Power And Prediction

Power And Prediction Plot Summary

Introduction

How will artificial intelligence reshape our economy and society? While predictions range from utopian to apocalyptic, understanding the true economic forces that determine AI's impact remains elusive. The disruptive economics of AI hinges on a fundamental insight: AI systems primarily enhance prediction capabilities, but their revolutionary potential emerges when they enable entirely new system designs rather than merely improving existing processes. The authors present a compelling framework distinguishing between three types of AI solutions: point solutions that enhance existing decisions, application solutions that enable new standalone decisions, and system solutions that transform entire organizational structures. This framework helps explain why, despite impressive technological advances, AI's economic impact has been uneven across industries. By examining historical parallels like electricity adoption and identifying how AI decouples prediction from judgment in decision-making, the authors illuminate why we currently exist in "The Between Times" – after witnessing AI's potential but before its widespread transformation of our economic systems. The theory reveals why some AI investments yield immediate returns while others require fundamental rethinking of organizational structures, and how this transformation will inevitably redistribute economic power.

Chapter 1: AI as Prediction Technology: Understanding the Core Function

Artificial intelligence, particularly in its modern manifestation, represents a dramatic reduction in the cost of prediction. Prediction, in this context, means taking information we have and generating information we don't have but need. While popular culture envisions AI as humanoid robots with general intelligence, today's practical AI is fundamentally about statistical pattern recognition and prediction – albeit extraordinarily powerful prediction. The breakthrough that catalyzed the current AI revolution came in 2012 when deep learning techniques demonstrated unprecedented accuracy in image recognition. These systems don't "understand" images in a human sense; rather, they predict what humans would say is in an image based on patterns learned from vast training datasets. This same predictive approach powers everything from language translation to medical diagnostics. The core insight is that prediction is not the entirety of intelligence, but a critical component that enables better decision-making when complemented by human judgment. Prediction and judgment work together in decision-making. Prediction tells us what might happen, while judgment determines how we should value different outcomes. When a self-driving car predicts a pedestrian might cross the road, judgment determines how to weigh the risks of sudden braking versus potential harm to the pedestrian. AI excels at prediction but lacks the value-based reasoning that constitutes judgment. This distinction explains why autonomous systems still require human involvement for setting parameters and handling exceptions. The economics of prediction technology follows predictable patterns. As prediction becomes cheaper through AI, we naturally use more of it, finding applications previously considered uneconomical. For instance, banks have long used prediction for fraud detection, which is why companies like Verafin found immediate success applying AI in this domain. However, in areas where prediction wasn't previously central to operations, organizations struggle to integrate AI effectively because they haven't built systems designed to utilize prediction. Amazon might use AI to predict your purchases so accurately that they could theoretically ship products before you order them, but implementing such a system would require redesigning their entire logistics infrastructure, particularly around returns. This illustrates why AI adoption often requires system-level changes rather than simple technology deployment.

Chapter 2: Point Solutions vs. System Solutions: The AI Implementation Spectrum

Point solutions represent the simplest way to implement AI – replacing an existing prediction task with a machine-learning alternative without changing surrounding processes. For example, a bank might upgrade its fraud detection algorithm to use neural networks instead of traditional statistical methods. The decision-making process remains unchanged; only the underlying prediction technology improves. Application solutions go one step further, enabling new standalone decisions that weren't previously feasible, such as a smartphone recognizing faces to unlock devices – something impractical before modern AI. System solutions, by contrast, require fundamental redesign of interdependent processes. They emerge when AI enables such significant improvements that maintaining existing workflows becomes counterproductive. When electricity first appeared, factories initially used it merely to replace steam engines (point solution), keeping the same shaft-and-belt power distribution systems. Only decades later did they completely redesign factories with distributed electric motors at each workstation, enabling entirely new layouts and workflows that dramatically improved productivity. This historical parallel reveals why AI's most transformative impacts remain largely unrealized. The distinction between these solution types explains the "AI paradox" – why despite remarkable technological achievements, productivity growth remains sluggish. The 2020 MIT Sloan Management Review found only 11% of organizations reported significant financial benefits from AI. These successful cases typically implemented point solutions in prediction-intensive contexts. The other 89% face the challenge that their existing organizational structures aren't designed to capitalize on improved prediction. They exist in what the authors call "The Between Times" – after demonstrating AI's capabilities but before redesigning systems to fully leverage them. Consider healthcare's slow AI adoption despite obvious potential. Medical systems are extraordinarily complex networks of interdependent decisions across providers, insurers, patients, and regulators. An AI that improves diagnosis (a point solution) provides limited value unless treatment processes, payment systems, and institutional structures evolve to accommodate the new capability. By contrast, digital platforms like Google, Amazon, and Netflix were designed from the ground up to leverage prediction, which explains their early, successful AI integration. Their systems were built to oil rather than glue – designed for adaptability rather than stability – making them naturally receptive to AI enhancements.

