Visual summary of operating lessons from Avi Goldfarb.

Lessons from Avi Goldfarb

Avi Goldfarb is an economist at the University of Toronto’s Rotman School of Management who studies how digital technology changes markets. In his co-authored books Prediction Machines and Power and Prediction, he reframes artificial intelligence from a vague concept into a measurable drop in the cost of prediction. This profile collects his arguments for how cheaper prediction changes the value of human judgment, forces companies to restructure, and affects data privacy.

Part 1: The Economics of Artificial Intelligence

  1. On reframing AI: "Thinking about AI as a drop in the cost of prediction is transformational because it shifts the conversation from science fiction to fundamental economics." — Source: [Talks at Google]
  2. On economic principles: "When a core input becomes cheap, we start using it for everything. Just as the internet made distribution cheap, AI makes prediction cheap." — Source: [Invest Like the Best]
  3. On defining intelligence: "From an economic standpoint, we shouldn't view AI as an artificial brain. We should view it as a statistical tool that fills in missing information." — Source: [Schwartz Reisman Institute]
  4. On the history of technology: "We tend to look at AI as unprecedented, but economically, it behaves like electricity or semiconductors—it requires complementary innovations to be useful." — Source: [Rotman Insights]
  5. On the limits of automation: "Economics tells us that when prediction becomes cheap, the complements to prediction—like human judgment—become more valuable." — Source: [Harvard Business Review]
  6. On decoupling tasks: "AI allows us to unbundle the process of making a decision, separating the prediction of an outcome from the judgment of its value." — Source: [The Mixtape with Scott]
  7. On productivity growth: "We often worry about AI destroying jobs, but the primary economic function of a general-purpose technology is driving long-term productivity growth." — Source: [NBER Working Papers]
  8. On cost curves: "The history of computing is a story of falling arithmetic costs. The history of AI will be a story of falling prediction costs." — Source: [Prediction Machines]
  9. On business strategy: "You don't need a computer science degree to understand AI strategy; you just need to understand what happens to your business when prediction costs fall to zero." — Source: [Rotman Management Magazine]
  10. On identifying opportunities: "Look for processes where the lack of accurate prediction is the primary bottleneck to efficiency. That is where the economic returns of AI will be highest." — Source: [Power and Prediction]

Part 2: AI as Cheap Prediction

  1. On the definition of prediction: "Prediction isn't just about forecasting the future. It is about using information you have to generate information you don't have." — Source: [Harvard Business Review]
  2. On medical diagnosis: "A medical diagnosis is fundamentally a prediction problem: taking data about symptoms and using it to fill in the missing information about the underlying cause." — Source: [Prediction Machines]
  3. On driving: "Autonomous vehicles do not program every possible rule of the road. They predict what a good human driver would do in any given situation." — Source: [Talks at Google]
  4. On translation: "Machine translation is not about understanding language; it is about predicting which word in English most likely corresponds to a word in French based on past texts." — Source: [Schwartz Reisman Institute]
  5. On the expanding scope of prediction: "As prediction gets cheaper, we will start treating problems as prediction problems that we never previously thought of as such." — Source: [Invest Like the Best]
  6. On data as an input: "Data is the raw material of prediction. Without vast amounts of historical data, the cost of generating accurate predictions remains high." — Source: [Rotman Insights]
  7. On the limits of machines: "Machines are terrible at knowing what to do with a prediction once it is made. They only output a probability." — Source: [The Mixtape with Scott]
  8. On resolving uncertainty: "The primary value of cheap prediction is that it reduces the uncertainty that prevents organizations from acting decisively." — Source: [Prediction Machines]
  9. On inventory management: "Retailers used to rely on rules of thumb to stock shelves. Now, they use cheap prediction to stock precisely what consumers will buy, reducing massive overhead." — Source: [Harvard Business Review]
  10. On creative industries: "Even creative work involves prediction—predicting which sequence of pixels will form a pleasing image or which sequence of words will form a coherent paragraph." — Source: [Talks at Google]

