The object of all work is production or accomplishment and to either of these ends there must be forethought, system, planning, intelligence, and honest purpose, as well as perspiration.

Thomas Edison

For over a decade, big-tech companies have invested heavily in artificial intelligence (AI), and numerous pre-digital incumbents are starting to do the same. There is a shared belief between researchers and business leaders that artificial intelligence will open the door to fantastic growth opportunities, and we agree, but with a caveat. Pursuing artificial intelligence for the sake of it is a dangerous route to take, as it will be filled with wastefulness and disappointment. In many cases, it seems that these investments are technology led, meaning the implementation of artificial intelligence is itself the primary concern, rather than the business outcome. Investing in technology must be business led, not technology led, and artificial intelligence is no exception. Throwing artificial intelligence at everything, or ‘AI washing’, won’t solve all your problems. You need a well-defined business case. Also, before making an investment in artificial intelligence, you need to understand the technology and its potential, and, more importantly, its limitations.

Artificial intelligence is software and can be broken down into two overarching categories: weak and strong artificial intelligence. Weak artificial intelligence is software that can undertake one specific task as well as, or better than, a human. It is designed specifically for that one task and will therefore fail miserably if applied to something else. Strong artificial intelligence is as smart as, or smarter than, a human across the board and can self-teach. This is the kind of artificial intelligence you see in the Terminator movies. Strong artificial intelligence doesn’t exist yet. But weak artificial intelligence does, and this is what we will be focusing on in this chapter.

Artificial intelligence will eventually permeate all industries. However, as we discussed in chapter 1, on its own it does not provide a point of difference. To realise the growth benefits of artificial intelligence, it must be considered in a systems context, or what we call a system of intelligence. A system of intelligence consists of an algorithm (a process or set of rules to be followed in calculations or other problem-solving operations) that provides the ‘smarts’, the necessary computer processing power and a fantastic user experience. All of this can combine to produce tangible business benefits, such as examining thousands of contracts then extracting key clauses, helping researchers unlock new insights from massive data sets or managing the power usage of large buildings. In this chapter, we discuss the key components of a system of intelligence, starting with machine learning. We also reveal five crucial factors in the successful application of AI and present a range of ideas on how to construct a highly defensible digital moat.

The rise of machine learning

Machine learning gives systems the ability to learn and improve automatically without being explicitly programmed. It is based on the principles of how a child learns — through experience, repetition and, in most cases, feedback. There are three broad categories of machine learning: supervised, unsupervised and reinforcement learning. Machine learning has actually been around since the 1950s but has only recently begun to take centre stage, as the algorithms that fall under the machine learning heading have become layered, allowing them to crunch more information than ever before, which has led to the rise of deep learning (a subset of machine learning that examines computer algorithms that learn and improve on their own).

While much traditional data science focuses on explaining the past, machine learning focuses on explaining what will happen in the future. Simply, machine learning algorithms identify patterns within large data sets and, given a detected pattern, perform an action or produce an output. The real art comes from applying the right machine learning technique to the right data. Machine learning is nothing without data. As more data is fed in, machine learning algorithms are able to continually refine their ability to make predictions, increasing their accuracy and precision. Understanding the data you have, and the problem you want to solve, will in turn dictate the techniques, tools and algorithms you use. The overall goal of machine learning is to create intelligent programs, often called agents, which at their core consist of algorithms that learn and evolve. As already noted, there are three broad categories of machine learning.

