Book

Prediction Machines

Ajay Agrawal, Joshua Gans & Avi Goldfarb

Three economists' reframing of AI, stripped of the hype: what modern machine learning actually does is make prediction cheap, and once prediction is cheap, the scarce, valuable thing left is judgment.

Prediction Machines Find this on the mind-map

Introduction

Agrawal, Gans, and Goldfarb make an economist's argument, not a technologist's: the breakthrough in modern AI isn't "intelligence" in any broad sense, it's a sharp drop in the cost of prediction: filling in a missing piece of information from related information you already have. Treat AI as a cheap-prediction technology, and a lot of confusing claims about what it will and won't do start making more sense.

Decisions Are Prediction Plus Judgment

Every decision breaks into pieces: prediction (what's likely to happen) and judgment (given that prediction, what outcome do we actually want, and what should we do about it). A loan approval needs a prediction of default risk and a judgment about how much risk the business is willing to accept. As prediction gets cheap and increasingly automated, judgment (the part that requires knowing what you actually value) becomes the scarcer, more valuable half of the equation, not the part that disappears.

Cheap Prediction Changes the Whole Decision, Not Just One Step

It's tempting to think of AI as a faster calculator bolted onto an unchanged process. In practice, when prediction gets cheap enough, the entire decision workflow often gets redesigned around it. A manager, Felix, assumes a new forecasting model just replaces a spreadsheet one of his analysts used to build by hand. But once forecasts are cheap and constant instead of rare and manual, the business starts making decisions (reordering, staffing, pricing) far more frequently than the old process ever allowed, because the cost of getting a new prediction dropped to nearly nothing.

Tasks, Not Jobs

AI typically automates a task inside a job, not the whole job. A radiologist's job includes image review, but also patient conversations, judgment calls on ambiguous cases, and coordination with other doctors. The practical question for a manager isn't "will this job be automated," it's "which specific tasks inside this job just got cheaper to do by prediction, and what does the job look like once those tasks are mostly handled by a model instead of a person."

The Risk of Leaving the Old Workflow in Place

When a business keeps running its old workflow unchanged around a new prediction capability, it ends up with a decision pipeline still organized around a world where prediction was expensive and rare, bolted together with a prediction engine that's now cheap and constant. The waste isn't the AI; it's the unchanged scaffolding around it that never gets redesigned to take advantage of what changed.

Data as the New Moat

Better predictions require better training data, and better training data mostly comes from running the system and observing outcomes. That means the businesses that get prediction into production earliest build a data advantage that's hard for a late mover to close, independent of who has access to the same underlying model architecture.

Putting It Together

The chain runs: cheap prediction changes the economics of a decision, which should trigger a redesign of the whole workflow around that decision, not a one-for-one swap of an old manual step. The business that actually redesigns around cheap prediction, rather than leaving the old workflow bolted onto a new engine, is the one that captures the data advantage that compounds from there.

Prediction Machines mind-map: Prediction Plus Judgment, Cheap Prediction Redesigns the Decision, Tasks, Not Jobs, the Risk of an Unchanged Workflow, and Data as the New Moat, branching from the book's cover

Why It Matters for Leaders

The managers who get the most wrong about AI right now aren't the ones ignoring it, they're the ones bolting it onto a process designed for a world where getting a forecast or a risk score took a person a day. If you just hand that old process a faster number and change nothing else about how often you act on it, you've paid for a new engine and kept the old car.

The useful question before adopting any AI tool isn't "what can this predict." It's "what decision does this prediction feed, and what should we do differently now that getting this answer costs almost nothing." Those are very different conversations, and most teams skip straight to buying the tool without ever having the second one.

It also reframes what you should be worried about losing to automation. The task that gets automated is rarely the whole job. The real planning question is which specific piece of someone's job just got a lot cheaper to do by prediction, and what that frees them up to spend their judgment on instead.

Picture an operations manager, Sana, whose team gets a new model that predicts weekly demand far more accurately than their old manual forecast. If she keeps the same monthly ordering meeting and just plugs the better number into the same old process, she's captured almost none of the actual value. The real move is to ask why ordering only happens monthly in the first place, now that a good forecast is available every day instead of once a quarter. Redesigning the cadence around the cheap prediction, not just swapping in a better number, is where the real gain sits.

Your chat with Kyoshi about the book Prediction Machines

What to do next

Reinforce your understanding of Prediction Machines

Books related to Prediction Machines

Skills related to Prediction Machines

Frameworks related to Prediction Machines