Part 3: The Rising Value of Human Judgment

  1. On defining judgment: "Judgment is the process of determining what the reward or penalty is for taking a specific action in response to a prediction." — Source: [Prediction Machines]
  2. On human comparative advantage: "As machines take over the math of predicting outcomes, humans are left to do what machines cannot: decide which outcomes actually matter." — Source: [Invest Like the Best]
  3. On ethical trade-offs: "If an AI predicts a 10% chance of rain, it cannot decide if you should carry an umbrella. Only a human knows how much they dislike getting wet versus the annoyance of carrying the umbrella." — Source: [Rotman Insights]
  4. On managerial roles: "The manager of the future is not someone who calculates probabilities, but someone who exercises judgment over complex, ambiguous organizational priorities." — Source: [Harvard Business Review]
  5. On the complementarity of AI: "Prediction and judgment are complements. The cheaper prediction becomes, the more frequently we will require human judgment to act upon those predictions." — Source: [The Mixtape with Scott]
  6. On automation anxiety: "Jobs that are heavily weighted toward prediction are at risk. Jobs that are heavily weighted toward judgment will see their wages and demand increase." — Source: [Talks at Google]
  7. On rare events: "Humans possess the unique ability to exercise judgment in novel situations where historical data is sparse—something prediction machines fundamentally struggle with." — Source: [Schwartz Reisman Institute]
  8. On engineering judgment: "The hardest part of implementing AI is often not the algorithm, but explicitly programming the human judgment of payoffs into the system's objective function." — Source: [Power and Prediction]
  9. On objective functions: "When you tell an AI to optimize a metric, it will do exactly that, ruthlessly. Human judgment is required to ensure the chosen metric actually aligns with the firm's true goals." — Source: [Rotman Management Magazine]
  10. On empathy: "Judgment often requires empathy—understanding how a decision will affect employees or customers emotionally. Machines have no mechanism for this." — Source: [Prediction Machines]

Part 4: The Between Times and Adoption

  1. On the adoption gap: "We are currently in the 'Between Times'—the period where the potential of AI is obvious, but the widespread economic impact has not yet materialized." — Source: [Power and Prediction]
  2. On historical parallels: "When electricity was introduced in factories, productivity didn't jump immediately. It took thirty years for managers to realize they needed to redesign the factory floor." — Source: [Invest Like the Best]
  3. On point solutions: "Most organizations are stuck implementing AI as point solutions, dropping new technology into existing workflows to do the exact same work, just slightly faster." — Source: [Schwartz Reisman Institute]
  4. On organizational friction: "The barrier to seeing massive returns from AI is not the technology itself; it is the friction of internal resistance and legacy organizational structures." — Source: [Talks at Google]
  5. On the necessity of patience: "Transformational technologies always require complementary innovations. The Between Times last exactly as long as it takes for an industry to invent those complements." — Source: [Rotman Insights]
  6. On misplaced expectations: "Companies often buy AI expecting an immediate ROI, failing to realize they are buying a capability that requires them to change how their business operates." — Source: [Harvard Business Review]
  7. On legacy systems: "You cannot simply plug a highly accurate prediction machine into a bureaucratic process designed around human guesswork and expect a revolution." — Source: [The Mixtape with Scott]
  8. On first-mover advantage: "The winners in the Between Times are not necessarily the companies with the best algorithms, but those most willing to endure the pain of reorganizing around the algorithm." — Source: [Power and Prediction]
  9. On the length of the transition: "We often overestimate the short-term impact of AI while severely underestimating the long-term, system-level transformation it will enforce." — Source: [Prediction Machines]

Part 5: System-Level Innovation

  1. On redefining the business: "System-level innovation means asking not 'how can AI make our current process faster?' but 'if prediction were perfect, why would we do this process at all?'" — Source: [Power and Prediction]
  2. On airport security: "A point solution uses AI to read luggage X-rays faster. A system solution predicts who is a threat before they arrive, eliminating the security line entirely." — Source: [Invest Like the Best]
  3. On shipping: "If Amazon's recommendation AI becomes accurate enough, it will be cheaper for them to ship goods to your house before you order them and handle the occasional return." — Source: [Talks at Google]
  4. On creating winners and losers: "System-level changes are difficult because they alter the power dynamics inside a firm. Middle managers who used to make forecasts lose status to algorithms." — Source: [Schwartz Reisman Institute]
  5. On architectural changes: "To get the most out of AI, you have to change the architecture of the organization. You have to move decision rights to the people who can best exercise judgment." — Source: [Harvard Business Review]
  6. On the role of the CEO: "Because system solutions require cross-departmental reorganization, they cannot be delegated to the IT department. They must be driven by the CEO." — Source: [Rotman Insights]
  7. On healthcare systems: "AI won't just help doctors read scans; it will eventually redesign the hospital, shifting care from reactive treatment to proactive prevention based on prediction." — Source: [The Mixtape with Scott]
  8. On risk tolerance: "Pursuing system-level innovation requires a high tolerance for risk because you are dismantling workflows that currently generate revenue for workflows that are untested." — Source: [Power and Prediction]
  9. On the inevitability of change: "Eventually, new entrants unburdened by legacy processes will build AI-native system solutions, forcing incumbents to adapt or die." — Source: [Prediction Machines]

