Just last week, I finished reading the book on Prediction Machines by Ajay Agrawal, Joshua Gans and Avi Goldfarb. Agrawal, Gans and Goldfarb are part of the academia at Toronto’s Rotman School of Management and are also actively involved in the Creative Destruction Lab. Together, they have done a great job of separating intelligence from prediction in this book and present a very interesting criteria for organizations that are looking to transform their business. So in this blog, I try to summarize some of the key items I found particularly interesting in the book and how it can help you transform your business by understanding the simple economics of Artificial Intelligence and Machine Learning.
AI doesn’t bring Intelligence. It brings prediction.
Prediction Machines – The Simple Economics of Artificial Intelligence focuses on a critical component of Artificial Intelligence (AI) – which is prediction. AI is everywhere – in the cars we drive, in the cellphones that we can hardly spend a day without, in healthcare for developing new drugs and for medical diagnosis in cancer and other diseases. The use of AI is constantly increasing with entrepreneurs and investors coining it as an exponential technology. The book, however, goes a step further in assessing that artificial intelligence does not really bring us intelligence. Intelligence existed decades ago, although expensive for commercial uses at the time. AI brings a rather critical component of intelligence which is prediction.
The authors of this book have done an excellent job in applying basic economic principles to understand the growth of AI and its impact on society and business. Simple rule of economics tell that as the price of something as fundamental as technology drops, i.e. as a commodity becomes cheaper, its demand will increase, and such significant price drops means opportunities the world has never seen before. For example: when arithmetic and light became cheap, its use cases increased significantly which created more shared value and innovation that wasn’t possible when these technologies were expensive. The same is true for prediction. Prediction existed decades ago with weather forecasts still being done in the 1940s and 1950s. However, with the advent of AI, prediction has and will continue to become increasingly cheap and we are going to see it emerge in many new places, perhaps even tackling on problems that weren’t traditionally understood as prediction problems (like navigation and translation). This leads us to understand that organizations can evolve their strategy around AI and prediction machines, although significant consideration should be given to the trade-offs associated with this technology. Many of these trade-offs surround around decision making, judgement, tools, organizational strategy and priorities, and society.
Prediction at its basic level is the process of filling information. It takes the information you have which is often called “data” and uses it to generate information that you don’t have, often referred to as prediction outcome. But for any prediction machine to employ machine learning or artificial intelligence and predict an outcome, it needs data and data is expensive.
Really? Is data expensive?
Machines learn through data. And the variety, quality and quantity of data has considerably changed in recent years. Companies have produced and consumed large amounts of data every day even before the recent enthusiasm over AI. And now with the AI revolution, specialized fields in data management and big data have emerged which solely focuses on how to efficiently collect, store and maintain data for machine learning and AI purposes. But collecting data is expensive and is perhaps considered an investment. This is because it is thought that having more data will create better prediction. However, this is not always true, and can be evaluated using statistical and economic factors that shape the value generated by data collection.
Data, for AI, can be classified into three main categories. First is the training data which is used to train the system i.e. to create a baseline foundation against which to predict. Second is input data which is used as an input to the model to predict the outcome. Finally, there is feedback data which is used to continuously improve the prediction accuracy. For organizations to develop a sustainable strategy, it is critical to view data collection from a statistical and economical perspective. Statistically, adding more data does not create any higher value than previous data. It has diminishing returns. From an economic point of view though, collecting data has enormous value if it pushes the application above a certain threshold, i.e. from unusable to usable, or from being regulator non-complaint to being fully complaint. Thus, organizations need to carefully evaluate this trade-off and create a balance between the cost of data acquisition and the benefit gained from enhanced prediction accuracy by looking at what they are trying to achieve and deciding if adding more data will in fact result in enhanced prediction.
Keep in mind though, that Prediction is not Decision…
It is also important for organizations to realize that prediction is not a decision and prediction machines does not end up doing decision making. Prediction is only a component of decision. Another significant component of decision is judgement which involves determining the payoff with each possible outcome of the decision. As the prediction technology becomes increasingly cheap, better, faster and cheaper predictions will drive the value of human judgement high which in simple words indicate that humans are still needed even after the advances in AI. However, the book presents an interesting notion about predicting judgement itself, a concept called “reward function engineering” – which in the future may very well be called Artificial Judgement where the machine is able to decide by itself. However, at this stage, this technology is costly as there are enormous combinations of action-situations which can create many possible decision outcomes. Circling back to our original discussion, when/if predicting judgement becomes cheap, prediction machines can, in fact, become decision making machines.
So, then the real question becomes – Is understanding trade-offs in AI tools and decisions enough to execute your strategy. The simple and short answer to this question is No.
One of the major learnings that the book presents is that AI is not buzz word, it is a real tradeoff and organizations would have to understand the alignment of AI strategy with their corporate business strategy to be successful. As an example, many companies in different industries are employing AI to stay competitive. Organizations in fragile Canadian Oil and Gas industry are experiencing digital transformation on such an enormous scale that understanding data has become a core required skill in almost all job openings. These companies are deploying AI and machine learning methodologies to increase operational efficiency and reducing asset downtime by predicting the timing of the maintenance needed for the assets and scheduling it such that no facility outage is required. Low facility downtimes generate high tariff nominations, resulting in both direct and indirect network effects. However, companies in this sector have not deviated from their topmost priority, which is to produce and transport oil and gas to markets safely while maintaining a low carbon footprint for the environment. Instead, they have evaluated the trade-offs between innovation and competitiveness of implementing AI and are carefully executing the strategy in phases.
And the authors of this book have stated exactly that. If organizations want to implement an AI-first strategy which makes maximizing prediction accuracy as the central goal of the organization, it means that compromising on other goals such as revenue, customers, and user experiences could be required. However, the authors also state that companies whose primary vision is not to execute an AI-first strategy should still be aware of the disruption that AI can lead to. Established companies often like to wait-and-see, observing the progress of AI in their industry and then acting upon it. However, the highly evolving dynamic space of AI can soon leave these established companies in a catching up mode if their competitors implement AI tools and get ahead in the training and deployment of their methodologies. As such incumbent firms are often left with weaker, and sometimes capital unjustified incentives as compared to startups when implementing Artificial Intelligence. This is evident from the fact that nearly 150 startups from the Creative Destructive Labs have AI as one of their main goals in some form or the other.
Another strategic concern for companies to implement AI revolves around the timing of AI models. Machine Learning and AI models need time to train. When in development stage, these models can take a long time until they are ready as they are not exposed to real operating considerations. This could pose risk on both established and startup firms since it directly places them on the edge when it comes to competition. On the contrary, if these firms deploy these AI models into real operating conditions rather quickly to stay ahead of competition, they are always exposed to a greater risk of reduced customer satisfaction since AI models could be immature in prediction and can further bring liability if incorrect prediction leads to discrimination. Thus, companies should also evaluate the risks that come along with the nature of AI technology, while at the same time understand the trade-offs that may come around innovation and competitiveness.
The rise of Artificial Intelligence comes with many choices and trade-offs. Agrawal, Gans and Goldfab do an amazing job by not focusing on the right strategy for AI or the right tools needed to execute an AI strategy for organizations. This is because there is none. Instead, emphasis is on the trade-offs associated with each AI-related decision and how both sides of a decision can be evaluated to understand any direct or indirect business network effects, impact to human and society and help decide what’s best for us. Finally, the book presents itself in the center of foundation of AI – prediction and the how managers and leaders can utilize the knowledge gained from this book to transform their business in AI technology and in broader general digital space. While the AI technology may evolve over time, concepts and frameworks will still continue to hold strong in the coming future.