Chapter 3: Rules as Glue: How Organizations Resist Decision Changes

Organizations operate on rules that provide stability and reliability. These rules – whether formal standard operating procedures or informal habits – reduce cognitive load and enable coordination. When Amazon warehouse employees follow specific picking protocols, or when airlines use preflight checklists, they're relying on rules that make operations predictable. While rules sacrifice optimization for reliability, they form the essential glue holding complex systems together. This organizational glue creates resistance to AI-enabled changes. Consider music streaming services like Pandora. While AI could personalize the number of advertisements each user hears based on their listening patterns and price sensitivity, implementing such a system would require changes to advertising sales strategies, user interfaces, and revenue models. The rule of "same number of ads for everyone" connects multiple organizational functions. Changing it impacts not just the user experience team but also advertising sales, content licensing, and financial forecasting departments. Even with compelling evidence that personalized advertising would increase profits, companies hesitate because the interdependencies make change risky. Educational systems provide another powerful example of rules as glue. Most schools operate on an age-based curriculum rule – students progress through grades based primarily on their birthdates. This rule connects teacher training, facility design, administrative systems, and parental expectations. An AI that personalizes learning content based on individual student progress might seem valuable, but its impact would be limited in the existing system. If faster-learning students quickly advance beyond grade-level material, most schools lack systems to support them, as teachers are trained for specific grade levels and resources are allocated accordingly. The scaffolding built around rules further complicates AI adoption. Organizations develop elaborate structures to manage the inefficiencies inherent in rules. Airports have transformed from transit points to entertainment centers partly because rigid security protocols create unpredictable waiting times. Rather than solving the underlying uncertainty, the system adapted by making waiting more pleasant and profitable. Similarly, insurance companies build complex actuarial systems around risk pooling rather than investing in personalized risk mitigation. This scaffolding represents substantial investments that organizations are reluctant to abandon. The key insight is that AI's potential to replace rules with decisions threatens not just the rules themselves but entire ecosystems built around them. Organizations resist such changes because they've optimized for stability within their current rule-based frameworks. True transformation requires dismantling both the rules and their supporting structures – a process that demands system-level thinking rather than incremental technology adoption.

Chapter 4: Decoupling Prediction from Judgment: The New AI Paradigm

Traditionally, human decision-makers integrate prediction and judgment as inseparable mental processes. When a doctor diagnoses a patient, she simultaneously predicts what condition might be present based on symptoms (prediction) and weighs the consequences of false positives versus false negatives (judgment). AI fundamentally changes this dynamic by separating prediction from judgment, allowing each to be optimized independently. This decoupling represents one of AI's most profound economic impacts. Consider Michael Jordan's decision whether to play basketball while recovering from a foot injury. Doctors predicted a 10% chance of career-ending reinjury if he played. Team owner Jerry Reinsdorf illustrated the stakes by asking Jordan: "If I gave you a bottle of pills and nine would cure you but one would kill you, would you take a pill?" Jordan's famous response – "It depends how bad the headache is" – perfectly captures the distinction between prediction (10% risk) and judgment (how much various outcomes are valued). AI excels at improving prediction, but judgment still requires human input about values and preferences. Decoupling creates new opportunities for specialization and efficiency. Insurance companies already use this principle when they convert complex risk predictions into simple premium prices for consumers. Rather than overwhelming drivers with statistical analyses of accident probabilities on different routes, navigation apps could simply state: "Taking the scenic route will increase your insurance premium by $1." The driver then applies judgment – is the scenery worth the extra dollar? – without needing to understand the underlying predictions. This separation allows prediction to be handled by specialized algorithms while judgment remains with the people best positioned to provide it. The implications extend far beyond insurance. Radiologists who once combined prediction (interpreting images) and judgment (deciding treatment plans) may find their role evolving as AI excels at image analysis. The key question becomes: once prediction is handled by machines, who is best positioned to provide judgment? Similarly, financial advisors who previously combined market prediction with client-specific judgment may need to evolve their services as AI handles more predictive tasks. Organizations must reconsider who makes decisions when prediction and judgment are provided separately rather than bundled together. Real-world examples of this transformation are emerging. Stitch Fix combines AI prediction with human stylist judgment to personalize clothing recommendations. CEO Katrina Lake describes their approach as "unique and personal selections by combining data and machine learning with expert human judgment." Companies that recognize the opportunity to redesign decision processes around decoupled prediction and judgment gain competitive advantage over those that simply try to automate existing workflows. This insight helps explain why some AI implementations create substantial value while others disappoint – success depends not just on prediction accuracy but on thoughtfully reallocating judgment responsibilities.