SUPERVISED LEARNING

The agent is trained using massive data sets that consist of examples of the correct answer to a particular problem. As data is fed in, the algorithm teaches itself how to infer the desired output given a specified input. The more training data the algorithm is fed, the more accurate the agent’s inference capabilities become. It is important to note that the training data is labelled, making it easier for the algorithm to ‘learn’. Once the agent is adequately trained, and provided with similar input data, it will classify the data as it was trained to, hopefully producing the desired output. Issues arise when the data used for training doesn’t represent the real world or contains biases that can have detrimental consequences. For instance, using historical data to create the algorithms that determine credit rating scores may give white males a better credit rating score compared with females or people of non-Caucasian background, because historically the employment market, and therefore income and ability to maintain strong credit, prejudicially favoured white males. Historical data sets can be riddled with human biases, which can lead the agent to make unfair or erroneous decisions. Although the agent may seem smart, it is only as intelligent as the data you feed it, and any human biases within the training data will flow through into its decision making.

UNSUPERVISED LEARNING

In this case, the agent is given unlabelled data and no defined desired output (unlike supervised learning, where the outcome is defined during the agent’s training). As data is fed into the agent, it makes its own classifications and mappings, attempting to cluster the data into similar buckets. The agent is given the freedom to find hidden patterns in the data without being given any direction. As in supervised learning, the data plays an integral role in the patterns the agent detects. This can lead to some very interesting insights. For example, online user behaviour can be monitored by the agent, which will learn over time what ‘normal’ user behaviour looks like, then be able to identify potential cyber threats by detecting abnormal behaviours. If, say, the agent picks up a log-on from a location that has never before been used by a user of the application, this may prompt the agent to issue a warning notification to the user or to shut down the account, depending on the nature of the application.

REINFORCEMENT LEARNING

Reinforcement learning agents actually mimic the way in which animals and humans learn. The agent is trained by receiving instant feedback on how it interacts with its environment. The actions that the agent can perform in its environment are predefined, as is the environment itself. Given an input, the agent will perform an action, and if it is the correct action, the agent is positively reinforced. As it interacts with its environment, it learns the correct action given a specific input, maximising its performance over time. The great thing about reinforcement learning is that the training environments for these agents can be simulated, meaning training data can be created at very low cost, such as an autonomous vehicle learning how to drive in a simulated environment. Google offered a powerful example of reinforcement learning by reducing the energy consumption of its data centres by 40 per cent. The company allowed its algorithm to experiment with different data centre configurations until it learned how to optimise power consumption.

The perfect companion to your data

As we discussed in chapter 4, data has become an invaluable commodity in the digital economy, driving insight and knowledge in a way that was previously unimaginable. This has been driven in part by networks that connect devices, such as computers, mobiles or sensors, enabling the rapid transmission of data from the source to a central storage location. The technologies and protocols that form modern networks are continuously improving, meaning more data can be moved at a quicker rate. Traditionally, all this data would be stored locally in expensive data centres. Today, however, organisations have the luxury of cheaply renting storage from a public cloud provider. The cloud enables applications to be deployed globally and massive data stores to be spun up, all at the touch of a button.

Machine learning algorithms are trained by data, and this learning process requires a huge amount of computing power. Another factor that has helped make AI accessible is the drop in price of hardware — specifically, the processors and computer chips needed for this power. As the hardware improves, more complex machine learning algorithms that demand more computing power can be used, slurping up massive swaths of data on a scale that was not previously possible. This has significantly improved the accuracy with which these algorithms can make predictions, as well as opening doors to new techniques and algorithms. The end result is a virtuous cycle, in which the improvements made to the speed and capacity of computer processors have enabled new applications, begetting more demanding and effective algorithms, driving the demand for more data, which in turn requires more demand for computing power. This virtuous cycle will continue to facilitate new products, business models and strategies, which will impact all markets on a global scale.

Now cast your mind back to chapter 1, where we introduced you to the flywheel, and to chapters 4 and 5, where we showed how user data is used to improve the experience of your product or service, and platforms induce network effects by creating demand. These concepts really gain momentum when you bring machine learning into the mix. Think about a system that autonomously teaches and improves itself as more people interact with it, creating a better user experience, leading to customer lock-in because the experience outmatches anything else on the market.

