A few weeks ago, I spoke with a healthcare technology executive caught in a familiar bind.  His entire team – from engineering to marketing – was eager to integrate AI into their product and service offerings, to stay ahead of their competitors and customer demands. But their CEO keeps forwarding articles about the entire AI industry being caught up in an investment bubble, and how the whole AI trend was headed for a crash.

And, honestly, this was a case where both sides were absolutely right.

When people talk about “AI hype” or “tech bubbles,” they often assume that just because investors have overblown expectations for a technology, the technology itself isn’t valuable.  But history shows that nearly every game-changing technology of the past two hundred years, from railroads to electricity to the internet, has experienced one or more bubbles between long years of grinding, gradual expansion.

The question isn’t whether we’re in a bubble (we probably are), the question is what kind of bubble, and whether smart organizations can benefit even as it pops.

To answer, let’s look at my favorite historical metaphor for AI adoption in modern organizations.  I’m not talking about smartphones, cloud computing, or even computers, but rather…  mechanized farming in the 19th and 20th century.   

Bear with me here.

Blowing Bubbles

Before we go back in time, let’s take a moment to ponder a deceptively simple question: “Are bubbles really so bad?”

The answer (as always) is, “it depends.”

The popping of a market bubble is painful for the investors who lose their money, and can be economically devastating if too many people are invested (like in the stock market collapse that triggered the Great Depression.)  That said, some bubbles are actually beneficial to society in the long run.

Let’s look at four different types of bubbles, their short term risks and (in some cases) long-term benefits:

Bubble Type 1 – Irrational demand for useless stuff

This is the purest kind of bubble, where the value of the assets in question is based on collective delusion, fashion, or misinformation.  Think tulip bulbs in the 1600s, Beanie Babies in the 1990s, and meme stocks / meme coins / NFTs in the 2020s.  And once the public comes to their senses, prices collapse to zero and nothing of substance remains.

Bubble Type 2 – Premature investment in technology that’s not ready for prime time

This is a case of “right idea, wrong decade”, where investors pour money into something that’s conceptually sound but technically impossible or economically infeasible at the moment (due to lack of infrastructure, materials, complementary tech, etc.)  A great example is early streaming video platforms like Pulse TV and RealMedia in the late 1990s when the average home Internet connection was far too slow for quality viewing experience, or electric cars in the 1900s, which lost out to combustion engines due to primitive battery technology and other limitations.  On the upside, even when these bubbles burst, the idea survives until the rest of the world catches up (see Netflix and Tesla.)

Bubble Type 3 – Unrealized investment in systems that require ubiquity

This type of bubble involves technologies that only deliver real utility if everyone participates, yet fail to catch on.  Examples would include video phones in the 1960s (they worked, but so few were installed that there was no one to call) and various VR social worlds like Second Life and the Metaverse in the 21st century (forgotten by all but a tiny handful of hardcore users.)  Like charcoal that never ignites, these bubbles end quietly, smoldering in small niches or getting absorbed into something else (e.g., VR becoming a mode of gaming rather than a world-changing social platform).

Bubble Type 4 – Overproduction of useful things 

This is the definition of a “productive bubble” where there’s a technology capable of delivering real value right now (even for isolated users), but providers and investors get way ahead of demand (in terms of production capacity, capital expenditure, or building out infrastructure.)  Some of history’s biggest technological transformations fall into this category including railroads in the 19th century (too much track was laid too soon, causing railroad lines to go bankrupt but leaving a transformative logistics network that enabled the next phase of the industrial revolution), electrification and telegraph lines (same dynamic as railroads), and the 1990s dot-com boom (which saw the failure of most early Internet companies fail but left enough excess fiber optic capacity to satisfy the public’s growing appetite for bandwidth into the 2010s.)  The hallmark of this type of bubble is the durable infrastructure, skills, and knowledge created that makes sustainable progress possible.

Reaping Rewards: Mechanized Agriculture in the 19th Century

Stop me if you’ve heard this one: there’s a hot new technology getting all kinds of hype in the press.  Its inventors talk about how this unprecedented innovation is going to change civilization itself.  But businesses are skeptical: the tech is expensive, unfamiliar, and relatively complex, and many early implementations fail.  Eventually the hype cools and the investment bubble around the new tech bursts, contributing to a larger, ongoing economic crisis.

Obviously, given the heading of this section, I’m not talking about AI in the 2020s.  

When Cyrus McCormick introduced his mechanical reaper in 1834, the media heralded it as revolutionary.  Imagine: a machine that allowed its operator to harvest as much grain as five men!  Advertisements promised these new mechanical reapers would “change the world.”   

Despite the hype, sales were slow.  Most farmers didn’t trust the machine.  Among farmers who bought those early reapers, a few saw immediate productivity gains.  However, the majority experienced “failed implementations” as the mechanical reapers only really worked on flat, dry fields: not the hilly, muddy terrain of many farms.  On top of that, the machine was expensive and complex to maintain, and rural users lacked access to repair shops or even replacement parts.

While McCormick used his investors’ money to manufacture thousands of reapers in anticipation of demand, only a few hundred reapers were actually sold in the first decade before the hype cooled.  Meanwhile a financial crisis in 1857 nearly ruined McCormick, who was borrowing heavily to keep the business running. 

It was only later in the 1860s to 1870s that lighter materials, better gearing, and standardized parts finally made reapers reliable enough for serious agricultural work at scale.  It was during this era that McCormick’s company rebounded and finally grew into the industrial giant investors believed it could be – about 15 years post-bubble, helped in part by farm labor shortages during the American Civil War. 

