The Slow Boom and the Fast Break

I keep hearing two confident predictions about AI. Both get the speed wrong. One camp says we are months away from an explosion that rewrites the economy. The other says this is electricity all over again: a big deal that builds quietly over decades. They share one mistake. They assume the good and the bad arrive on the same clock. They don't. The upside is slow. The damage can be fast. That gap is the whole story, and almost nobody is planning for it.

Here is what I want you to take away. First, why the boom is slow even when the technology is good. Second, an honest answer to the best objection, the one that says this time the bottlenecks melt away. And third, the one stance that makes sense once you see the gap. It beats both hype and dread.

The argument both camps are having is the wrong one

Watch the two camps long enough and they start to look like mirror images. The optimist sees the capability curve shooting up and decides everything changes next year. The skeptic sees the same curve, shrugs, and says big technologies always take decades, so relax. Both are arguing about the destination. Boom or bust. World remade or barely nudged.

But the destination was never the interesting part. Speed is. And speed is not one number, because the upside and the downside do not run on the same engine. The optimist is roughly right about how fast harm can move, and wrong about the benefit. The skeptic is roughly right about how slowly the benefit builds, and wrong to be that patient about the harm. Each camp holds half the picture. They have glued the wrong halves together.

So stop thinking about AI as a thing that gets better. Start thinking about value as something you can either build up or break. Those are not the same kind of move. That difference is the whole game.

Value is a chain, and chains are stubborn on the way up

Picture anything worth money as a chain of steps. A loan does not get approved because one clever step happened. Someone finds the borrower. Someone gathers the documents. Someone judges the risk. Someone prices it, checks it against the rules, funds it, and services it for years. Each step is a link. The finished loan is the whole chain holding together.

Here is the key thing about these chains. They are pass-or-fail, not add-it-up. A loan that fails the compliance check is worth nothing, no matter how good the risk model was three steps earlier. You get no credit for the links that worked. The chain only pays off if every link holds. So the total tracks the weakest link, not the average one.[1]

Now suppose a tool makes one of those links a hundred times better. Say risk judgment becomes nearly free and nearly perfect. How much better is the loan business? Not a hundred times. Your gain is capped at what that one link was worth in the first place. Everything you did not touch now sets the limit.[2] You made one link brilliant and inherited a chain that is only as strong as its next weak point.

This is the part people skip. Making most links stronger barely moves the total, because the weakest link rules it. You can pour genius into nine steps. The tenth still sets the pace: the slow human sign-off, the clumsy data handoff, the rule nobody has touched since the last crisis. The chain does not really get stronger until that laggard does. And the laggards are usually the messy, trust-heavy links that resist automation the longest.

So when someone tells you a model now matches an expert at some narrow task, believe them. Then ask the only question that matters: how much of the chain was that task? Usually a sliver. Make the sliver perfect and you have bought the value of a sliver. The rest of the chain carries on at its own pace.

The returns run toward whatever is still scarce

There is a second move here, and people find it surprising. When you make a link cheap and abundant, the money does not stay there. It moves on.

Think about long-distance communication. Sending a message across the planet went from slow and expensive to instant and nearly free. You would expect the people who move the bits to have captured a huge share of the economy. They did not. Moving information got cheap, so the value drained out of it. It pooled around the things that stayed scarce: deciding what to say, and being trusted enough that someone acts on it. When one link becomes cheap, the returns flow to the links that are still hard. Scarcity pulls the money. AI does not remove scarcity. It just moves it.

AI will do this again, faster and in more places. As one kind of thinking work gets cheap, that work stops being where the value sits. The value shifts to what the cheap tool still cannot do: framing the problem, and making the call when the inputs disagree. This is why "AI will take your job" and "AI will make your job more valuable" can both be true of the same job at the same time.[3] Automate most of a role's tasks, and the person left doing the few hard ones becomes the scarce link in a chain that now runs faster everywhere else. That lasts until the last few tasks fall too. Then the role does not get more valuable. It ends. But the long middle of that shift pushes people toward the deciding links and away from the doing links. Hold on to that. It is where this is heading.

Slow to build, quick to break

Now for the heart of it.

A chain like this is stingy on the way up. To get the full boom, you have to strengthen every link, including the human, institutional, trust-heavy ones that move at the speed of a generation, not a release cycle. That is the slow boom. It is slow for built-in reasons, not because anyone is dragging their feet.

