts of the loop. You need to answer the following two questions:
- Is there a systematic way to think about your business that helps to find domain actions that can benefit from AI?
- What are the high-level capabilities of AI?
Once you know the answer to these two questions, you can perform an analysis like the one shown in section 2.4.1 to find viable use cases for AI.
Examples in this chapter (and chapter 1) show why, to achieve success with AI, linking AI with business is much more important than specific algorithms and technology. AI isn’t bringing superhuman intelligence to the table; it possesses humanlike capabilities in limited domains, such as image recognition. It can also apply those capabilities economically and operate with larger datasets than any human can. But you still need to figure out how AI’s capabilities translate into improving your business.
AI can sometimes find insights that escape human intelligence because of its ability to process large datasets. However, when operating in complicated domains, AI still lags behind humans. By itself, AI can’t figure out how to make money.
Peter Drucker believed that it’s more important to do the right thing than to do things right.[a] AI’s job is to help you do better analyses, and it could help you do things right, but only you can ensure that AI is applied to the right problem.
From the article “Managing for Business Effectiveness” [4]: “It is fundamentally the confusion between effectiveness and efficiency that stands between doing the right things and doing things right. There is surely nothing quite so useless as doing with great efficiency what should not be done at all.”
2.5. Finding domain actions
Now that you understand that the application of AI is simply a matter of applying the Sense/Analyze/React loop to some domain action, the next question is how you can systematically find domain actions that you can take. This section shows you how to find them.
There’s a limited set of high-level roles that AI can play in your business. Figure 2.5 shows those roles.
Figure 2.5. AI taxonomy based on the high-level role it plays in business. You could use this taxonomy to guide you in eliciting available business actions you can help with AI.
You can use AI as a part of the following:
- Decision support system— AI helps an employee or manager of your organization to make better decisions. Uses of such systems range from helping management make decisions affecting the whole organization to helping line employees in their day-to-day tasks.
- Larger product— AI could be just part of a larger product. Such a product has capabilities that AI may enable but that aren’t purely AI capabilities. An example here would be house cleaning robots (like Roomba [51]) or smart thermostats (like ecobee [37] and Nest [36]). In the case of a fully autonomous system, AI guides the system’s operation and makes its decisions without needing human involvement.
- Automation of the business process— AI automates some steps in the business process. Sometimes this is done to replace human labor; other times, it’s done to process datasets that are so large that humans can’t possibly handle them.
- AI as the product— You can package AI tools as a product and sell them to other organizations. An example is an AI product capable of recognizing images on traffic signs that will be sold to manufacturers of autonomous vehicles.
The rest of this section provides discussion of each of these bullet points.
2.5.1. AI as part of the decision support system
One of the most common scenarios for the use of data science in the enterprise today is the one in which AI is used as a decision support system. This section shows you how to use AI as a part of such a system to find domain actions.
AI as a part of the decision support system is the easiest scenario for elicitation of domain actions. In any decision support system, you’re already focused on the options you need to decide on. When using AI as a part of the decision support system, you should consider the user (or management team) whose decisions you’re supporting. Then you enumerate a spectrum of the decisions they can make. Finally, you ask yourself this question: “What information is needed to choose between these possible options?” The project is then organized around providing that information.
Suppose you’re supporting a large manufacturing operation. The operation has multiple large suppliers that ship thousands of components to it every day. A big cost concern for your organization is that if a certain percentage of the supplier’s components are faulty, the organization will spend a lot of time in the manufacturing process on troubleshooting problems. Such troubleshooting is costly. Even worse, the quality of the manufacturer’s own end product could suffer.
Although individual suppliers are a big part of your business, this sector is dominated by a few large suppliers, and your ability to force suppliers to improve the quality of their product is limited. How can AI help the management team for your manufacturing operation?
Start by enumerating the options regarding what the management team can implement and which ones are viable. Because your organization has little leverage on an individual supplier, the only viable business action your organization can take is to change suppliers.
