Prediction Machines

Prediction Machines Plot Summary

Introduction

In today's rapidly evolving technological landscape, artificial intelligence has emerged as a transformative force reshaping industries, business models, and everyday decisions. Yet beneath the hype and complexity lies a surprisingly simple economic principle: AI is fundamentally making prediction cheaper. This cost reduction in prediction—filling in missing information based on available data—has profound implications for businesses, decision-making processes, and the future of work. The core framework presented here strips AI of its mystique by viewing it through an economist's lens. Rather than focusing on science fiction scenarios or technical jargon, this approach reveals how AI affects strategic decisions by changing the cost equation of prediction. As prediction becomes cheaper, businesses face critical trade-offs regarding automation, job redesign, data strategy, and risk management. Understanding these trade-offs allows organizations to determine when AI should remain an operational tool and when it should drive strategic transformation, helping leaders navigate the uncertainty of this technological revolution with clarity and purpose.

Chapter 1: AI as a Prediction Technology

Artificial intelligence, particularly in its current form, is not the sentient being of science fiction but rather a sophisticated prediction technology. Prediction, in this context, means using information you have to generate information you don't have—whether about the future, present, or even the past. When Amazon recommends products you might like, Google translates languages, or a medical AI identifies tumors in scans, they're all making predictions based on patterns in existing data. What makes modern AI revolutionary is its dramatically improved prediction accuracy at decreasing costs. Consider image classification: in 2010, the best AI systems made errors in identifying objects 28% of the time, while by 2017, error rates had fallen below human levels. This improvement isn't merely incremental—an increase from 98% to 99.9% accuracy means mistakes decrease by a factor of twenty, which can be transformative in critical applications like medical diagnosis or autonomous driving. The economic impact of cheaper prediction extends far beyond obvious forecasting applications. Just as cheap arithmetic through computers transformed unexpected areas like photography (shifting from chemical processes to digital calculations), cheap prediction is finding applications in domains not traditionally associated with prediction. Self-driving cars reframe navigation as predicting what a human driver would do in each situation. Language translation becomes predicting the equivalent sentence in another language rather than applying grammatical rules. When prediction becomes cheaper, businesses use more of it, but they also discover entirely new uses. This creates a virtuous cycle where more prediction leads to better performance, which enables more applications, further driving down costs. The seemingly mundane process of filling in missing information becomes transformative when it crosses thresholds that make previously infeasible tasks possible, much like how cheap artificial light fundamentally changed human civilization by enabling activities regardless of natural light conditions.

Chapter 2: The Value of Human Judgment

As prediction machines become increasingly powerful, human judgment grows more valuable, not less. Judgment involves determining the relative payoffs of different outcomes—the rewards and penalties associated with making particular decisions. While AI excels at prediction, it cannot independently establish what outcomes matter or how much they matter relative to others. This fundamental distinction helps clarify what will remain uniquely human as AI advances. Consider a medical diagnosis scenario: an AI might predict a 70% probability of condition X, a 20% probability of no disease, and a 10% probability of condition Y. The physician must then apply judgment to determine whether to order additional tests, begin treatment, or wait and observe—weighing factors like the costs of misdiagnosis, patient preferences, and treatment side effects. The AI predicts probabilities; humans judge what those probabilities mean for action. Judgment requires understanding context, objectives, and values that aren't easily codified. When you use a navigation app like Waze, it optimizes for speed, but a human driver might override its suggestions based on other considerations—needing to stop for fuel, preferring scenic routes, or avoiding stressful driving conditions. The AI lacks this broader understanding of personal objectives beyond the single dimension it was designed to optimize. The complementary relationship between prediction and judgment means they rise in value together. As prediction becomes cheaper through AI, the economic return to judgment increases. Organizations will need to identify where judgment is most crucial and ensure those roles are filled by humans with appropriate expertise. This explains why despite automation in financial services, human advisors remain valuable—they understand client objectives, risk tolerances, and life circumstances that inform investment decisions beyond mere market predictions.

