A bubble tells you prices were wrong, never that direction was.
Every era of waiting has its sophisticated excuse, and this era's is spoken in the CFO's register: we will let the bubble sort itself out. It sounds like prudence. It quotes real numbers: valuations priced for perfection, capital expenditure racing ahead of revenue, circular deals where chipmakers fund model labs who buy chips, startups worth billions on products younger than a lease. The person saying it has usually lived through 2000 or 2008 and carries the scar tissue honorably. And after four chapters of demolition, it is the last dignified reason left for doing nothing.
So let us treat it with respect and assume the skeptic is entirely right. Assume the valuations are wrong, the capex is overbuilt, the froth is froth. Grant every clause of the bubble case, because this chapter's argument does not need to win that debate. It needs you to see that the debate is about the wrong question.
Whether AI is a bubble is a question about prices. Whether AI changes how companies run is a question about direction. The public conversation treats them as one question, and they are separable: a bubble tells you capital overpaid, and overpayment can sit on top of a real direction as easily as an imagined one. A bubble tells you prices were wrong, never that direction was.
The sharp reader will raise a fair objection: promoters of past bubbles also insisted the direction was real. Worth granting fully. The dot-com era is useful because commerce continued moving online through the market crash. Market collapse and operational adoption can coexist. So the objection does not settle anything by itself. What settles it is a working standard, and this book keeps applying the same one: judge by what can be run, measured, and repeated in your own workflow now, at a visible cost, and give projections about what people will surely do someday the weight projections deserve.
Run the tape on the example everyone already knows. The dot-com crash was real and severe, and companies failed. If a pop could settle a direction argument, that one should have.
One narrow fact is worth carrying away from it. Through the crash itself, commerce kept moving online. Census estimates show U.S. retail e-commerce sales rising 13.1 percent year over year in the fourth quarter of 2001, and total e-commerce sales for 2002 rising 26.9 percent over 2001. Households and companies went on adopting the thing whose stocks were burning. That is the claim, at full size: the crash corrected prices and did not reverse adoption. It is not a law that a durable transformation hides inside each bubble. Plenty of manias have had little underneath.
Carlota Perez supplies the most useful frame for the pattern. Across several technological surges, she distinguishes an installation period, led by finance and prone to frenzy, from a deployment period, led by production. A bubble and collapse often mark the boundary between them. Treat her framework as a lens for asking where in a surge you might be standing. It is a reading of several historical cases, and it does not prove that AI must follow the same path.
One more correction, this time against my own side. It is tempting to close with a ratchet: prices oscillate, capability only moves forward. Too strong. A demonstrated capability can regress in the next release, disappear from a product for commercial, legal, or safety reasons, be restricted by a regulator, or prove uneconomic to serve at scale. What a crash cannot take back is more modest and still decisive over time: the knowledge that the task has been done. A capability that has been demonstrated can be rebuilt when the economics allow, by whoever finds it worth rebuilding. Prices record what capital believed. Demonstrations record what is possible. Only the second survives a correction, and it survives as knowledge, with no promises attached about price, availability, or timing.
Now assume the crash actually comes
This is where the chapter turns, because the waiting executive has not thought one step past the pop they are waiting for.
Suppose it happens: funding winter, valuations halved, the weekly model announcements gone quiet, a graveyard of AI startups. The waiting executive imagines that world as vindication. Walk through what it might contain, because the futures fork, and only one branch is the one the analogy prepared you for.
In the first branch, some deployed capabilities may stay in place. Talent from failed companies may become available to other employers. Some costs may fall as surviving vendors compete for customers. None of that is guaranteed, and this branch is the more forgiving of the two; for a prepared company, it can be survivable and even generous.
The other branch costs this book's thesis something, and you should hear it from me. A deep enough winter can slow the direction itself. Capital scarcity can starve the research that was driving capability. Surviving providers concentrate, and concentrated providers price accordingly. Products you depend on get shut down. Regulation written in a crisis can restrict what remains, and infrastructure that was priced for infinite growth stops getting built. In that branch, intelligence stays expensive for years, diffusion slows, and part of what this book treats as arriving is postponed. I cannot tell you which branch you get, and neither can anyone selling certainty about it.
So plan on what survives both branches, because a short list does. Clean data is worth having whether intelligence is cheap or dear. Legible processes pay under any vendor landscape. Knowing which decisions are yours, and who owns each one, keeps its value in either future. So does the habit of evaluation, measuring a tool against a baseline before believing it, and so do bounded probes, small enough to close down without ceremony. That list is Movement II of this book, and its worth does not depend on how the bubble resolves. The startup that rebuilt itself around cheap intelligence does not get less dangerous when intelligence gets cheaper. Your bet on the bubble popping was never a bet against AI. It was a bet against your own future costs falling, and you win it in the one way that does not help you.
A word to the executive whose winter plan is acquisitions: buying cheap in the crash works for tools, talent, and customer books, and history will hand you bargains in all three. What it cannot buy is the one asset this book is about. Another company's product does not clean your data, and an acquired team does not rewire how your decisions flow; transformations are notoriously the thing acquisitions fail to import. The winter shopping list is real. It is just not a substitute for the rebuild, and it rewards precisely the companies whose base is already ready to absorb what they buy.
Even the forecasting establishment holds both futures at once. Gartner has forecast that more than 40 percent of agentic AI projects will be canceled by the end of 2027, and also that 33 percent of enterprise software applications will include agentic AI by 2028. Both are forecasts from the same research company, and neither has happened yet. The pair is worth quoting because it is coherent: particular projects can be canceled while agentic features still appear in a growing share of enterprise applications. Whether these particular numbers come true is open. Treat them as one firm's structured guess, and check the guess against your own deployments before repeating it.