What questions do you need to answer to change the supplier? Here we concentrate on one possible answer:
Ideally, you want to be proactive and switch suppliers before their quality deteriorates—by then, our manufacturing operation has already incurred costs. It’s difficult to decide to terminate the relationship with the supplier, because you don’t know what the cost of staying with the supplier will be. Ideally, you want to act proactively, based on where the trend of the supplier’s quality is heading. You don’t want to cut a supplier whose trend in quality is improving. Nor do you want to wait to switch suppliers if the trend is diving.
Based on management response, you now know that if you could use AI to analyze the historical trend of quality and to predict a trend in future quality, you’d have a system that would be useful to management. This is an example of using AI as a part of the decision support system.
This example also illustrates why customizing AI’s use to your own business case is better than applying AI solutions that worked for someone else. If you were a much larger customer to those suppliers, your management team might be able to negotiate the terms of the relationship, as opposed to just switching suppliers. Examples of adjusting such relationships might include escalating problems to the supplier’s management or asking for monetary compensation for defective parts.
While those might be viable actions for customers of your suppliers that are much bigger than your organization is, they aren’t viable actions for your organization. Generic AI solutions tailored to much bigger organizations might focus on actions you can’t take.
One final question in this scenario: Supposing that you’re a large organization with many departments, on which level of granularity should you request that business actions be supported by the decision support system? You should consider the options that are directly within the scope of responsibility and execution of the team that’s performing the analysis.
Warning
It’s critical that you choose the right level of organization to look for possible actions you can take. If you finish with a list that enumerates 20 or more different options that you believe you can take, the granularity level on which you performed analysis was wrong.
When discussing the use of AI as a part of a decision support system, the danger lies in diving in too deep. If you’re applying AI as part of a decision support system for the team of senior managers, then you should analyze the actions that senior managers take, not the actions that each individual worker working in their organization can take. Don’t analyze the actions an intern can take on their first day.
Tip
The decision maker doesn’t have to be a high-level manager. Imagine an AI that recommends to your sales force which customers to approach and displays a dashboard with the further information about each customer. However, that AI leaves the final choice to the individual sales professional. Such an AI is a decision support system.
2.5.2. AI as a part of a larger product
Another common situation occurs when AI capabilities are part of a larger product. In this situation, a key characteristic is that the end customer isn’t buying the AI itself; they pay for some capability that the larger product wouldn’t have without using AI. This section shows you how to use AI in the context of a larger product.
AI as a part of the product itself is already extremely important. Examples include products that range from smart speakers (Amazon Alexa [52], Google Home [53,54], and Apple HomePod [55]) to autonomous vehicles [38]. Although you can think of AI as a way to differentiate products, you’re generally better off thinking about it as an enabler of your value proposition to the customer.
Tip
Few people would buy a product specifically because it uses AI. The key question is, “What value is the product providing for your customers?”
There might have been a time when saying “We use AI” was a viable marketing/fundraising technique, but that time is over. Over time, AI will play the same role in autonomous products as an engine plays in a car today: you can’t go anywhere without it. However, most car buyers don’t care about a particular engine, but rather its ability to move the car from point A to point B.
An example of an AI product that also loops in humans is a home security company that uses an AI-powered device as a part of the security system. What are the relevant actions that such a system can take? For one, it can sound an alarm if it believes there’s an intruder.
For various cost and liability reasons, management will probably require that the final action of sounding the alarm or calling the police must always be initiated by a trained, live operator in a monitoring center. Management could also decide how many operators will be assigned to monitor the properties. This business would be much more profitable if a single person could monitor multiple secured properties.
AI could be leveraged in such a system to help the operator seated in the monitoring center. If AI can recognize faces, it can also sound an alert when humans are in the house who are not part of the family that lives in the home. The AI can then notify the operator so that a check can be made and, if necessary, the operator can raise an alarm.
When AI functions as part of a larger product, that product operates somewhere within the physical world. Because the customer is paying for some capability the product has, not for the fact that the product is using AI, start with how the product functions. Which potential actions could the system carry out? Once you know the set of possible actions, the next question is, “When should the system take each one of those actions?”
Note
When AI is part of a larger product, the product itself could be fully autonomous, or it could be a hybrid product that performs some functions automatically, while depending on humans for other tasks.
An example of a fully autonomous system would be a vacuum cleaning robot like the Roomba [51]. In this case, the vacuum needs to clean the whole room. The relevant domain action is, “Where should I go, and what areas should I avoid?”
