Capítulo 1
The Digital Oracle: How AI Prediction Will Transform Everything
When Ajay Agrawal and his co-authors published "Prediction Machines" in 2018, they thought they'd covered everything about AI economics. They were wrong. The real story wasn't about individual AI applications but how entire systems must transform to harness AI's power. This realization came when they incorrectly predicted Canada's first AI unicorn would emerge from AI research hotspots like Montreal or Toronto. Instead, it came from St. John's, Newfoundland-Verafin, a fraud detection company acquired by Nasdaq for $2.75 billion. Why? Because financial institutions were already designed for machine prediction. We're now in "The Between Times"-after witnessing AI's potential but before its widespread adoption. While point solutions happen quickly, system solutions require redesigning entire organizations. This book explores why some AI applications deploy rapidly while others advance slowly, examining not just the technology but the systems in which it operates.
Capítulo 2
The Electricity Parallel: History Repeats Itself
Despite Edison's invention of the light bulb in 1879, only 3% of US households had electricity twenty years later. After another two decades, that number accelerated to 50%. These forty years represented electricity's "Between Times"-the period between demonstrating a technology's capability and achieving widespread adoption. We're experiencing the same phenomenon with AI today.
The pattern follows three entrepreneurial approaches. Point solution entrepreneurs simply swap one prediction method for a better one, like Verafin replacing traditional fraud detection with AI. Application solution entrepreneurs redesign products around AI capabilities, like smartphone cameras identifying faces. But the highest value comes from system solution entrepreneurs who reimagine entire operations-similar to how Henry Ford couldn't have invented the production line without electricity enabling flexible factory layouts.
This parallels the productivity paradox observed with computers in the 1980s, when economist Robert Solow noted computers were "everywhere but in the productivity statistics." Similarly, a 2020 MIT study found just 11% of organizations reporting significant financial benefits from AI despite widespread deployment. The explanation lies in system-level change-AI will only reach its potential when entire systems of decision-making adjust to leverage its prediction capabilities.
The challenge is that such system changes disrupt existing power structures. When David Friedberg started The Climate Corporation, he transformed farming by moving decisions from rural America to San Francisco. His company's AI could show farmers precise field conditions, optimal planting times, and best harvest dates. This wasn't just a tool farmers could use-it was replacing how they made decisions. As Michael Lewis observed, no one asks "If my knowledge is no longer useful, who needs me?" Yet the industry recognized the potential, with Monsanto acquiring The Climate Corporation for $1.1 billion in 2013.
Capítulo 3
AI Is Simply Better Prediction
When most people think about AI, they imagine intelligent machines from science fiction-entities that think, reason, and have agency like humans. But today's AI is actually an advance in statistical techniques that dramatically reduces the cost of prediction-something we do constantly in everyday life.
The breakthrough came in 2012 when Geoffrey Hinton's University of Toronto team used deep learning to significantly improve image recognition. Their approach conceived image identification as a prediction problem: guessing what a human would say was in an image. By allowing for numerous attributes and their combinations, deep learning outperformed other algorithms and eventually most humans.
But predictions aren't the only input into decision-making. Judgment-determining the reward to a particular action-remains equally crucial. In the movie I, Robot, Detective Spooner resents a robot for saving him instead of a young girl during an accident. The robot had predicted Spooner had a 45% chance of survival versus the girl's 11%, and was programmed to value all human lives equally. Spooner believed the girl's life was worth more-a judgment call. When using prediction machines, we must be explicit about judgment.
AI's correlation capabilities must also be distinguished from causation. Predicting beyond your data's "support" can lead to inaccuracy. The toy industry shows strong correlation between advertising and revenue during December, but increasing April advertising wouldn't necessarily boost spring sales. The correlation might exist because Christmas drives both advertising and sales, not because advertising directly causes sales.
Verafin represents the exception rather than the rule for AI adoption. Three factors made it successful: prediction was already central to its business, its financial institution customers required minimal changes to adopt its products, and these institutions already knew how to handle predictive errors. Most businesses face a more challenging path to AI adoption-not just adding point solutions but undertaking system-level changes.
Capítulo 4
The Hidden Cost of Decision-Making
Economists don't actually believe people are perfectly rational calculating agents. They know from experience that real people don't match the rational models, but treating people as if they make deliberate decisions based on interests remains useful for understanding collective behavior.
People make countless decisions daily, from what to wear to what to eat. Some, like Barack Obama who wore only gray or blue suits as president, deliberately limit certain choices to preserve mental energy for more important decisions. As Obama explained to Vanity Fair, "I don't want to make decisions about what I'm eating or wearing. Because I have too many other decisions to make."
This creates a challenge for AI implementation. AI prediction is only useful when making actual decisions, not following rules. Organizations build systems with reliability through rules, but AI transforms rules into decisions, potentially disrupting system reliability unless redesigned to accommodate this shift.