There is no doubt you are already learning from your customers. However, if you can automate the process, you create an extremely scalable system that is able to learn without human intervention. And as we discussed in chapter 5, frictionless scalability is an extremely important aspect of building out your platform. The more autonomous your system is, the easier it is to get your flywheel spinning, and the faster it spins. Remember, first-mover advantage is critical, as, just like network effects, the growth associated with data network effects is exponential.

As you can see, machine learning has great potential to augment the value of your data. However, putting the right algorithm in place is a difficult task. The rest of this chapter will focus on helping you successfully implement machine learning.

Five crucial factors in applying AI successfully

Investment in artificial intelligence and machine learning technology is only going to increase. As more money is invested, the value of the resulting algorithms and the associated knowledge actually diminishes over time, as they become more abundant and less rivalrous in nature. Algorithms themselves will become less of a competitive advantage. According to entrepreneur and artificial intelligence expert Beau Cronin, the most important factors in the successful application of artificial intelligence and machine learning are data, computing power, algorithms and talent (see figure 6.1, overleaf). He argues that data is the hardest to come by, followed by talent, computing power and algorithms.

With respect to data, we completely agree that this is the hardest part to come by. This is because the data sets required for machine learning need to be large and complete. Typically, pre-digital incumbents do not have these types of data sets for many of the potential use cases of machine learning because they haven’t had the infrastructure in place to collect them. We also believe that data is fundamentally the most important, not only with respect to artificial intelligence and machine learning, but also to your economic engine, for without data, there can be no learning, no data network effects and no flywheel.

Many algorithms are accessible online and free, but the data has to be unique if it is to provide any sort of competitive advantage. Acquiring unique data sets that can be useful here requires a considerable amount of effort and time. This is why data is one of the three key enablers that make up the Digital Maturity Index (DMI), which we discussed at the beginning of part II. Given that we already have a chapter dedicated to data, we’ll focus on the other three components here.

There is a vast array of algorithms and techniques that can be used in the application of machine learning, depending on the problem. For instance, the algorithm to determine the likelihood of something happening, such as converting a potential lead into a sale, is called a classification model, while predicting how much of something is likely — for example, the total customer spend on your ecommerce site — is called a regression model. The algorithm is the brain of artificial intelligence, and in machine learning it evolves over time as it learns from the data. Applying the right algorithm is integral to the solution’s success. Applying the wrong algorithm can lead to lacklustre results and wasted money. Knowing which algorithm to apply, and applying it effectively, is a function of having talent that is knowledgeable in both artificial intelligence and your business.

Up until this point, we wholeheartedly agree with Beau. But his artificial intelligence thesis is missing a key component that is integral to the successful application of artificial intelligence, and that is the user experience. Artificial intelligence needs to be implemented in a systems context. Bolting on artificial intelligence for the sake of it does not provide the user with a fluid and seamless experience. In fact, it probably detracts from the experience. The application of artificial intelligence should seem natural and enhance the customer’s interactions in an unobtrusive way. Currently, the application of artificial intelligence has been more technology driven than business driven, which has led to unnatural experiences — for example, online chatbots that accept only a small number of commands and spit out error messages if a user enters anything unexpected. These underwhelming results are caused by a lack of understanding on the demand side and overselling on the supply side.

Returning to this ‘systems’ approach, the application of AI has to be integrated so it fits snugly into the customer’s lifecycle, as well as the underlying business infrastructure, processes and personnel. Machine learning can be applied to many activities in the value chain, enhancing the value of that specific activity, and thus the overall value provided to the customer. One example of machine learning is forecasting demand based on historical data, ensuring the adequate amount of stock is available in the right location to meet seasonal demand. Another common example is the implementation of automatic processes to send customers targeted advertising based on their past purchase behaviour or interaction with a retailer. In these ways, algorithms can be used to enhance customer interactions, whether that be entirely digital, a combination of machine learning and human, or further along the value chain, sitting as a layer above your data, used to unearth new patterns and drive decision making. As we discussed in chapter 3, it’s all about starting small, gaining buy-in, then building out your machine learning capabilities. Over time, intelligent algorithms can become commonplace throughout your organisation, creating a system that evolves through learning, and providing your customers with ever-increasing value. We call this a system of intelligence.

