The question this book answers is the one that most organisations trying to adopt AI are actually stuck on, and almost nobody addresses it honestly. You run a pilot. The model works. The accuracy is better than the process it was meant to improve.
And then nothing happens, or the deployment produces a fraction of the value the business case promised, and the project quietly ends. The usual explanations are about change management or data quality or executive sponsorship, and they are usually wrong. The explanation offered here is structural and it is convincing. The framework distinguishes three kinds of adoption.
A point solution drops a prediction into an existing task without changing anything around it, and it captures a small amount of value because the surrounding process was built around not having that prediction. An application solution changes the workflow around the prediction. A system solution changes the arrangement of decisions and responsibilities across the organisation, and it is the only one that captures the large gains, and it is also the hardest, because it requires changing things that currently work and that people are attached to. Most organisations do point solutions, measure disappointing returns, and conclude the technology is overrated.
The insight here is that the technology was fine and the deployment asked almost nothing of the organisation. The concept the book calls the between times is the other genuinely valuable idea. There is a long, uncomfortable period between a technology working and the systems being redesigned to exploit it, and during that period the measured returns look poor and the sceptics look correct. The parallel with electrification is used properly.
Factories bought electric motors and installed them where the steam engine had been, driving the same central shaft, and got very little benefit. The gains arrived decades later when someone realised that electric motors could be small and distributed, which meant the factory could be laid out around the work rather than around the driveshaft. That took a generation and a new set of buildings. The argument that we are currently in the equivalent period is the most useful thing you can take from this book, and it should change how you read any claim about AI productivity statistics either way.
The authors being economists is an underrated advantage. They ask about incentives, about who bears the cost of a change and who captures the benefit, about why a manager might rationally block a redesign that would help the organisation and harm their department. Technologists writing about adoption tend to treat resistance as ignorance. Economists treat it as a response to incentives, which is more useful because it suggests what to do about it.
Now the reservations. This is a slighter book than Prediction Machines, and if you have read that one you will find the first section largely familiar. The core new material could have been a long essay, and there is padding around it. That is a common fate for sequels and it is still worth noting before you buy both.
The evidence is thin. The case studies illustrate the framework rather than testing it, and they are the sort of retrospective examples that can be arranged to support most theories. The central argument is persuasive on its own logic and it is not demonstrated in the way an empirical reader would want. Timing has been unkind.
The book was written before generative models became the centre of the conversation, and the examples are drawn from a period when AI in business mostly meant predictive analytics. The framework transfers, and it transfers well, and the reader has to do the transferring themselves, which slightly undercuts the impact. The prescriptive chapters are the weakest. Having established that system redesign is the hard and necessary thing, the guidance on how to actually do it inside a real organisation with real politics is largely a restatement of why it matters.
That is the pattern of most business books and it is disappointing here because the diagnosis was so much better than average. My three point eight is for one genuinely important and widely applicable idea, explained with an unusually good historical parallel, from authors whose training makes them ask better questions than technologists do. Marked down for thinness relative to its predecessor, for illustrative rather than evidential support, for examples that predate the current wave, and for a prescriptive section that does not deliver. Read it after Prediction Machines, and read the system solutions chapters twice.