Part 6: Decision-Making Under Uncertainty

  1. On the anatomy of a decision: "Every decision can be broken down into data, prediction, judgment, action, and outcomes. AI only handles the prediction component." — Source: [Prediction Machines]
  2. On managing uncertainty: "The primary economic cost of uncertainty is that it forces organizations to build expensive buffers—extra inventory, extra time, extra staff." — Source: [Harvard Business Review]
  3. On reducing buffers: "When prediction improves, uncertainty drops. When uncertainty drops, companies can strip away the costly buffers they built to protect against the unknown." — Source: [Invest Like the Best]
  4. On default actions: "Organizations rely on rules and defaults when prediction is expensive. AI allows firms to replace rigid rules with dynamic, context-specific decisions." — Source: [Schwartz Reisman Institute]
  5. On the limits of past data: "A prediction machine assumes the future will look like the past. When structural breaks occur—like a pandemic—these machines fail and human intervention is required." — Source: [Rotman Insights]
  6. On the illusion of certainty: "No matter how good AI gets, it produces probabilities, not certainties. Leaders must still learn how to act confidently on probabilistic information." — Source: [The Mixtape with Scott]
  7. On asymmetric payoffs: "AI is incredibly useful in situations where the cost of a false positive is drastically different from the cost of a false negative, provided human judgment sets those costs." — Source: [Talks at Google]
  8. On real-time decision making: "By lowering the cost of prediction, AI increases the frequency at which decisions can be made, shifting strategy from annual planning to real-time adjustment." — Source: [Power and Prediction]
  9. On the paradox of better data: "Having better predictions often makes the remaining uncertainty more glaring, forcing executives to confront the limitations of their own judgment." — Source: [Rotman Management Magazine]

Part 7: The Privacy Paradox and Regulation

  1. On the privacy paradox: "Consumers consistently say they value privacy highly in surveys, yet they willingly trade their personal data for tiny conveniences or small discounts." — Source: [Journal of Political Economy]
  2. On the cost of privacy: "Privacy regulation is not free. When you restrict data flow, you inevitably decrease the efficiency of digital advertising and online markets." — Source: [NBER Working Papers]
  3. On advertising effectiveness: "Empirical evidence shows that when strict privacy regulations prevent behavioral targeting, the effectiveness of digital ads drops precipitously, especially for smaller publishers." — Source: [Management Science]
  4. On entrenching incumbents: "Complex privacy laws like GDPR often inadvertently benefit large tech monopolies, because only they have the resources to comply and the massive first-party data to survive without third-party tracking." — Source: [Rotman Insights]
  5. On the trade-off with innovation: "Policymakers must realize that aggressively restricting data access to protect privacy directly slows the development of accurate prediction machines." — Source: [Harvard Business Review]
  6. On contextual sharing: "People do not value privacy uniformly. They are highly protective of medical data but largely indifferent to sharing their geographic location with a weather app." — Source: [The Mixtape with Scott]
  7. On consent fatigue: "Forcing users to click 'I agree' on endless cookie banners does not necessarily protect privacy; it mostly increases transaction costs and induces consumer apathy." — Source: [Invest Like the Best]
  8. On data as a barrier to entry: "In a world heavily regulated for privacy, a firm's historical stockpile of user data becomes an insurmountable moat against new competitors." — Source: [Schwartz Reisman Institute]
  9. On the future of regulation: "The challenge of the next decade is designing privacy laws that protect consumers from real harms without accidentally outlawing the data-driven systems that run the modern economy." — Source: [Talks at Google]

Part 8: Market Competition and Digital Strategy

  1. On economies of scale: "Prediction machines exhibit massive economies of scale. The algorithm gets better with more data, attracting more users, which generates more data." — Source: [Prediction Machines]
  2. On the value of data vs. algorithms: "In many industries, the algorithms are open-source and commoditized. The true competitive advantage lies in proprietary access to training data." — Source: [Harvard Business Review]
  3. On competitive dynamics: "When prediction is cheap, the basis of competition shifts from who can forecast the market best to who can act on the forecast fastest." — Source: [Power and Prediction]
  4. On substituting capital for labor: "AI allows firms to substitute capital in the form of prediction machines for human labor in tasks that require forecasting and pattern recognition." — Source: [The Mixtape with Scott]
  5. On digital moats: "A strong digital strategy does not just accumulate data; it identifies the specific data that will generate predictions valuable enough to lock in customers." — Source: [Rotman Insights]
  6. On the cold start problem: "The hardest part of building an AI business is getting the initial data to train the model before you have a product good enough to attract users." — Source: [Invest Like the Best]
  7. On liability and risk: "As AI takes over more decisions, companies face new strategic risks regarding liability. If the machine makes a disastrous prediction, who is legally responsible?" — Source: [Schwartz Reisman Institute]
  8. On specialized AI: "Broad, general AI makes headlines, but the most profitable corporate strategies involve highly specialized, vertical AI trained on niche industry data." — Source: [Talks at Google]
  9. On the ultimate commodity: "As prediction becomes universally cheap and widely available, the only true scarcity left in business will be human vision and judgment." — Source: [Prediction Machines]