Recalling chapter 1, artificial intelligence by itself will not be enough to construct a wide moat, but a system of intelligence will be. If you can create a system that integrates computing power, the right algorithms, vast and varied types of data, business talent and a great user experience, you have a very effective moat. The fuel of these systems is obviously data, but what generates the data is the day-to-day activities of your business and the interactions your customers have with you. The more customer interactions you can support, the more data you can ascertain.

This is where the power of the platform comes into play. Think of your platform as the distribution channel that creates new interactions and acts as a two-way conduit in which data can travel both up and down between parties. A system of intelligence enhances the experience of your platform, driven off the back of the data collected from your company’s day-to-day activities. It can also be used to encourage more customers and more interactions. More platform usage generates more data, which your system of intelligence can use to further improve itself and the offering. The more data you produce, the smarter your product becomes. This is the data flywheel we’ve spoken about many times in this book, and your system of intelligence plays an important role in this virtuous improvement cycle. There will also be a large human aspect in this improvement process, which will also be driven by the data you collect, but the more automation you build into the loop, the more your flywheel will spin.

Creating a system of intelligence

So how do you go about creating and implementing a system of intelligence? The key is to keep it simple to begin with. Author and systems theorist John Gall wrote:

A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be patched up to make it work. You have to start over, beginning with a working simple system.

This is especially true in creating a system of intelligence. Trying to over-engineer from the outset is destined to fail. It’s also important to start small and build momentum internally to win over key stakeholders. The creation of a system of intelligence is a highly iterative process that requires data experts and business domain experts to combine their knowledge and work together. When undertaking the task of integrating AI into your organisation, we recommend you take the following approach.

UNDERSTAND

It is critical that before you invest any money into AI, you first understand AI and its implications for your organisation and market sector. Make sure you do your research and upskill the necessary business leaders before you start consulting with vendors or hiring AI experts.

IDENTIFY

Second, identify a series of potential business-led AI projects — systems of intelligence. For each project, define the key metrics you want to improve through the application of these systems. These metrics must be easily measurable so you can continue to track the optimisation of your algorithms, as well as help define the success criteria for each project.

Look for areas where a decision needs to be made or an output produced given a specific input. Typically, these are repetitive tasks that require some form of human judgement. According to leading AI expert Andrew Ng, if it takes a human about one second to perform a mental task, it can probably be automated with AI, either now or in the future. Other opportunities lie in large data sets, as AI is fantastic at crunching through data to help make predictions and recommendations. Remember that the successful implementation of AI requires large data sets regardless of the activity, and that there are also ethical and legal considerations associated with using and storing some forms of data, such as customer data.

RANK AND PLAN

Once you have selected your projects and defined their success criteria, rank the projects based on their impact on the bottom line. When doing this, you need to consider how much it is going to cost to implement and maintain them. Conducting this research typically requires someone with AI knowledge, so if possible hire someone with this skill set to lead the initiative. There are many ways to go about implementing AI solutions, including using vendors or open-source solutions, so make sure those leading your AI initiative evaluate all possible options before moving forward (we discuss open-source software in more detail in chapter 8). Finally, include personnel costs, as training these systems requires the help of your expert staff.

In conducting this analysis, you should have enough information to put together a series of project plans. We don’t recommend executing them all at once. First choose the projects that aren’t too costly but will provide good returns. If successful, these initial projects will play an important role in gaining buy-in from key stakeholders within the organisation. Consider your first project as a proof of concept.