The downside is not that same story run backward, and I will not pretend it is. It works in a different way, and it is worth saying plainly. For decades, a bad output in one link got caught by a person in the next. Someone looked at the number, frowned, and said that cannot be right. Those people were buffers. They absorbed a link's failure before it spread. Now look at what the push for the slow boom actually does along the way. To make the chain run faster everywhere, you automate link after link. Each time you do, you quietly remove a human who used to stand there and catch things. You end up with a chain that is fast and tightly linked, with almost no one left in between.[6]

In that chain, a single failure does not get caught. It spreads. A poisoned data feed reaches a thousand decisions before anyone notices. A faked approval sails through, because every step trusts the one before it. Or the model is confidently wrong in the one spot where no human is left to flinch. You do not have to break the whole system to do real harm. You break one link. The buffers are gone, so the damage runs the length of the chain at machine speed. The same push that makes the upside crawl is what makes the downside sprint. It gets there by removing the very people who used to make failure survivable.

So here is the trap, in plain terms. Building value means lifting every link, and that takes a generation. Breaking it means cutting one link in a chain we have carefully emptied of anyone who could stop the fall, and that takes an afternoon. We are about to spend years exposed to the fast clock while we wait on the slow one.

The objection that almost works

The strongest pushback I get goes like this. You are describing the old world. This time, the weak links are exactly what AI fixes. Judgment, coordination, the messy human stuff that used to slow everything down, is precisely what the new systems are starting to do. So the bottlenecks dissolve from the inside, and the boom is fast after all.

I think this is right about the ceiling and wrong about the clock. And that difference is everything.

Grant the whole point. Say AI does get good at the soft links. It still has to get built into how real organizations work, and that is where the time goes. Being able to do the job is not the same as being used for it. A tool that can make the judgment is not the same as a bank that has rebuilt its process around it, trained its people, satisfied its regulator, survived the first embarrassing failure, and earned enough trust for someone to sign off on letting the machine decide.[4] Consumer apps can spread in a few years. The regulated, high-stakes, trust-heavy rewiring this piece is about has usually taken close to a generation, even when the technology was ready early. That is simply how long it takes to change institutions and the habits of the people inside them.[5]

There is a sharper version of the objection, and it is worth taking seriously. New players do not wait. A company built from scratch on AI has no old process to rewire and none of the incumbent's caution. It can move at the speed of code. Grant that too. But look at what a brand-new, fully automated operation also is. It is tightly linked and barely buffered. The same thing that lets it skip the slow climb is what leaves no human standing at the link that matters. So the fastest adopters are also the easiest to break. The objection that looked like it closed the gap actually widens it. Fast to build, in the rare case it happens, buys you fast to break as the price.

What to do with a slow boom and a fast break

Right now almost everyone is planning for the destination. Will it take my job? Should we buy the platform now or wait a year? Those are destination questions, and the destination was never the part we could see clearly. Meanwhile the shape of the path, slow on one clock and fast on the other, sits in plain view and goes unmanaged. That is the gap I opened with. It is a planning gap, not a forecasting one.

Seen that way, the slow boom stops being a letdown and becomes the one thing you actually have: time. Stop treating it as a delay to wait out. It is the only window you get to prepare, and the prep is not a mystery.

Worry less about which jobs vanish next quarter. Worry more about where you have quietly let a chain run fully automated, with no one standing at the link that matters. The buffer you removed for speed is the brake you will wish you had kept. So put it back on purpose. Treat the human in the loop not as a cost to cut, but as the deliberate stop in a system that now moves too fast to fail safely on its own. The scarce, valuable person in the years ahead is the one who can say stop. The one who decides who runs the increasingly capable pieces, instead of letting them run alone. This is risk management, not sentiment. A chain that has lost its slack needs someone who can still reach the brake.

Abundance is genuinely possible at the end of this. It is just not automatic, and it is not safe by default. Who it serves, and how badly it breaks along the way, are still ours to decide. The window to decide well is open now, while the boom is still slow and the break has not yet found us.

References

[1] Michael Kremer, "The O-Ring Theory of Economic Development," The Quarterly Journal of Economics, 1993. academic.oup.com

[2] Daron Acemoglu, "The Simple Macroeconomics of AI," NBER Working Paper No. 32487, 2024. nber.org

[3] David H. Autor, "Why Are There Still So Many Jobs? The History and Future of Workplace Automation," Journal of Economic Perspectives, 2015. aeaweb.org

[4] Erik Brynjolfsson, Daniel Rock and Chad Syverson, "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies," NBER Working Paper No. 25148, 2018. nber.org

[5] Paul A. David, "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox," American Economic Review, 1990. ideas.repec.org

[6] Charles Perrow, Normal Accidents: Living with High-Risk Technologies, Basic Books, 1984. en.wikipedia.org