AI can be used to provide navigation capabilities for the device in its environment. Note that such an AI could range from a sophisticated navigation system to relatively simple operations. A robotic vacuum can use AI to learn the layout of your rooms and recognize changes in that layout. You can also trade sophisticated mappings of the room for a bigger battery, allowing obstacle avoidance using a time-intensive trial-and-error approach.
That bigger battery is another example of the whole system being more important than the choice of AI algorithms. A few years back, it was simpler (and cheaper) to add a larger battery to increase run time than to spend a lot of time and money on significantly improved AI navigation.
In the context of a fully autonomous product, you also need to consider not only what actions the product can take, but that some actions and outcomes are neither desirable nor permissible. You don’t want to watch an expensive robotic vacuum such as the Roomba crashing down the stairs.
It’s important when using AI as a part of a larger product to consider not only the capabilities you’re planning to add in the initial product, but also the whole roadmap of product capabilities that you plan to add later.
Often, your product is a physical system shipped to the customer. For example, in the case of the AI-powered autonomous vehicle [38], you’d ship the vehicle itself. Once the vehicle is delivered, an additional capability could be added to it as a software upgrade. But you’re stuck with the sensors and effectors (engine, brakes, steering mechanisms, horn, signal lights, headlights, and so forth) that are shipped with the car. Once you distribute physical systems to your customers/users, it’s often impossible (or expensive) to add the capacity to perform new actions that you didn’t envision at design time. Whatever autonomous cars we have in the future, it’s a safe bet that some of their capabilities will be fixed at the time the car is manufactured and will be difficult to change later.
2.5.3. Using AI to automate part of the business process
One of the uses of AI that’s getting increasing attention in both industry and the popular press is the use of AI to perform actions that previously required humans. This section shows you how to apply AI to optimize existing business processes.
Suppose you have a facility that’s using CCTV cameras and security guards to monitor it. Looking at the screens is part of the workflow of the security guards. AI could be used to make this part of the security guard’s workflow more efficient by monitoring the video streams and highlighting unusual situations.
When you’re looking at using AI to automate part of the business process, start by sketching out that process and then ask, “Can any of these steps be made more efficient or eliminated using AI?” This is using AI to perform a one-to-one task replacement: the task that used to be done by humans is now done by AI.
As the capabilities of AI and humans differ, a one-to-one replacement of tasks performed by people with performing them using AI is complicated and expensive. In most workflows, a few tasks are essential, and they prove to be difficult to automate, even if the most time-consuming function of the job is automatable!
In practice, it’s usually necessary not only to apply AI to steps in an existing process, but also to re-engineer business processes. Re-engineering should separate out operations that are easy to automate with current technology into a separate step of the workflow. Then you assign AI to only those parts of the process that are easy for AI but time-consuming or error-laden for people.
The use of AI for automation is a controversial topic. If AI replaces a human in performing some action, and that action is the primary purpose of that person’s job, that job can now be in jeopardy.
There are significant costs that should be considered when eliminating jobs. Foremost are the costs to the people whose jobs disappear. There are also costs to your company, not just monetary, but also in the goodwill of both the public and your remaining employees. It’s important to keep that human perspective in mind when you talk about automation of your processes, and to understand that this scenario is often a zero-sum game.
If you’re limiting yourself when thinking about AI only to scenarios in which you’re replacing jobs with AI, then you’re actually missing an opportunity. AI can allow you to create new businesses that weren’t possible or economical before. This scenario generates new jobs—not only jobs building and supporting that AI system, but also all other jobs that come with such a business.
Take, for example, using AI to monitor the behavior of pets when owners are at work. At the moment, no one is doing this, because such monitoring isn’t economically viable as a service if it has to be done by humans. An AI that’s capable of monitoring the behavior of pets and entertaining them requires people in the loop to handle some rare situations that AI can’t (for example, situations in which the pet appears to have a medical issue). Such an AI creates jobs for the people monitoring those pets. These jobs weren’t economically viable at all when 100% of the work had to be done by humans. Such jobs become viable once an AI handles most of the monitoring and humans handle the exceptions.
