What the book argues

The foundational economics: when something becomes cheap, we use more of it - and we discover prediction hiding inside tasks never framed that way (translation, driving, diagnosis, hiring are all prediction problems once data is sufficient). Simultaneously, the complements to prediction appreciate: data (the new input with decreasing-returns caveats honestly treated), judgment (deciding what to predict and what outcomes are worth - irreducibly human in their framing), and action. The first managerial consequence is decomposition: break any decision into prediction, judgment, data, and action, and ask which parts the cheap new input transforms.

The tools-and-workflow chapters operationalize it: the AI canvas forces teams to specify, for each candidate deployment, the prediction (exactly what is being predicted), judgment (how outcomes are valued, including error asymmetries), action, outcome, and the three data types (training, input, feedback). Workflows then get re-engineered rather than sprinkled: some tasks absorb AI as a tool, others decompose and recombine - their recurring examples (from radiology's re-bundling to fulfillment's anticipatory shipping) show jobs as task-bundles whose composition shifts with prediction costs.

The strategy chapters contain the book's sharpest idea: thresholds. Improving prediction quality changes business models discontinuously - Amazon's hypothetical flip from shop-then-ship to ship-then-shop once return-rate predictions clear a bar - so executives should ask where their industry's thresholds sit and who reaches them first. Trade-offs close it honestly: more AI means data-feedback loops that advantage incumbents, new risks (bias at scale, security of input data, homogenization), and the C-suite reallocation of what remains scarce - judgment about what the organization should value. The 2022 update folds in the generative wave without changing the frame: generation, too, is prediction - of the next token - and the economics hold.