Chapter 5: System Mindset: Redesigning Organizations for AI Integration

The system mindset represents a fundamental shift in how organizations approach AI implementation. Rather than focusing on task-level substitution – identifying specific human activities that machines might perform – system thinking considers how entire organizational structures could be reimagined around AI capabilities. This distinction explains why leading technology companies have successfully integrated AI while traditional industries struggle despite substantial investments. Task-level thinking dominates current AI discourse. When examining radiologists' work, for instance, analysts identify thirty distinct tasks and evaluate which might be automated. This approach suggests incremental change: perhaps AI handles image interpretation while humans manage patient interaction and treatment planning. System thinking, by contrast, asks more fundamental questions: If we had powerful prediction machines, how would we design healthcare delivery from scratch? This perspective might lead to entirely different organizational structures where diagnostic AI enables preventive interventions before symptoms appear, or where medical expertise is distributed differently across specialties. The healthcare system illustrates both the promise and challenges of system-level AI integration. Despite numerous potential applications – from diagnosis to treatment customization and drug discovery – healthcare ranks near the bottom in AI job creation across industries. The system's complexity creates significant barriers to change. When researchers developed an AI that outperformed doctors at predicting heart attacks, implementation proved difficult because it challenged existing divisions of responsibility between physicians and administrators. The AI could identify which patients truly needed expensive cardiac catheterization, potentially saving costs and improving outcomes. However, doctors concerned about liability preferred overtesting, while administrators focused on resource allocation. Without system-level redesign addressing these tensions, even demonstrably superior AI prediction provides limited value. Building system solutions requires understanding how decisions interconnect. Digital twins – virtual replicas of physical systems – enable experimentation with new configurations before implementation. Team New Zealand used this approach to win sailing's America's Cup by creating an AI-powered simulator that tested both boat designs and sailing maneuvers. The simulator identified optimal coordination between engineering and sailing decisions that would have been impossible to discover through traditional methods. Similarly, Virtual Singapore allows planners to simulate changes to transportation systems before physical implementation. These examples demonstrate how system thinking enables organizations to design coordinated changes rather than isolated improvements. The challenge isn't technological but organizational. Most established businesses are optimized for reliability rather than flexibility, with carefully balanced interdependencies between divisions. Changing one element without adjusting others creates friction. Startup advantages in AI often stem not from superior technology but from freedom to design systems without legacy constraints. Organizations that develop a system mindset – understanding interconnections between decisions and designing coordinated changes – will capture disproportionate value from AI. This requires leadership that can envision transformation beyond departmental boundaries and manage the inevitable tensions that arise when established roles and responsibilities evolve.

Chapter 6: Power Redistribution: How AI Creates Winners and Losers

AI adoption inevitably redistributes economic power – the ability to extract value based on controlling scarce resources. This redistribution occurs not because machines themselves gain power, but because AI changes which human skills, organizational capabilities, and business models create the most value. Understanding these power dynamics helps explain both resistance to AI adoption and strategies for navigating the transition. Economic power shifts emerge clearly in historical parallels. When electricity transformed manufacturing, vertically integrated producers lost power to more flexible competitors who redesigned factories around distributed motors. Similarly, AI enables new approaches that challenge incumbent advantages. Financial fraud detection, previously handled by individual merchants evaluating customers, shifted to centralized credit card networks with superior data and algorithms. Weather forecasting evolved from localized predictions to sophisticated modeling that benefits large-scale operations. In each case, power shifted from those with local knowledge to those with superior prediction capabilities and data access. The redistribution dynamic explains why AI adoption faces resistance beyond mere technological challenges. In Flint, Michigan, an AI system accurately predicted which homes had lead water pipes, enabling efficient replacement. However, local politicians initially rejected the system because it undermined their control over resource allocation across districts. Only after a court mandated its use did the city fully implement the AI, dramatically improving efficiency. This pattern repeats across domains – predictions that could improve overall outcomes often threaten established decision-makers' discretion and authority. First-mover advantages further concentrate power in AI-intensive industries. Unlike physical production where economies of scale eventually plateau, prediction systems benefit from virtuous cycles of data accumulation. More users generate more data, which improves predictions, attracting still more users. This feedback loop creates powerful network effects, explaining why companies race to deploy AI even before profitability is clear. Google's search engine maintains dominance partly because its vast user base provides continuous feedback that improves prediction accuracy, making it difficult for competitors like Bing to catch up despite substantial investment. Crucially, machines themselves don't gain power – humans who control AI systems do. When Amazon tracks warehouse productivity, the decision authority remains with management, not algorithms. However, by enabling decisions to be made at different times and places, AI changes which humans exercise judgment. Judgment at scale becomes possible when standardized parameters replace case-by-case discretion. This explains why some industries consolidate around platforms with superior prediction capabilities while others remain fragmented. Understanding these power dynamics helps organizations anticipate how AI adoption will reshape competitive landscapes and internal authority structures.