Herbert Simon, who won both a Nobel Prize in economics and the Turing Award in computing, observed that people face the same constraints as primitive computers-limited computing resources force us to "make do." He called this "satisficing"-accepting "good enough" solutions rather than optimal ones. Instead of constantly processing new information and updating choices, people adopt rules, routines and habits that allow them to ignore information entirely.
Two factors determine whether we actively decide or default to rules: the consequences of the decision and the cost of information. When consequences are limited, deliberation isn't worth the cognitive tax. For high-consequence decisions like choosing a life partner, we willingly invest time in deliberation. Similarly, if information is costly to obtain, we often default to rules rather than making case-by-case decisions.
AI prediction provides better information for decision-making. With reliable weather forecasts, you might abandon your umbrella rule and make daily decisions based on accurate predictions, avoiding both the cost of getting wet and carrying an unnecessary umbrella.
Breaking rules through "forced experimentation" can reveal hidden value. During COVID-19, many discovered unexpected productivity benefits of working from home. Similarly, during a 2014 London tube strike, 5% of commuters permanently changed their routes after discovering more efficient alternatives, saving an average of six minutes daily-a 20% time reduction.
Capítulo 5
Uncertainty Hiding in Plain Sight
Economist George Stigler once said, "If you never miss the plane, you're spending too much time in airports." Yet modern airports have transformed into destinations with spas, casinos, and shopping. People now spend an hour longer at airports than a decade ago, not because they want to, but because uncertainty in travel has increased.
The wealthy who fly privately experience an entirely different airport universe. Private terminals are spartan because there's no waiting-the plane leaves when the passengers arrive, not according to a fixed schedule. Without uncertainty, there's no need to invest in making waiting pleasant.
AI could transform airports by addressing traffic and security uncertainties. Navigation apps already estimate travel times, and future versions could account for actual flight departures rather than scheduled times. AI could also predict security line wait times, allowing travelers to make informed decisions about when to leave for the airport rather than relying on conservative rules.
Rules arise because it's costly to embrace uncertainty, but they create their own problems. The Shirky Principle states that "institutions will try to preserve the problem to which they are the solution." Businesses that profit from waiting passengers have little incentive to eliminate wait times. Finding AI opportunities requires looking beyond the guardrails protecting rules from uncertainty.
Consider greenhouse farming, which reduces weather uncertainty but creates new problems-particularly pests that thrive in controlled environments. Companies like Ecoation use AI to predict pest infestations, allowing farmers to deploy the right pest-control tools at the right time. Beyond cost savings, AI prediction could enable system-level changes: farmers could grow pest-sensitive crops, build larger greenhouses, and implement alternative energy-saving strategies.
Capítulo 6
Rules as System Glue
Checklists and standard operating procedures are essential for complex systems but represent rules to follow rather than decisions to make. Surgeon Atul Gawande's "The Checklist Manifesto" argues that even highly skilled specialists need checklists in complex environments, citing how Boeing's Model 299 bomber succeeded after implementing a checklist system. These procedures ensure reliability and reduce error, but each represents hidden uncertainty that could potentially be addressed with AI prediction.
Rules treat everyone the same despite fundamental differences between people. Marketers traditionally segment populations due to lack of information, but with better data, they can provide personalized products and services. Radio stations broadcast the same songs to all listeners, while streaming services create personalized playlists. Pandora researchers discovered they could use AI to predict how much different users disliked ads, allowing them to personalize ad frequency rather than following a uniform rule.
Education is filled with rules that create uniformity, from seating arrangements to homework policies. While these rules provide structure, they can also stifle personalization. John Stuart Mill warned that state education could mold people to be "exactly like one another." Though education standards acknowledge the need for personalized learning, implementation remains challenging. AI offers a solution, as demonstrated by economists Jin and Sun who used AI to personalize entrepreneurship training for e-commerce sellers. The AI analyzed sellers' operations and recommended specific training modules, increasing revenue by 6.6 percent.
Long-standing rules become embedded in systems so deeply that everything must move simultaneously for change to occur. Consider how television advertising rules dictate program length, writing structure, and commercial breaks. YouTube's AI-driven system allows content of any length and matches viewers with content and ads they'll likely enjoy-a capability far more valuable in YouTube's flexible system than in rigid network television.
Capítulo 7
Oiled Systems vs. Glued Systems
AI could have helped address the Covid-19 pandemic, but many countries relied on rules-based public health procedures rather than decision-making approaches. The core prediction problem wasn't primarily about health for most people-it was about not knowing who was infectious. By January 2021, only 9 million Americans had Covid-19, while 320 million were affected by restrictions despite not being sick.