EXECUTE AND MONITOR

Once you have selected your first project, it’s time to execute. The most time-consuming aspect of implementing a system of intelligence is preparing the data. Choosing well-labelled and -structured data sets means much less preparation work is required. If your organisation is lacking in data expertise and infrastructure, you will struggle to implement any machine learning solution effectively, so ensure these foundations are in place before moving forward (see chapter 4 for more on data).

As with any software implementation, you will need to undertake rigorous testing before launching to ensure your system is producing the correct outcomes, so factor in adequate time for this. Because this is your first project, lay the foundations to building out the necessary infrastructure and skill base for all future projects.

Machine learning is a process of continuous refinement and thus not something that can be set up and left. Over time, as consumer behaviours or internal processes change, a machine learning algorithm can become less accurate. Therefore, these algorithms will need to be checked regularly to assess relevance and, if necessary, redeveloped.

There are a couple of other things you also need to consider.

Assembling your team

When it comes to choosing your team, remember that both business and technical skill sets are required. When selecting the technical members, it’s imperative that they understand your business, the market in which you play and the root of the problem you are trying to solve. They must also be able to explain in lay terms what they are doing and what they have done. If they cannot articulate how this will help your bottom line or the business imperative, then you’re already starting on the back foot. The question to ask here is: does the business case stack up? This is a two-way street, and it’s important that you and your business team take the time and make the effort to understand at a high level how artificial intelligence and machine learning work, as well as their potential and limitations. People have varying expectations, and you need to be realistic about what AI can do to ensure the project runs as smoothly as possible.

The five Vs of big data

One of the more challenging aspects of the application of artificial intelligence is in finding the right problem to tackle. You may have a number of business processes or customer touchpoints that could be optimised, but if you don’t have the data to train an algorithm in these areas, then this isn’t a good place to start. Instead, focus on lower hanging fruit. We will show you how to identify a use case for machine learning and a system of intelligence. To begin, and building on what you learned in chapter 4, you need to think about the five Vs of big data for your systems of intelligence:

1. Velocity. This is the pace at which data flows from its sources. A system of intelligence that has a data feed closer to real time is much more responsive to changes within its environment, enabling better decision making and customer experiences.
2. Volume. For best results, machine learning requires large data sets. The more data you have to train your system of intelligence, the better the output.
3. Variety. Having a large amount of variation in your data means your system of intelligence has a wider scope or environment in which to operate. This, in turn, means it can understand more variables and consider more aspects to make better decisions.
4. Veracity. If your data cannot be trusted, then obviously your system’s output cannot be trusted. You need to implement appropriate checks to ensure the data your system of intelligence is receiving is trustworthy.
5. Value. Will the intelligent algorithm you use, the data you feed it and the system of intelligence you create provide value? If so, who to? You should always be thinking about this in the application of artificial intelligence.

Now you’ve refreshed your sense of the role of data as fuel for your system of intelligence, you can focus on identifying a business-led use case for machine learning.

Constructing a highly defensible moat

Systems of intelligence can be used to replace or supplement expert judgement and manual decision-making processes. Predictive models are faster because automated decision making can be applied to millions of data sets simultaneously. For humans to do this work would be extremely expensive and time-consuming. These models are often more accurate than their human counterpart and don’t suffer from being inconsistent. Once the model has been trained, it is cheaper to run than a human employee, especially at scale (think digital distribution).

Artificial intelligence harnesses the power of these predictive models, but, as we’ve mentioned before, the algorithm alone is not enough to act as your digital moat. Your moat comes from creating a system of intelligence that drives customer interactions, and creates lock-in as the experience becomes more personalised and effective. So what constitutes a real system of intelligence that creates a defensible moat? Here, in no particular order, are five different ideas.

VERTICAL EXPERTISE

In many industries, there are specific domains where human capital, business-specific knowledge and intellectual property can act as a moat. Through harnessing this expertise, you can create a system of intelligence that will provide a competitive advantage relative to a system of intelligence created by a team without this expertise. Note that artificial intelligence does not tell you where and how it should be applied — that is a human job. This relates back to having both the technical and business capabilities in one team, and having them work together. So where and with whom does your domain expertise lie? Can it be combined with data to create a system of intelligence?