Chapter 7: Building AI Systems: From Blank Slate to Implementation

Implementing system-level AI solutions requires methodical approaches that overcome organizational inertia. The blank slate method provides a powerful starting point by temporarily setting aside existing constraints to envision optimal systems built around AI capabilities. This approach reveals opportunities obscured by incremental thinking and helps identify the minimum viable changes needed for transformation. The AI Systems Discovery Canvas offers a structured framework for blank slate thinking. Organizations begin by articulating their fundamental mission, then identify the fewest possible decisions required to achieve that mission if powerful prediction machines were available. For each core decision, they specify what prediction is needed and what judgment must be applied. This exercise strips away accumulated complexity to focus on essential functions. In home insurance, for instance, the mission reduces to three key decisions: marketing resource allocation, premium pricing, and claims validation. With AI predictions for customer lifetime value, claim probability, and fraud detection, an insurance provider could design a far more efficient system than today's bureaucratic structures. Moving from canvas to implementation requires understanding the balance between coordination and modularity. When decisions interact, changing one affects others, potentially creating a "bullwhip effect" where improved prediction in one area increases uncertainty elsewhere. System designers must choose whether to coordinate interrelated decisions or create modular structures that insulate them from each other. Successful implementations typically combine both approaches – using coordination where synergies are high and modularity where independence creates flexibility. Redesigning emergency medical systems, for example, might coordinate paramedic dispatch with hospital resource allocation while keeping billing systems modular. Digital twins enable experimentation with system designs before physical implementation. These virtual replications simulate how changes to one decision affect others, helping organizations identify unintended consequences and optimal configurations. Team New Zealand's America's Cup victory demonstrates this approach – their digital twin tested both boat designs and sailing techniques, revealing how these decisions should be coordinated for maximum performance. Similarly, Doosan Heavy Industries used digital twins to optimize wind farm operations by connecting turbine performance predictions with maintenance schedules and energy grid commitments. The ultimate challenge is managing the transition from existing to new systems. Most organizations can't "turn off" current operations while building replacements. The most successful approaches typically combine centralized vision with decentralized implementation, allowing flexible responses while maintaining strategic direction. Starting with clearly defined boundaries – perhaps a new customer segment or geographic region – provides space for system experimentation without disrupting core operations. As the new system proves its value, it can gradually expand, eventually replacing legacy approaches. Organizations that master this balance between visionary redesign and practical implementation will capture disproportionate value as AI transforms their industries.

Summary

The transformative potential of AI lies not in automating individual tasks but in enabling entirely new system designs built around enhanced prediction capabilities. By decoupling prediction from judgment in decision-making processes, AI creates opportunities to redesign organizational structures, business models, and competitive strategies. However, this transformation faces resistance because existing systems are optimized around rules that provide reliability at the expense of adaptability – what the authors call "glue" rather than "oil." We currently exist in "The Between Times" – after witnessing AI's technological potential but before its full economic impact materializes. History suggests that truly revolutionary technologies like electricity require decades to transform economies because their greatest value comes from system-level changes rather than direct substitution. Organizations that develop a system mindset – understanding how decisions interconnect and designing coordinated changes rather than isolated improvements – will navigate this transition most successfully. Those who recognize that AI redistributes economic power rather than merely reducing costs will be best positioned to thrive in the emerging landscape where prediction becomes abundant while human judgment remains the essential scarce resource.

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

Related Books by Ajay Agrawal

Loading...

Prediction Machines

Ajay Agrawal

The Simple Economics of Artificial Intelligence

★★★½☆3.9