2.5.4. AI as the product
Sometimes you have an AI solution, or an infrastructure solution supporting AI, that you believe to be applicable to many business contexts and many different customers. When that happens, such an AI solution is valuable in and of itself and could be packaged and sold as a standalone product. This section talks about some special considerations that apply when you’re intending to sell an AI solution as a complete product.
You have a complete product when you have customers who are willing to pay for the AI capability that you can develop. There’s a long history of companies offering various analytical products (such as SAS [56] or IBM’s SPSS [57]), and AI-based products could be considered a continuation of this tradition, wherein complicated analytical capability is packaged in a format that customers can use.
Tip
You’re selling a product. The key question is whether you can find customers that are willing to purchase this product. With regards to the sales cycle, the fact that the product itself is based on AI is secondary to all other sales considerations.
But there’s a specific consideration that you must address when you base your product on AI. You must correctly assess the capabilities of your organization and your team regarding their knowledge of AI. There’s a vast difference between developing an unprecedented AI solution and applying known AI capabilities in a new and specific context.
Developing new AI solutions and capabilities is a different ballgame that requires significant prior expertise in the field. When you’re selling AI as a product, you must assess not only the ability to deliver an initial version of the product, but also your ability to out-innovate the competition.
Warning
Unless you have a team of experts in AI research working for you, stick to applying an existing AI capability to a new context. Avoid AI products that require you to develop new AI capabilities that no one else has demonstrated yet, because they’re unpredictable, difficult, and risky to develop.
On the other hand, if you understand a general AI capability, then there is much less risk in applying that capability to a product in some new field. For example, it’s known that AI is getting very good at recognizing the context of an image—that’s a general capability. If you can apply that capability to a specific area, you might have a viable product. One example would be software that’s able to recognize defects on a factory line. This could be invaluable, provided you know to whom you would sell it.
Some AI products are general frameworks that are (clearly) widely applicable, but others are specific to one category of problems.
If your AI solves one category of problems, it can stand alone as a product if you can find multiple examples where the use of your AI solution makes new business actions viable. If, instead, you find only a single example of AI producing a new business action, then you’re better off thinking about what you have as an example of AI being part of a larger product.
When trying to figure out what new business actions AI makes viable, you’ll be applying the techniques in sections 2.5.1, 2.5.2, and 2.5.3. However, instead of applying them to your own business, you’ll apply them to your potential customer’s business.
2.6. Overview of AI capabilities
Section 2.5 showed you how to find a business question on which you can act if you can pair it with the appropriate AI capabilities. This section presents the taxonomy of AI capabilities that helps you answer the question, “Is there a broad area of AI capabilities that could address my business problem?” Figure 2.6 presents such a taxonomy of AI methods.
Figure 2.6. Taxonomy based on AI capabilities. This framework groups broad areas of AI capabilities so that you can quickly check if any of them are applicable to the business problem you’re addressing.
This taxonomy is a modification of the taxonomy originally presented in Bill Schmarzo’s books [58,59] with the “Use of unstructured data” category expanded to highlight the use of AI in perceptual tasks. The main goal of this taxonomy is to guide a discussion between the AI expert and the business expert. Categories in this taxonomy follow:
- Know results faster. Here, AI helps you discover a result more quickly, and that has a business value in many scenarios [58,59]. Suppose that you run a car manufacturing plant and you’re assembling cars from parts that are made in one area of your factory. If you know that some car part is defective as soon as it’s made, you can discard it at once and never install it in the car. This is much better than learning that the part was defective after you’ve already installed it in the car and shipped that car to the customer.
- Predict some event that occurs in the future, based on current trends. You saw this technique used in section 2.5.1 when predicting the future quality of the supplier based on historical trends.
- Use structured data. Sometimes you can find the answer you’re looking for in one of the relational databases that you already have, especially
if you have a large volume of data [58,59]. There are also AI methods that work well with data that’s already in tabular format.[2]An example of such a method is gradient boosting. If you’re interested in the technical details of this method, see the discussions on Wikipedia [60] and the Kaggle website [61].
- Use unstructured data. AI methods can also help you process and comprehend a large quantity of unstructured data, such as text, images, video, and audio [58,59]. In this case, you can use AI methods to recognize the context of the image, video, or audio recording.