Just a quick note. Some don’t categorise focusing on one vertical as a ‘system’, but we certainly do. Even in one vertical there are numerous data sets from varied sources to collect, and these can be used to unlock huge amounts of value when coupled with machine learning.

PERSONALISATION AT SCALE

A system of intelligence must be scalable, in that the owner shouldn’t have to make large changes whenever there is a new user. The point is to build a system that can easily bring in new users, while providing them with a personalised experience from the first interaction. The marginal cost of providing a user a personalised experience through a system of intelligence is zero as it is powdered by software. Having to create this bespoke experience manually, however, isn’t a system of intelligence and is costly with respect to marginal costs. This attribute is probably more critical with customer-facing solutions, but is important for internal systems as well.

MAN AND MACHINE

In many circumstances, the best outcome will be achieved through a mix of human and machine intelligence. This is similar to the vertical expertise moat, but different in that an expert and a trained intelligent algorithm will work together to provide optimal outcomes. Your system is then half man, half machine. This has proven to work well where there is a need for human emotion, judgement or creativity. For instance, AI can help a customer service representative troubleshoot problems faced by customers and find solutions extremely quickly, while the customer service representative can convey the solution in a manner and tone that is agreeable to the customer. This combination creates highly defensible moats, as both the intelligent algorithm and the human expert evolve over time, creating a hybrid flywheel.

NETWORK INTELLIGENCE

The network effects of data partners with an algorithm to provide incremental value to players across the value chain. The more parties involved, the more data it can use to provide increased value. The flywheel kicks in, and more value attracts more parties and more interactions. Network intelligence on a platform can improve the experience on both the demand side (recommendation engines recommend products to customers based on purchasing behaviours of similar customers, as well as other data sources like weather or seasons), and the supply side (pricing optimisation engines or forecasting engines help sellers set the price of their goods and forecast demand based on data sources like time of the year, the purchasing habits of similar customers, weather, and remaining and future stock).

INTEGRATION INTELLIGENCE

When smart algorithms meet physical hardware, an almost infinite range of opportunities present themselves. Integrating hardware and software so they are built to run optimally together creates tangible products that provide new experiences and open up new business opportunities. Incredibly strong moats can be created through building products that live in both the digital and physical worlds. A prime example of this is the iPhone 8, which was built with an Apple-designed GPU (graphics processing unit), allowing Apple to optimise the machine learning workloads running on the iPhones, which is especially handy for Apple’s intelligent personal assistant, Siri.

The barriers you raise will not come from intelligence alone, but from the system you create. These ideas offer five ways to think about how to create a system of intelligence. There are varying degrees of overlap between them, but they suggest a number of different ways to approach the creation of a highly defensible digital moat. Before proceeding, always ask yourself: will this improve the bottom line in a sustainable way? Whether through cost saving, creating a better customer experience that encourages more users, or helping you make faster and better decisions, all of these factors affect the bottom line. Technology for the sake of technology does not.

The power of these intelligent systems is the continual, progressive, automated learning from data. The question that then needs to be asked is: is my company’s organisational design and culture ready to respond proactively to this evolving intelligence? An intelligent system continues to improve over time, and your teams need to move with it. Otherwise, the investment in machine learning is pointless. In fact, it could cost your company valuable time and resources.

If you’re going to implement machine learning and create a system of intelligence, make sure your teams can respond and move with agility. Windows are small, opportunities are big, and missing out can be devastating. In the end, to utilise artificial intelligence, you need a culture that is willing to learn continuously. Therefore, the implementation of your system of intelligence should certainly be a key priority of your Engine B, which isn’t shackled to your old ways of doing business. Rather, your new agile team can operate in an agile way — continuously learning, adapting and improving with the technology, which in turn is driven by the people and hardware that interact with it.