Your feelings about this are wrong, in both directions
The bubble debate runs on conviction, so it is worth watching what happened when conviction met a stopwatch.
In early 2025, METR ran one of the few controlled trials of the era: experienced open-source developers, working in their own mature codebases, completing real tasks with and without AI tools. The developers expected a 24 percent speedup. Even after the study, they believed AI had sped them up by 20 percent. Measured, their completion time increased by 19 percent. Skilled people, close to the technology, misjudged its effect on their own output, on those tasks, at that moment, with confidence.
Handle the result with its own discipline, because METR does. It is a bounded finding about particular developers, particular work, and early-2025 tools, and the organization's February 2026 update reported that a newer experiment had selection and measurement problems severe enough to offer only weak evidence about current effects. So the study does not tell you what AI does to developer productivity today. What it demonstrates has a longer shelf life: perception of AI's effect, even expert perception at close range, can miss the measured direction entirely.
The tempting next move is symmetry, and I will not make it. That the developers in the study misjudged the effect on their own work does not license the conclusion that skeptics are wrong by an equal and opposite margin; nobody has run that measurement, and errors do not arrive in matched pairs by courtesy. What the evidence supports is narrower and sufficient. Unmeasured conviction is not a reliable operating instrument in either direction, and the executive dazzled by a demo and the executive dismissing the field on impression alone are running on the same fuel.
The cure is indifferent to which mood you arrived with. Pick one real workflow in your own operation. Establish its baseline: what it costs today, in money or minutes, at what error rate. Introduce the tool under conditions you control, and measure something a decision turns on. A surprise in either direction teaches you about your operation. Declining to measure is choosing a mood, and moods compound.
The clock that broke
There is a quieter mechanism underneath the waiting position, and it involves clocks.
Corporate planning assumes the world holds still long enough to plan against: an annual budget presumes that the capabilities available at its end resemble the ones at its start. In parts of this domain, that assumption is currently unsafe. Stanford's 2026 AI Index documents large gains on several difficult benchmarks within a single year, movement fast enough to outrun an annual planning cycle where it occurs. The same report documents the other half, which deserves equal weight: major failures, uneven reliability, capability that is jagged across tasks that look adjacent. Fast benchmark progress does not tell you the technology will work in your workflow, and it does not tell you the pace will hold. Some of this may stabilize inside your planning horizon. Some of it has not yet.
The managerial problem is less the speed than the memory. Somewhere in your organization is a verdict formed some time ago: an evaluation that concluded the technology hallucinated too much, or could not meet compliance requirements, or was not ready for customer contact. That verdict may well have been correct when it was made. It may be governing decisions today with the full authority of due diligence behind it, in areas where the evidence beneath it has gone stale. The fix is unglamorous: date the evaluations that gate real decisions, and reopen the dated ones on a schedule instead of in a crisis. Organizations are better at producing conclusions than at expiring them. Companies do not just plan more slowly than a fast capability moves. They remember more slowly, and an org chart has no mechanism for expiring its own conclusions.
This breaks "wait for clarity" less cleanly than I would like, so let me say precisely where waiting is right before saying where it fails.
Waiting is a rational position for irreversible commitments. If the decision in front of you is a large platform contract, a multi-year infrastructure buildout, or a tool purchase that locks you into one vendor's architecture, then price uncertainty is a real cost, option value is real value, and the CFO counseling patience is doing the job well. Prices set in a frenzy are bad prices to lock in. On those decisions, this book sides with the skeptic.
The position weakens as the decision becomes reversible, and the work this book cares about is mostly reversible. Cleaning the data you already own. Making your processes legible. Clarifying which decisions belong to whom. Building the habit of evaluation. Running probes bounded enough to shut down without ceremony. None of that locks you to a vendor, a model, or a forecast, and each pays under a boom, a crash, and the muddled middle you are most likely to get. Using the bubble argument against this work is using a price argument against decisions with almost no price exposure, and the mismatch is worth noticing when you hear it, including from yourself.
Then why does the mismatch survive in competent organizations, run by people who can tell an option from an obligation? Because the bubble argument does two jobs, and only one is analysis. The other is protection. Existing commitments, the systems already bought, the processes already staffed, the roadmaps already promised, benefit whenever a rival claim on resources is deferred, and the bubble supplies a deferral that asks no one to defend the status quo on its merits. It also performs a quiet conversion: an operating question, whether to make your own work more legible and better tested, becomes a market question, whether AI valuations will hold, which has no answer on your calendar and therefore no deadline. This is not an accusation against everyone who says the word; some of the people saying it are simply right about prices. The test is what the word is being used to postpone. If it postpones a locked-in purchase, it is prudence. If it postpones learning about your own operation, it is protection wearing prudence's clothes.
And the protection is worth a chapter of its own, because it is larger than this one argument. The deferral comes from somewhere: from the system that built your company's current success, defending the commitments that success is made of. A competent organization protects what works, and most of the time that protection is the correct behavior. The next chapter is about what happens when the same machinery meets a change that requires the organization to question its own working parts, and why the immune system that keeps a company alive is also the thing most capable of killing its adaptation.
The same discipline, turned around, is a career instruction. If you are waiting for the noise to settle before changing how you work, ask what the wait defers. Making your repeatable work easier to inherit, and building the judgment a handoff cannot carry, are reversible moves with no vendor risk, and they hold up under both of the futures on offer. If the capability keeps moving fast, you arrive with your week already shifted toward the work machines cannot yet inherit. If it stalls, you have still become someone whose methods are teachable and checkable and whose judgment is visibly their own, which is a durable operating position in its own right. A Blank Collar is making the repeatable part easier to hand over and becoming more useful beyond it. From the outside that looks like restlessness. It holds its worth in either future. It holds either way.