Most business books about artificial intelligence are useless because they treat it as a magic ingredient that transforms things, without ever saying what it does. This one is useful because it answers that question in a single sentence and then follows the implications honestly. The claim is that recent AI is, economically, a sharp fall in the cost of prediction. Prediction meaning the act of filling in missing information from information you have.
Not forecasting in the narrow business sense, but the general operation of going from what you know to what you do not. Once you accept that framing, a lot of things fall into place, because economists know what happens when the cost of an input drops sharply. You use more of it. You use it for things that were never worth doing before.
Things that are substitutes for it fall in value and things that are complements to it rise. The complement argument is the most valuable thing in the book. If machines get very good and very cheap at prediction, then the scarce, valuable human contribution is judgement, meaning the ability to decide what a good outcome is and what a mistake is worth. A model can tell you the probability that this customer will not repay.
Deciding what to do about that, given the cost of a bad loan, the value of a good customer, and what your regulator and your reputation will tolerate, is judgement, and it does not get cheaper when prediction does. That distinction is a genuinely good tool for looking at a workflow and asking which parts of it are prediction and which parts are judgement, and it explains a lot of failed automation projects, which usually turn out to be attempts to automate judgement with a prediction machine. The economic discipline shows throughout. These are academics who reason in terms of tradeoffs and complements rather than in terms of transformation and disruption, and the writing is noticeably more careful than the genre standard.
There is also a good treatment of when the value of better prediction is high, which is when the decision changes as a result, and how often that condition fails. A more accurate forecast that does not change anything you do is worth nothing, and a great many analytics projects have foundered on exactly that. Now the criticisms. It is one idea.
A very good idea, established clearly in the opening chapters, and then applied to workflows, to job design, to organisational structure, to strategy and to society, with the same argument re-explained at each stop. By the halfway mark you are ahead of the book, and the last third can be read at speed without losing anything. This is a common failing of business books built out of a paper and it is present here. The 2018 publication date is the real problem.
The book was written when applied machine learning meant classification and forecasting on structured data, and the examples reflect that. Everything that has happened since, meaning generative systems, language models, coding assistants and agents, is absent. To the authors' credit the framing partly survives the transition, since predicting the next token is prediction in exactly their sense, and the book cannot help you think about the things people currently want help thinking about, and their follow up volume only partly closes the gap. The case studies have aged in the usual way.
Companies held up as exemplars have since had mixed fortunes, and applications presented as imminent either arrived differently or did not arrive. Anyone reading a business book from 2018 should expect this and it still undermines the concrete sections. And there is an assumption running underneath that prediction problems come with clean feedback. You predict, you observe the outcome, you improve.
Enormous numbers of real decisions do not work like that. You reject the loan applicant and never learn whether they would have repaid. You choose one strategy and never observe the other branch. The book's economics are cleanest where feedback is clean, and the messy majority gets less attention than it deserves.
My three point seven is for a genuinely clarifying idea, argued with more rigour than this genre normally attempts, dragged down by repetition, by a publication date that puts it a whole technological era behind, and by an assumption about feedback that does not hold in most of the decisions its readers actually face. Read the first hundred pages properly. Skim the rest with a clear conscience.