If AI is so good at prediction, where are the results? That’s the uncomfortable question Chapter 2 opens with. Over the past decade, measured productivity growth has fallen by roughly half, and for a majority of Americans real income has barely moved since the late 1990s. Machines can now read X-rays, translate languages, and drive cars, yet the aggregate numbers look like nothing much is happening.
The authors are quick to point out that this productivity paradox is not new. In the 1980s, computers were spreading through every office while the productivity statistics stayed flat, prompting economist Robert Solow’s famous observation that the computer age was visible everywhere except in the numbers. The gains arrived eventually, but only after businesses stopped using computers to speed up old processes and started redesigning the processes themselves.
The book gives this stretch a name: the Between Times. It’s the period after we’ve seen what a technology can do and before it has been widely adopted in ways that change how things actually work. We are in AI’s Between Times now. Point solutions and application solutions can be designed and deployed quickly, which is why we see so many of them. System solutions, the ones that will unlock AI’s real potential, take far longer.
Three kinds of solutions, and what actually separates them. The book’s distinction between point, application, and system solutions is the spine of the whole argument, so it’s worth making it concrete.
A point solution takes a decision your organization already makes and makes it better, without touching anything else. The decision was there before the AI arrived; the AI just improves the prediction feeding it. A bank already decides which transactions to flag as fraud; swap the old rules engine for a machine-learning model and you catch more fraud with fewer false alarms. A retailer already forecasts demand; replace the analyst’s spreadsheet with a model and the forecast gets sharper. Nothing downstream has to change. Same people, same workflow, same approvals, just a better input. This is the AI equivalent of bolting an electric motor onto the old central shaft.
An application solution uses prediction to create a decision that didn’t exist before, but still slots into the system as it stands. Nobody at a record label used to decide, song by song, what each listener should hear next. Streaming recommendations made that a decision, and a valuable one. A thermostat that learns when you’re home and adjusts itself is making a decision no human was making. These are new capabilities, not upgrades, but they’re self-contained. You can add them to the existing business without rethinking how the rest of it operates. This is the individual electric drive on each machine: new and useful, but the factory is still built the old way.
A system solution is what happens when a better prediction changes what everyone else should be doing. Consider a hospital that deploys an AI to predict which emergency patients will deteriorate. As a point solution, it’s a flag on a screen that a nurse may or may not act on. But if the prediction is reliable, the right response is to change how triage works, how beds are allocated, how staff are scheduled, and which patients get admitted at all. The value of the prediction isn’t in the flag; it’s in the reorganized hospital. Or consider retail: if you can predict demand well enough, you might ship products toward customers before they order, which upends inventory, logistics, returns, and pricing all at once. The prediction is the trigger, but the payoff lives in everything that has to be rebuilt around it. This is the single-story factory designed around the flow of work.
The test is independence. The authors boil the distinction down to one question: does the value of the improved decision stand on its own, or does it depend on other decisions changing too? A useful way to check is to imagine switching the AI off. If the only thing that happens is that one decision gets a bit worse, you have a point or application solution. If switching it off would break a whole set of processes that were redesigned to rely on it, you have a system solution. Independence is what makes the first two fast to deploy; interdependence is what makes the third slow, and what makes it valuable.
Why system solutions are hard, and why they matter. Point and application solutions tend to reinforce the existing system. They’re welcome because they don’t threaten anyone; they make the current way of working run a little better. System solutions, by definition, upend it. One better prediction changes what the next person should do, which changes what the person after them should do, and so on down the line. That means new workflows, new roles, new organizational structures, and resistance from everyone invested in the old arrangement.
Yet in many cases, system solutions are where the greatest returns on AI investment will be found, just as redesigned factories, not cheaper motors, were where electricity finally paid off. And because they restructure how industries operate rather than just how a single task gets done, system solutions will create winners and losers. The firms that figure out the new system first won’t just be more efficient; they’ll change the terms of competition for everyone else.
The Between Times, in other words, are not a sign that AI has failed to deliver. They’re the gap between seeing the technology and rebuilding around it, and the rest of the book is a guide to closing that gap.