Chapter 3: AI Canvas: Deconstructing Decisions

To effectively implement AI in an organization, business leaders need a structured approach for identifying where prediction machines can create the most value. The AI Canvas provides this framework by decomposing decisions into their constituent elements: prediction, judgment, action, outcome, and three types of data (input, training, and feedback). This methodical breakdown helps organizations determine which components of a decision can be enhanced or automated through AI. The canvas begins by identifying the core prediction at the heart of a task. For example, a pharmaceutical company might need to predict which molecular compounds will bind effectively to disease proteins before costly laboratory testing. The AI Canvas would specify what information is needed to make this prediction (input data), what historical examples can train the algorithm (training data), and how results will be used to improve future predictions (feedback data). Judgment represents another critical element of the canvas—determining the relative value of different outcomes. In pharmaceutical development, this means weighing the benefits of targeting a specific disease against potential side effects, with different tolerance thresholds for life-threatening versus cosmetic conditions. The canvas forces clarity about these trade-offs that only humans can assess. The practical application of the AI Canvas often reveals that seemingly simple questions like "What are we trying to predict?" trigger profound discussions about organizational objectives. A business school using AI to select applicants must first determine what "best" means—highest test scores? Most diverse class? Greatest career potential? This specificity about objectives, often missing from mission statements, becomes essential when implementing AI. By methodically working through each element of the canvas, organizations can identify precisely where AI will deliver the highest return on investment while maintaining human judgment where it remains essential.

Chapter 4: Redesigning Jobs and Workflows

The integration of AI into business processes necessitates a fundamental redesign of workflows and jobs. Rather than simply automating existing tasks, organizations must deconstruct workflows into their component parts, identify where prediction machines add value, and then reconfigure tasks into new roles that maximize human-machine collaboration. This process often reveals surprising opportunities and challenges that wouldn't be apparent with a simplistic "automate or not" approach. When VisiCalc, the first electronic spreadsheet, emerged in the late 1970s, it didn't eliminate bookkeepers as many might have expected. Instead, it transformed their roles by automating arithmetic calculations while elevating the importance of financial analysis and decision support. Similarly, AI's impact on jobs follows multiple patterns: augmentation (enhancing human capabilities), contraction (reducing certain tasks), reconstitution (adding some tasks while removing others), and skill shift (changing which abilities matter most). Consider radiologists, whose traditional role centered on examining medical images to detect abnormalities. As AI becomes increasingly capable at this predictive task, the radiologist's role evolves rather than disappears. They now focus more on communicating findings to primary care physicians, helping determine when invasive procedures are necessary, training AI systems on new imaging technologies, and handling complex cases where human judgment remains crucial. The tasks within the job change, emphasizing different skills and responsibilities. The fulfillment industry demonstrates another pattern of workflow redesign. While robots now efficiently transport shelves of products to human pickers in warehouses like Amazon's, the "grasping problem"—picking up varied objects without crushing or dropping them—remains a challenge for automation. This missing link in the automation chain creates opportunities for humans and machines to collaborate, with AI handling prediction (identifying objects and determining the appropriate grasp) while humans provide feedback that improves the system over time. Understanding these workflow interdependencies helps organizations identify where to focus AI investments and how to redesign roles for maximum effectiveness.

