Prediction Machines: The Simple Economics of Artificial Intelligence

Ajay Agrawal, Joshua Gans, Avi Goldfarb — 2018

Category: Tecnología, Datos & IA

Prediction Machines (2018) presents a simple yet powerful idea: AI lowers the cost of prediction. When the price of something drops drastically, it changes what we do, who does it, and how the economy is organized. The authors—management economists—provide a framework for translating AI into business decisions: where it creates value, when to automate, what to complement with human judgment, and how to redesign processes.

What is it about exactly? The book breaks down a decision into components: Data → 2) Prediction (what AI does best when there is historical data) → 3) Judgment (defining the “payoff”: what a good decision means) → 4) Action → 5) Results → 6) Feedback (system improvement). The central thesis: if prediction becomes cheap, the bottleneck shifts to quality data, judgment, actions, and organizational design. Key Ideas (in brief) AI = cheap prediction: classify, estimate, fill gaps. If you can frame a task as prediction, AI will likely make it cheaper. Judgment as a complement: AI does not define objectives or costs of error: that is judgment (ethics, regulation, brand). The cheaper the prediction, the more valuable the judgment. Quality and ownership of data: not all “quantity” is useful: representativeness, bias, and rights matter. Real “moats” combine data + processes + relationships. Automate vs. augment: automate when the cost of error and variability are low and the feedback is clear. Augment (AI as co-pilot) when the error is costly or ambiguous. Process redesign: improvement does not come from “tacking AI on at the end”: you need to distribute work between prediction, judgment, and action; perhaps change who decides and when. Threshold economics: many decisions are “yes/no” with a threshold (minimum acceptable probability). Choosing that threshold is strategy (trade-off precision/recall, risk/income). Value shifts: when prediction becomes cheaper, businesses emerge in orchestrating data, designing actions, and ensuring/regulating risk. Policy and responsibility: bias, explainability, privacy, and employment become design decisions, not externalities. What you will learn from “Prediction Machines” To reframe problems as concrete predictions (what variable is missing that, if we knew it, would lead to better decisions?). To separate prediction from judgment and action, to know what to automate and what to keep human. To choose thresholds and metrics (precision, recall, cost of error) aligned with business and regulation. To build feedback loops that make AI increasingly useful. To decide make/buy/ally: proprietary models vs. APIs, and where your real “moat” lies. How to apply the book in iGaming 1) Reframe decisions as predictions KYC/Verification: What is the probability that this document is legitimate? Fraud and chargeback: Probability of a disputed transaction? Churn/retention: Probability that an active player will leave in 14/30 days? Risk of harm (RG): Probability of an emerging problematic gaming pattern? Customer service: Probability that this ticket will escalate if not taken by a senior agent in 10 min? 2) Define judgment (cost of error and thresholds) False positives (blocking someone you shouldn't) vs. false negatives (letting risk pass). For RG and fraud, the threshold should prioritize protection; for churn, you might prefer higher recall (risk “useless” incentives rather than losing the right customer). 3) Choose to automate or to augment Automate: obvious duplicate detection, initial document scoring, ticket prioritization. Augment: decisions with high reputational or regulatory costs (account closures due to RG, complex disputes): AI suggests and humans decide with traceability. 4) Data and governance Data catalog with lineage and permissions by country; elimination of sampling biases (e.g., underrepresentation of certain payment methods). Drift and fairness dashboard: monitor when the model degrades or discriminates. Privacy by design: minimization, limited retention, anonymization where applicable. 5) Operational integration Complete cycle: data → model → recommendation → action in systems (PSP, CRM, limits) → labeled feedback. Useful explainability: reasons for the recommendation for agents (main features, evidence) and for auditing. 6) Business metrics (not just AUC) Time to first withdrawal, CES/NPS post-withdrawal, % with active limits, chargebacks/1,000, retention by cohort. Connect each model to 1–2 KPIs of the journey; if it doesn't move the needle, rethink it. Minimum toolkit: AI case template: variable to predict, threshold, cost of error, who decides, automatic/human action, available feedback. Decision map by journey (KYC, deposits, withdrawals, CRM, RG) indicating what is prediction and what is judgment. Threshold guide by country/risk with examples of trade-offs. Data contract: sources, rights, quality, biases, expiration. Model health dashboard: performance, drift, fairness, impact on business KPI. Incident runbook: what to do if the model fails or causes harm (rollback, communication, corrections). “Prediction Machines” turns AI into management: it teaches you to think in terms of costs, thresholds, complements, and process redesign. The promise is not magic, it is applied economics: if prediction is cheap, the value lies in choosing wisely what to predict, how to decide, and how to act—with metrics, ethics, and continuous learning.