If we had known who was infectious, we could have kept only them away from others while allowing normal life to continue for everyone else. Various prediction tools emerged-from AI-powered cough detection to trained dogs in Thailand-but rapid antigen tests proved most efficient. These tests weren't perfect but could reliably identify infectious individuals.
The authors partnered with twelve major companies employing over half a million workers to create the CDL Rapid Screening Consortium. These companies committed executive resources to remove rule-based barriers and implement workplace testing. Beginning with a Toronto pilot in January 2021, the system successfully identified infected individuals before they entered workplaces, keeping facilities open that might otherwise have closed.
To leverage prediction machines effectively, we must transform rules into decisions-but the system must accommodate this change. If rules are tightly glued together for reliability, inserting decisions may prove futile. The pandemic's "stay home" rule created cascading problems: business closures, unemployment, mental health issues, and disrupted healthcare. Prediction tools like rapid tests could have enabled targeted isolation decisions rather than blanket restrictions, minimizing economic and social costs while maintaining health outcomes.
Capítulo 8
The System Mindset
Every year at Bletchley Park, contestants compete against computers in the Turing test, trying to prove they're "the most human human." Such human-versus-machine competitions fuel anxiety about AI replacing people. An industry has emerged analyzing jobs task-by-task to evaluate AI replacement potential. A 2013 Oxford study claimed nearly half of US jobs are vulnerable to automation, yet a decade into the current AI wave, technological unemployment hasn't materialized.
Stanford professor Tim Bresnahan argues that task-level substitution misses how new technologies drive radical adoption: through systemwide change. Leading tech companies like Amazon and Google haven't simply replaced tasks with AI-they've built completely new systems. This "system mindset" contrasts with the "task mindset" by recognizing that generating real value requires reconstituting entire systems of decisions involving both machine prediction and humans.
While economists tend to focus on how AI reduces prediction costs, the real opportunity lies in new applications enabled by these cost reductions. AI startups initially pitched their products as cost-saving replacements for human labor, but this proved to be a tough sell. The more successful pitches focused on value creation-showing how AI could help businesses generate more profits by providing higher quality products to customers.
The healthcare system's complexity presents significant barriers to AI implementation. AI solutions in healthcare often fail because they provide predictions nobody can use (due to unavailable treatment options) or enable actions nobody can take (due to liability rules) or wants to take (due to misalignment with compensation systems). The challenge isn't that the predictions are inaccurate or actions useless, but that coordinating all moving parts is extraordinarily difficult.
Capítulo 9
AI's Greatest Impact: Transforming Innovation Itself
AlphaFold, an AI that predicts protein structures, has been hailed as "the most important achievement in AI-ever" by Forbes and described as something that "will change everything" by Nature. By predicting protein structures from amino-acid sequences, AlphaFold helps scientists discover new facts about the building blocks of life.
AI's greatest potential for economic transformation lies in the system of innovation and invention itself. At the 2017 NBER Economics of Artificial Intelligence Conference, economists argued that AI "has the potential to change the innovation process itself." By making data science better, faster, and cheaper, AI enables new types of predictions that open new avenues of inquiry and improve laboratory productivity.
Innovation typically involves a structured process of trial and error. An organization specifies an objective, generates hypotheses, designs experiments, learns from failures, runs successful pilots, and deploys at scale. AI could transform this process by generating thousands of possible solutions from existing data.
For example, an innovation-focused AI could generate thousands of possible recommendation engines rather than manually hypothesizing about the best type. This would enable innovation on more impactful measures than short-term purchasing, such as customer churn or long-term sales. With better hypotheses, more experiments could be run with higher yield and greater return on investment. If predictions become good enough, it might even be possible to skip experiment or pilot phases entirely.
By changing the innovation process itself, AI could ultimately have a greater impact than all other AI applications combined. However, many industries haven't yet recognized the need for change. Large companies rarely find it worthwhile to transform their industry's operations, especially when currently profitable. The risk of getting it wrong is too high.
Capítulo 10
Power Shifts and AI Disruption
The adoption of AI in the context of system-wide change is aptly described as "disruption" for three reasons: opportunities for AI application can be hidden, creating blind spots for existing industries; the challenges of replacing old systems with new ones are part of creative destruction accompanying transformational change; and as old systems are displaced, economic power shifts, making power accumulation both the reward for system innovation and something to be feared and resisted.
When AI drives organizational redesign, existing organizations face a critical challenge: their current structures are optimized for existing technologies, not for AI-driven systems that may require greater modularity or coordination. The real difficulty comes from power shifts within organizations. Those who expect to lose power in the resulting reallocation will actively resist change. The Blockbuster case exemplifies this problem-management understood the need to adopt a Netflix-like subscription model without late fees, but franchisees (who earned 40% of revenue from late fees) successfully resisted this change, ultimately leading to the company's demise.