Chapter 5: Strategic Impacts on Business Boundaries

AI fundamentally changes the calculus of where business boundaries should be drawn—what activities to perform in-house versus outsource, what data to own versus purchase, and what predictions to generate versus buy. These boundary decisions become strategic when prediction machines are powerful enough to transform a company's business model rather than merely enhance productivity within the existing model. Consider how Amazon might evolve from its current shop-then-ship model to a ship-then-shop approach as its prediction accuracy improves. If Amazon could predict with sufficient accuracy what customers want to buy before they order, it could send items preemptively, allowing customers to keep what they want and return the rest. This strategic shift would require not only better prediction algorithms but also investments in return infrastructure, potentially including a fleet of trucks for item collection. The strategic dilemma hinges on prediction accuracy—when it crosses a threshold where the increased sales outweigh return costs, the entire business model transforms. Ownership of data becomes another critical boundary decision as AI advances. While past data (training data) quickly becomes obsolete as customer preferences and markets evolve, organizations must decide whether to own the systems that generate ongoing input and feedback data. This explains why companies like Google and Meta have built or purchased advertising networks rather than simply buying third-party data. The key strategic question becomes: Is prediction core to your competitive advantage or merely an input to other processes? The airline industry illustrates how prediction machines affect make-versus-buy decisions in unexpected ways. Major airlines historically kept uncertain routes in-house rather than outsourcing to regional partners because weather disruptions required quick, flexible responses that contracts couldn't easily specify. Better weather prediction might shift this boundary by allowing more comprehensive contracts with partners, enabling major airlines to take advantage of partners' lower costs without sacrificing control. These subtle shifts in boundary decisions emerge as prediction machines reduce uncertainty across various business activities.

Chapter 6: Managing AI Risk and Implementation

Implementing AI introduces distinct risks that require thoughtful management strategies. Unlike traditional technologies, prediction machines can generate unexpected biases, security vulnerabilities, and quality issues that organizations must proactively address to realize AI's benefits while minimizing its dangers. Algorithmic bias represents one of the most significant risks. Studies have shown that AI systems can inadvertently discriminate based on race, gender, or other protected characteristics—like when Google's advertising system showed arrest-related ads more frequently for searches of Black-associated names or when Facebook's ad targeting system showed fewer STEM job ads to women because of cost differentials in reaching different demographic groups. These biases often emerge from statistical patterns in training data rather than explicit programming but can nevertheless create legal liability and reputational damage. Organizations must invest in testing AI outputs for disparate impacts and developing methods to mitigate biases when discovered. Security risks take multiple forms with AI systems. Input attacks involve manipulating the data fed to prediction machines, causing them to generate false outputs. Training data theft occurs when competitors systematically query an AI system to reverse-engineer its algorithm. Feedback manipulation happens when bad actors deliberately feed misleading examples to a learning system, as occurred when Microsoft's Tay chatbot quickly learned to post offensive content after interacting with users on Twitter. Each vulnerability requires specific safeguards, from monitoring unusual query patterns to implementing adversarial training techniques. The implementation process itself requires careful consideration of the stakes involved in different predictions. Low-stakes applications like product recommendations can tolerate higher error rates and benefit from early deployment to gather feedback data. High-stakes applications like medical diagnosis or autonomous vehicles demand more extensive testing and potentially keeping humans in the loop. The key implementation decision becomes when to deploy AI tools "in the wild"—earlier deployment accelerates learning but increases risks, while delayed deployment reduces risk but slows improvement. Organizations need thoughtful frameworks for making these trade-offs based on their risk tolerance and the specific context of each AI application.

Summary

The economics of artificial intelligence fundamentally reframes our understanding of this technological revolution: AI is making prediction cheaper, and this simple cost change carries profound implications across business and society. As prediction becomes more affordable and accessible, the value of complementary factors rises—particularly human judgment, data ownership, and effective action implementation. The organizations that thrive in this new landscape will be those that recognize these economic relationships and strategically redesign their operations around them. Beyond individual businesses, the broader societal implications of cheap prediction present important trade-offs that require thoughtful navigation. The tension between productivity and distribution, innovation and competition, performance and privacy—these balances will shape how AI's benefits are distributed and what risks emerge along the way. By applying a clear economic framework to AI decisions, from tactical implementations to strategic transformations, leaders can move beyond both techno-optimism and fear-based reactions to make reasoned choices that capture value while managing downside risks in this pivotal technological transition.

About Author

Loading...

Ajay Agrawal

Ajay Agrawal is a professor at the University of Toronto’s Rotman School of Management as the Geoffrey Taber Chair in Entrepreneurship and Innovation as well as the Professor of Strategic Management. Agrawal co-founded NEXT Canada, previously The Next 36 in 2010.

Read more