- Replace humans in perceptual tasks. This subcategory of unstructured data use is based on the fact that, in recent years, AI has matched and even exceeded human abilities on many simple recognition tasks, such as image recognition [62,63]. You can think about this category of AI as having the ability to perform simple perceptual tasks that humans easily and instinctually perform. An example of such a task is recognizing objects in a photographic image.
- Replace experts in perceptual tasks. This subcategory of AI capability also comprehends unstructured data, but here AI performs perceptual tasks that otherwise require a high-level human expert. Such an expert uses skills that, after years of training, have become instinctual. An example of this would be using AI to interpret medical imaging. In recent years, AI has demonstrated an ability to interpret medical images on a level that in some cases rivals human experts [64,65].
Now you see how we’ve found AI solutions applicable to the business problems presented in section 2.5. In all those examples, you start by finding an actionable business problem and then the domain actions that could be taken. You ask the question, “Can we apply any of the six categories of AI capabilities shown in figure 2.6 to this business problem?”
There’s no way to describe all the capabilities that AI has in any single book, including this one. AI is a rapidly developing area, and AI capabilities transform daily with the development of novel methods and applications. If you’re interested in the details of individual AI methods, you need an experienced data scientist or consultant to guide you through the details of the latest capabilities of AI.
The taxonomy presented in this section isn’t a substitute for AI expertise, but it’s a systematic way to frame a discussion between a business expert and an AI expert.
It provides common terminology and concepts in a way that’s easy to comprehend for the business users. If you’re an expert in AI, you can use the taxonomy presented in this chapter as a quick checklist for a class of methods and algorithms that should be checked for applicability to business questions.
2.7. Introducing unicorns
This chapter has shown you how to determine which business problems can benefit from AI techniques, but does your particular development team have the knowledge necessary to implement the solution you just proposed? This section helps you answer that question.
The skills we’re using on AI projects are still new (and rare), and there’s still some amount of confusion in the industry about the skillsets that data scientists and data engineers should possess. Because of the rarity of those skills, there’s a joke that such experts are unicorns. In this section, I’ll start by describing the skills that are often attributed to unicorns. Then I’ll explain why most real-world teams will never have all those skills. Finally, I’ll show you how to make sure your team possesses all the skills that the specific AI project you’re running requires.
2.7.1. Data science unicorns
Data science could be considered an umbrella term that covers many skills. A survey performed in 2013 lists 22 different areas that are part of data science [66]. Examples of those areas include topics like statistics, operational research, Bayesian statistics, programming, and many others. It gets worse! Today, there are new areas that would certainly be considered important (for example, deep learning).
Note
Clearly, a data science unicorn should be a world-class expert in each one of those areas, right? No, these are individually very complex areas. Many distinguished professors in leading universities spend the totality of their time and effort to be an expert in just one of those areas. Most likely, no one in the world has expertise (defined as being on a level comparable to the skills of the aforementioned professors) in all those disciplines. Even if such a unicorn exists, which AI projects would have the budget for one?
Why are so many different skillsets part of data science? Because different practical problems benefit from different skills. No single ML method beats out all other methods across all possible datasets.[3] Each of these methods emerged because when the AI community tackled real, practical problems, some of them worked better than others. After many years, we use a combination of many methods from different disciplines.
This is also known as the No Free Lunch Theorem [67].
Did the leading data scientists start by learning all the methods that they know today? To be an accomplished data scientist, must you first build a skillset that emulates the skillset of a famous data scientist? No. Often, the skills of two accomplished data scientists don’t match. Even among accomplished data scientists, it’s virtually guaranteed that one will have expertise in at least one area with which the other is unfamiliar.
The skillset of accomplished data scientists is often acquired by working through problems that benefited from specific types of AI methods. They had to learn those particular AI methods because they were necessary to solve a concrete problem in the domain in which they worked. Each new project brings them new skills, sometimes in new areas that weren’t previously part of their core domain of expertise. For example, in 2011, few people in the world, in business or academia, were working on what today is known as deep learning.
If you want to become a unicorn, work on problems worth solving. You’ll acquire a strong skillset along the way.