Even then, McCormick’s reapers could only automate part of the harvesting process and it wasn’t until the introduction of combine harvesters in 1910 that it was possible to fully automate grain harvesting, and some U.S. farms didn’t properly mechanize until 1940, a full century after McCormick filed his original patient.  

Applying our categories, one could say the reapers began as a combination of a type 2 bubble (“idea ahead of its time”) and a type 4 bubble (“producing too much of a good thing”).  Many early adopters did benefit (especially those who happened to live in flat regions where early reapers performed well) though uptake and integration was uneven until the technology – and society – caught up. 

History Rhymes: the Hypothetically Imminent AI Bubble

Let’s be clear: every historical moment is unique and today’s AI mania is different from reaper-mania in many important ways.  Still, applying the old maxim “history doesn’t repeat, but it often rhymes” – what does the story of McCormick’s reapers portend for generative AI?

If (that’s “if”) generative AI follows a similar pattern and if (once again “if”) we are due for some kind of bubble burst in first quarter 2026, we could say that first quarter 2022 (when ChatGPT 3.5 completed training) to fourth quarter 2025 maps to the same phase the reaper went through from 1834 to 1857, but in just 3.5 years.  So about 6.5x as fast.  

How does AI adoption compare to reaper adoption in this first “pre bubble” stage?  The parallels are striking: 

• Right now a few organizations and industries in fields that naturally lend themselves to AI automation are achieving significant productivity gains while others are struggling to implement the tech effectively. 
• Today’s AI tools don’t automate entire workflows yet, but they massively multiply human capacity (by about 6.5x) for many tasks — coding, writing, document review, customer service triage, etc.
• The industry is churning out more data center capacity than customers actually need, and unlike railroads or fiber optic cable many of today’s servers will go obsolete before AI’s moment fully arrives – like early reapers from the 1840s rusting in some farmer’s barn.
• The AI industry isn’t making enough to pay for itself, such that any dip in investor confidence could bring the whole thing tumbling down and possibly trigger a “Panic of 2025/6” which won’t be fun to live through.

If and when it comes, the crash will likely be even more devastating than what McCormick experienced.  Imagine if, instead of building 10,000 reapers versus sales of a few hundred, McCormick had somehow manufactured 100,000 with no corresponding increase in sales: that’s the current level of overinvestment in AI tech versus demand we see today. Also the crash will be worsened by the fact that  – like reapers that eventually broke down from rust or wear – a lot of today’s AI data center infrastructure will be obsolete in a few years, as opposed to the railroads, electrical lines and fiber optic networks left by other tech bubbles, which could remain useful for decades or even centuries.

That said, if the metaphor extends beyond the bubble (yet again “if”), the subsequent 4 or 5 years post-crash could see mind boggling expansion of AI as a technology and industry, with wealthy economies achieving the near total automation of knowledge work around 2040 (approximately ⅙ as long as it took to automate agriculture in the 19th and 20th centuries) with a few big, vertically integrated players dominating the field just like International Harvester –  the corporation that grew out of McCormick’s company – with its massive combine machines.

Clouding the Crystal Ball

Taking this tenuous comparison to its logical conclusion – what does the end game of “full mechanization” of knowledge work even look like?

First, the tech will be different, either because AI itself evolves or complementary technologies allow it to reach its full potential.  Combine harvesters – capable of not just reaping but also threshing and winnowing (i.e., later steps of the harvesting process) – existed from the beginning in the 1830s, but were completely impractical until the invention of diesel powered tractors capable of pulling them.  Likewise, it might take some other complementary enabling technology (Robotics?  Truly ubiquitous IoT?  Neural interfaces?  New types of databases?  Quantum computing?  An alternative to SharePoint that actually works?) to maximize the value of AI. 

And what will become of human workers?

In 1800, roughly 70–80% of the U.S. workforce worked in agriculture.  By 1900, that number had dropped below 40%. By 1950, it was under 10%.

Mechanized agriculture didn’t eliminate work: it transformed it, as farmhands became machine operators, mechanics, and eventually industrial workers in cities, raising middle class standards of living to levels previous generations of farmers couldn’t have imagined (though we could fill an entire series of 500-page airport bookstore non-fiction bestsellers with all the other history  glossed over in that sentence.)

Conclusion

Returning to the present, the impact of an AI bubble won’t just be the billions of investor dollars lost and economic aftershocks when it pops. It will also be the useful technology, infrastructure, and institutional knowledge that survives once the hype deflates.

AI, like the reaper, is clearly a type 4 bubble (“overproduction of useful things”) with elements of type 2 (“right idea, wrong decade”). The technology isn’t just promising: I’ve seen firsthand how a well-designed AI agent can allow a single user to accomplish what previously required five researchers, financial consultants, or social service providers. And it will work even better as complementary technologies and supporting infrastructure click into place.

So, what does all this mean for that healthcare technology executive and his skeptical CEO, deciding whether or not to infuse AI technology into their products and services?  What would history have them do?

Probably the same thing smart investors have done in every tech bubble: distinguish between the speculation and the actual capabilities of the technology involved.  

The reaper didn’t disappear when McCormick’s company nearly went bankrupt in 1857. It got better, cheaper, and more reliable, and eventually delivered on its promise of changing the world.  Your AI strategy should probably assume that, even if history doesn’t exactly repeat, it will most likely rhyme. 

Finally if you’re disappointed that we didn’t make a joke about that Will Ferrell “Don’t Fear the Reaper” skit… here you go: https://www.youtube.com/watch?v=cVsQLlk-T0s