Blog · Review Essay · May 2026

Prediction Machines and the Price of Automated Judgment

Ajay Agrawal, Joshua Gans, and Avi Goldfarb make artificial intelligence look less mystical by calling it cheap prediction. That move is useful. It is also incomplete unless the reader keeps asking what happens when prediction enters institutions that already know how to turn scores into authority.

The Book

Prediction Machines: The Simple Economics of Artificial Intelligence was written by Ajay Agrawal, Joshua Gans, and Avi Goldfarb and first published by Harvard Business Review Press in 2018. Harvard Business Review Press lists an updated and expanded edition published on November 15, 2022, with 304 pages.

The authors are economists associated with the University of Toronto's Rotman School of Management and the Creative Destruction Lab. Their argument grew out of a 2016 Harvard Business Review article that treated machine intelligence through a familiar economic lens: when a technology makes something cheaper, people use more of it and reorganize other activities around it.

The book's plainest claim is that AI lowers the cost of prediction. That means it is not only about robots, consciousness, or general intelligence. It is about making guesses cheaper: demand forecasts, fraud risk, medical images, loan default, hiring fit, text completion, translation, logistics, targeting, routing, classification, and recommendation.

Cheap Prediction

The frame works because it cuts through spectacle. A system does not need to be wise to predict which customer is likely to churn, which warehouse will run short, which applicant looks similar to past hires, which post will hold attention, or which word probably comes next. Much of AI's practical power comes from turning uncertainty into an input that organizations can price, compare, and automate.

This is why the book remains useful after the generative-AI wave. Large language models do more than narrow forecasting, but they still operate as prediction engines in a broad sense: they infer continuations, likely answers, useful actions, plausible summaries, probable code, probable intent, and probable next steps from context. The visible magic is conversation; the institutional use case is often reduced uncertainty at scale.

The important consequence is substitution. When prediction gets cheaper, complementary goods become more valuable. Data, judgment, action, workflow design, liability management, human override, and institutional trust become the scarce pieces. The prediction is only one part of the decision system.

Judgment Does Not Disappear

The book's best distinction is between prediction and judgment. Prediction estimates what is likely. Judgment decides what matters, what tradeoffs are acceptable, which errors are tolerable, and who bears the cost when the system is wrong.

That distinction should be central to AI governance. A model may estimate that a patient is low risk, an applicant is unlikely to repay, a worker is underperforming, a student essay is machine-written, a neighborhood deserves more patrols, or a user is vulnerable to a certain message. None of those predictions settles the moral question. The institution still has to decide whether the predicted event is the right target, whether the data are legitimate, whether the action is fair, and whether the person classified by the system can understand or contest the result.

Cheap prediction can therefore make judgment more important while making judgment easier to hide. The decision may be presented as technical because a model supplied a score. But the score only becomes power when a hospital, school, employer, court, platform, insurer, or agency decides what to do with it.

The Workflow Is the Politics

Prediction Machines is a management book, so it repeatedly asks readers to decompose decisions and redesign workflows. That is more politically charged than it sounds. A workflow encodes who sees the evidence, who can object, who has discretion, who is monitored, who performs repair, who gets blamed, and which part of the system is allowed to remain opaque.

AI rarely enters a neutral workplace. It enters a call center with time metrics, a school with scarce staff, a hospital with liability pressure, a warehouse with productivity tracking, a welfare office with eligibility rules, a newsroom with traffic incentives, or a platform with engagement goals. In each case, prediction amplifies the surrounding institution's theory of the person.

That is the bridge to labor. The book's language of job redesign is clear and practical, but the social outcome depends on bargaining power. Prediction can augment workers by giving them better information. It can also deskill them, turn them into exception handlers, compress tacit judgment into prompts and checklists, or make their every action legible to management.

The AI-Age Reading

Read in 2026, the book is strongest as a discipline against AI mysticism. It asks a useful first question: what uncertainty is this system making cheaper? If the answer is unclear, the deployment may be theater. If the answer is clear, the next question is sharper: who gains from reducing that uncertainty?

A hiring system reduces employer uncertainty about applicants while increasing applicant uncertainty about evaluation. A predictive policing system reduces institutional uncertainty about patrol allocation while increasing public exposure to suspicion. A recommender reduces platform uncertainty about attention while increasing user exposure to behavioral shaping. An AI companion reduces user uncertainty about response while increasing dependence on a private system's memory, incentives, and constraints.

The book's economics helps name the mechanism. The site's recurring concern is what happens afterward: prediction becomes a reality loop when the system acts on the world, records the response, retrains on the changed world, and then treats its own effects as new evidence.

Where the Frame Strains

The cheap-prediction frame is clarifying, but it can understate the cultural and institutional force of AI. Generative systems do not only predict demand or classify risk. They produce language, images, code, synthetic evidence, social presence, managerial instructions, educational feedback, and official explanations. Their outputs can become the environment in which people form beliefs and make choices.

The frame can also make adoption sound too smooth. Economic complements do not automatically appear in humane form. Judgment may be scarce, but institutions can respond by automating around it, centralizing it, outsourcing it, or pretending it has been embedded in policy. Human review can become a checkbox. Appeals can become scripts. Accountability can become a dashboard.

Finally, the book is mostly written for managers and policymakers who want to act. That is its strength and its bias. The people classified by prediction systems often do not get to redesign the workflow. They meet the system as a score, denial, recommendation, route, ranking, shift assignment, or chatbot answer.

The Site Reading

The most useful lesson is that prediction is not knowledge by itself. It is a lever. Once attached to action, it can create the world it claims merely to anticipate. A fraud model changes behavior. A ranking model changes culture. A hiring model changes careers. A feed model changes attention. A companion model changes self-disclosure. A procurement model changes public capacity.

That means AI governance has to inspect the whole decision chain: data provenance, model output, threshold setting, institutional incentives, human discretion, appeal rights, audit records, error distribution, labor effects, and feedback loops. The question is not whether prediction is accurate in the abstract. The question is whether the prediction system is answerable to the people it changes.

Prediction Machines belongs in the catalog because it offers a clean economic grammar for AI without treating AI as destiny. Its best use is as a first pass: demystify the machine, then refuse to let demystification become permission. Cheap prediction still needs expensive judgment.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Return to Blog · Return to Books