As a manager, you should look for two things when hiring data scientists for your team. You should look for a candidate who has skills in the core domain that your initial AI project is likely to use, but you also need them to have a demonstrated ability to learn new skills. Chances are good that, along the way, your data scientist will need to learn many new methods. When hiring senior data science team members, don’t just look for a strong background in one set of AI methods. Senior data scientists should have a history of solving concrete problems using a diverse set of methods.
Tip
Data science is a team sport. To completely cover all of the knowledge that’s part of data science, you need a whole team, so you must assemble a team with complementary skillsets.
How should you assemble your initial data science team? Your team needs both enough business expertise to understand your business problem and enough proficiency in AI methods to perform an initial analysis and determine if AI can address your problem. Keep in mind that on the way to delivering a full AI solution, the team will have to learn some new skills.
2.7.2. What about data engineers?
When we discuss AI, we often talk about operating with some datasets so large that they don’t fit on a single machine and require a big data framework to manage. While data scientists are proficient with the use of big data frameworks, they’re rarely experts in the details of those frameworks. As a result, you’ll need specialists who are primarily focused on the use of the big data frameworks themselves. We call them data engineers.
Just like data science, big data is a large area. Let’s take the example of just one popular product in the big data space—the Apache Hadoop framework [15]. A few years back, the distribution of one of the leading Hadoop vendors consisted of 23 separate components, each one of which was large enough that a separate book could be (and often has been) written about it [68].
The body of knowledge that falls under the umbrella of data engineering is much larger than any single big data framework. Data engineers often need to be able to operate in both an on-premise and cloud environment. Cloud services like Amazon AWS [11], Microsoft Azure [13], and Google Cloud Platform [12] have different platforms with significant differences between them. That means that in addition to the specialist skills in big data frameworks, data engineers you hire may also need to have a skillset in the cloud platform of your choice.
Clearly, the same limits that apply to data scientists also apply to the data engineers: they’re also humans and can’t know everything. Data engineers are characteristically experts in a few of the components of the leading big data frameworks.
2.7.3. So where are the unicorns?
I hate to break this to you, but it’s highly unlikely that you’ll find any single human that has a strong expertise in each one of the methods, products, and technologies that are part of data science and data engineering. At best, you could hope to find a couple of senior people who have strong experience in individual data science and data engineering topics and who have enough familiarity with other related subjects to talk with specialists in areas in which they themselves aren’t experts.
Although universities have started offering programs and degrees in data science and data engineering topics in recent years, it’s not likely that this problem is solvable through better education. The field of knowledge is simply too large, so you should have a realistic expectation of what these institutions can teach their students.
Warning
As a project leader, you must differentiate between the skillsets that your team possesses and skillsets that you need for your project. You should identify and close skill gaps. Don’t assume that your senior data scientists and architects know everything in their fields, and don’t impose expectations on them that they should. Such expectations only make people less likely to acknowledge skill gaps.
Project leaders must know where the knowledge gaps are in the team. Piloting an AI project that requires skills your team doesn’t already possess means that you need to close these knowledge gaps. You do that by applying gap analysis [69]. An example of a gap analysis between the skillset a team presently has and the skillsets that are needed is shown in figure 2.7.
Figure 2.7. Gap analysis between skills the team has and the ones needed. This analysis allows you to create a plan for how to address missing skills.
You perform gap analysis by applying the following steps:
- You first work with your technical team to sketch a high-level technical solution. Match the time spent on this technical solution with the likelihood that you’ll implement it. If you’re just considering the project, then the solution should be a high-level one. If you’re planning to initiate the project soon, then your initial technical solution needs significant detail.
- Based on the solution, take an inventory of the technical skills you expect the project will need to address your identified use cases. This summary of skills should be made by people who both are familiar with the business problem you’re trying to solve and have enough technical expertise to quickly identify a high-level technical approach to solving it.
- Which skills do your team members already have? Ask your team about their skillsets and avoid assumptions in this area—AI and data engineering are highly technical fields, and it’s easy to have unfounded assumptions.
- Find any gaps between the needed and current skillsets. These gaps are useful in estimating the project’s difficulty level for your team. Keep this list. If you decide to proceed with a project that addresses this business question, you’ll need to make a plan for how to close the gap. (You close knowledge gaps by training your team, hiring new team members, or hiring consultants.)