Despite alarming headlines suggesting robots automatically fire workers without human oversight, the reality is more nuanced. Machines do not decide anything and therefore do not have power-humans remain the decision-makers. While AI cannot transfer decisions to machines, it can change which humans make decisions, potentially shifting power dynamics and necessitating system changes.
System-level innovation creates first-mover advantages with AI because the technology learns from data, creating a powerful feedback loop. The sooner AI is deployed, the sooner it learns, improving prediction accuracy and system effectiveness, which attracts more users, generating more data. This explains why venture capitalists aggressively invest in early-stage AI projects-they're betting on this learning flywheel effect.
Capítulo 11
The Great Decoupling of Prediction and Judgment
AI fundamentally separates prediction from judgment in decision-making processes, a distinction that was previously blurred in human reasoning. This decoupling is powerfully illustrated through Michael Jordan's 1986 injury decision - team doctors predicted a 10% chance of career-ending reinjury if he returned to play. While the prediction was clear, Jordan and Bulls owner Jerry Reinsdorf interpreted this same probability differently. Jordan valued potential achievement over safety, while Reinsdorf prioritized protecting his star asset. Jordan ultimately played with careful restrictions, defying the odds to set an NBA playoff scoring record with 63 points against the Celtics - a decision that highlighted how identical predictions can lead to different choices based on personal values.
AI prediction technology forces judgment to become explicit rather than remaining implicit in our decision processes. Consider how firefighters traditionally make split-second emergency decisions - they unconsciously combine probability assessments (likelihood of building collapse) with value judgments (potential lives saved versus risk to crew). AI transforms this by handling the prediction component separately, requiring humans to consciously and explicitly apply their judgment. This separation is revolutionizing industries like insurance, where companies now use sophisticated telematics systems to price premiums based on detailed driving behavior data. The AI predicts accident risk with remarkable precision, while customers must explicitly judge whether certain driving behaviors are worth the premium adjustments - a decision that weighs convenience against cost.
The critical distinction between bad decisions and bad outcomes becomes especially important when working with probabilistic AI systems. Annie Duke, professional poker player and author of "Thinking in Bets," emphasizes that we must learn to distinguish between unfortunate luck and poor decision-making. Amateur poker players frequently fall into the trap of "resulting" - modifying their strategy based on immediate outcomes rather than the quality of their decision process. For instance, a player might abandon a mathematically sound strategy after losing several hands, despite the strategy being correct in the long run. This mirrors Jordan's injury situation - while his return to play worked out brilliantly, the positive outcome alone doesn't prove it was the right decision at the time.
Embracing uncertainty in the AI era means shifting from seeking absolute certainty to making decisions based on probability thresholds and risk tolerance. The refugee adjudication process provides a compelling example - currently, adjudicators make binary decisions about complex cases despite having limited evidence and almost no feedback on the accuracy of their past decisions. An AI prediction system could transform this process by providing detailed probability assessments of claim legitimacy, making uncertainty explicit and quantifiable. This would allow adjudicators to set appropriate thresholds based on policy goals and humanitarian considerations, while reducing the psychological burden of false certainty. Some jurisdictions are already experimenting with risk assessment tools that provide probability scores rather than binary recommendations, allowing human judges to apply their judgment more transparently.
Capítulo 12
Building Reliable AI Systems
When AI transforms one decision in a system, it affects how all decisions coordinate. This challenge centers on reliability. Using Thomas Schelling's coordination problem as illustration, shared expectations enable coordination without constant communication. But when AI improves prediction in one area, it can create unpredictability elsewhere-what the authors call the "AI bullwhip."
When a restaurant uses AI for demand forecasting, it orders precisely what it needs-sometimes 30 pounds of avocados, sometimes 300 pounds. This variability, while optimal for the restaurant, creates unpredictability for suppliers who previously received consistent 100-pound orders. As this effect cascades through the supply chain to distributors and growers, small changes at the restaurant level create massive fluctuations upstream.
To counter the AI bullwhip, one approach is better coordination. Like a coxswain in an eight-rower team who monitors and adjusts strategy during a race, coordination systems ensure that AI-driven decisions remain synchronized across an organization. When coordination isn't possible, modularity offers an alternative approach. By building walls around AI-driven decisions, organizations can prevent misalignment with other decisions.
Team New Zealand's America's Cup victory demonstrates how AI can revolutionize system design. While sailing simulators had long been used for boat design, the team partnered with McKinsey to overcome their innovation bottleneck: human sailors who limited simulation speed. They developed AI sailors that could run hundreds of simulations in the time humans could run just a handful. After eight weeks, these AI sailors began outperforming humans and teaching them new techniques.
Digital twins-virtual representations of physical systems-provide risk-free environments for innovation and system-level simulation. When combined with AI, they enable organizations to design entirely new operational approaches and test coordination between decisions without costly physical implementation.