Understand that gap analysis is always performed based on the current situation. If you’re just thinking about an AI project as a possibility, you should perform this gap analysis on a coarse level with just an outline of a technical solution. For projects that are in progress, you need a much more detailed solution as a starting point. That means that throughout the project life cycle, you’ll typically perform gap analysis multiple times.
When you ask about gaps in technical skills, you’re asking your technical staff to admit to areas in which they personally don’t have expertise. If poorly handled, they might rightly consider it a landmine. Put some thought into this before you ask; it’s your job as a leader to make sure that you create an atmosphere in which it’s easy for team members to admit that they don’t possess some technical skills.
One preferred technique to create such an atmosphere is based on building trust among team members so that they can talk about issues like this. Other techniques you might find useful are asking for the skillset in private, creating an anonymous survey, or asking a trusted intermediary to approach the subject with your team members.
2.8. Exercises
The goal of this book is to help you develop practical skills you can use when running your project. To help you with that, the exercises in this section ask you to apply skills learned in this chapter to new business scenarios.
2.8.1. Short answer questions
Please provide brief answers to the following questions:
Think about a failed project in your enterprise. Would that project have failed in the exact same way if it also had a component based on AI?
Do you personally have enough knowledge of data science and data engineering to understand the gap between the technical skills that your team has and the skills that they need for this project?
Do you have a good enough relationship with your team members that they’re comfortable admitting the limitations of their skillset to you?
2.8.2. Scenario-based questions
Answer the following questions based on the scenarios described:
One of the important skills in applying a Sense/Analyze/React loop is to identify who will execute on the React part of the pattern. For the following scenarios, answer this question: Who or what will carry out the action and fulfill the React part of the Sense/Analyze/React loop?
- Scenario 1: You’re making an automated car, and the AI that you’re using will allow fully autonomous driving under all conditions (so-called Level 5 autonomy [38], in which there are no available controls for the driver).
- Scenario 2: You’re writing a recommendation engine in which products are suggested to the customer.
- Scenario 3: You’re writing an AI program to regulate a smart thermostat that controls the temperature in your home.
Use AI to create a new job. Find an example of an AI capability that would let you offer a new service that your organization doesn’t yet provide. (For the job to count as a solution to this exercise, it must be a job that’s so unrelated to the software development team that’s building the AI, that the person hired for the job is unlikely to ever meet that team.)
Suppose you’re using an AI algorithm in the context of a medical facility—let’s say a radiology department of a large hospital. You’re lucky to have on the team the best AI expert in the field of image classification, who has you covered on the AI side. While you’re confident that expert will be able to develop an AI algorithm to classify medical images as either normal or abnormal, that expert has never worked in a healthcare setting before. What other considerations do you need to address to develop a working AI product applicable to healthcare?
Apply the previous example from a hospital setting to a classification problem in your industry. What are the new considerations that exist in your industry as compared to the healthcare industry?
Provide an example of an AI that has replaced a human role but doesn’t provide as good of an experience as a human would.
You’re a manufacturer of security cameras, and you’ve developed an AI algorithm that can detect a person in a picture. Regarding the taxonomy of its role in your business, how would you classify this use of AI?
You’re an insurance company, and you’ve developed an AI program that, based on static images from an accident site, could recognize which parts of the car are damaged in a wreck. Can this replace an insurance adjuster?
Summary
- Managing AI projects doesn’t require expertise in the details of AI algorithms. Instead, you need to know how to explain the benefits of an AI project in business terms. What business problem is being solved? What business benefit does AI provide? How is that benefit measured?
- You can discover business actions you can take and those that may benefit from AI using a systematic process. Apply the taxonomy described in figure 2.5 to your organization.
- AI capabilities are based on being able to know sooner, predict, process structured and unstructured data, and perform perceptual tasks.
- AI can help your business by performing analysis that informs some concrete business action. AI opportunities arise when you can apply a Sense/Analyze/React loop, with the Analyze part based on AI capability and the React part based on concrete business actions you can take.
- No individual is an expert on all topics of AI, data science, and data engineering. Project leaders must identify and close any relevant gaps in the knowledge and capabilities a team has.