For most of the software era, there has been an implicit rule governing what kinds of markets were worth building technology companies around:
The market had to be large enough to justify the company required to serve it.
In other words, good software was pricey to build. Even a relatively narrow vertical SaaS company needed engineers, PMs, salespeople, customer support, infrastructure, and enough domain expertise to understand how customers actually worked. Before the first dollar could be earned, a meaningful amount of investment had to be made.
That imposed a largely invisible minimum viable market size on software.
However, that never changed the fact that there are thousands of industries with painful workflows, meaningful economic activity, and terrible technology. They just never attracted much venture-backed attention. Not because the problems were unimportant; the economics just didn’t math.
A software company serving a giant market like healthcare, financial services, construction, or logistics could amortize millions of dollars of product development across tens of thousands of customers, whereas a company serving 2,000 specialty businesses could not.
AI may be changing that equation.
Most of the conversation around artificial intelligence focuses on expanding software capabilities (e.g., better interfaces, autonomous agents, new products, and entirely new categories).
But one of AI’s most important consequences may be more mundane and (arguably) more economically significant: AI is reducing the cost required to build and operate specialized software businesses.
And as that cost falls, the minimum market size worth serving falls with it.
Software Has Always Had a Minimum Efficient Scale
We tend to think of software as having extraordinary economics because the marginal cost of distributing another copy is close to zero. That is true once the product actually exists, but creating, selling, implementing, and supporting software calls for considerable fixed investment.
Consider a hypothetical niche industry with 3,000 businesses. Its operators may spend tedious amounts of time moving information between spreadsheets, PDFs, emails, databases, and physical paperwork. They may employ people specifically to reconcile records, prepare reports, schedule work, check compliance requirements, or manually enter data.
The pain can be very real, but pain alone does not a good software market make. If the average customer can only justify paying $10,000 per year, the theoretical market may top out at $30 million in annual spend. Against that scale, a traditional software company is very difficult to justify. You still need engineers, an understanding of obscure industry workflows, sufficient integrations, onboarding and customer support, and someone to actually sell the software. Even if the product works, the resulting business is likely much too small to support the effort and systems required to create it.
This is one of the stranger aspects of the modern tech scene. We have extraordinarily sophisticated software for some parts of the economy while huge chunks of the collective economy still run on Excel, email, phone calls, paper forms, and software built twenty years ago.
Usually the explanation (or excuse, among VCs) is that these industries are too technologically resistant or slow to adopt. That may be true at times, but the more accurate explanation is that these were never attractive enough markets for sophisticated tech companies to pursue.
AI Changes the Denominator
Most tech investing is oriented around the numerator. (e.g., “how large is the market?”, “how fast is it growing?”, “how much spend can customers support?”, etc.)
It’s not wrong to assume that AI may create enormous opportunities by increasing those headline numbers. But it may matter equally as much because it changes the denominator.
Historically, the denominator was pretty stubborn. A software company needed a certain amount of organizational infrastructure regardless of whether it served a $100 million market or a $10 billion market, so always swing big.
AI potentially compresses that cost across much of the organization as software development speeds up, customization becomes cheaper, and support becomes increasingly automated. Increasingly, software can perform meaningful portions of the underlying work rather than merely helping humans execute it better. The result is not purely that existing software companies become more profitable. The more interesting impact is that markets that previously failed the attractiveness test above begin to pass it.
A market capable of supporting $20 million of high-margin revenue is not especially interesting if building the company requires 150 employees.
It becomes much more interesting if it requires 20 or fewer.
The Long Tail of the Economy
Imagine ranking every potential vertical software market in the economy from largest to smallest (there are countless market maps that do exactly this, so I will refrain from doing so now). At the far left sit the obvious categories: healthcare, banking, insurance, construction, transportation, retail, manufacturing, etc.
These industries rightfully support enormous software markets, which is why generations of entrepreneurs and venture investors have repeatedly attacked them.
Move farther to the right, and the markets become increasingly specialized: Commercial fire-alarm inspection, environmental testing, elevator maintenance, niche distributors, certification providers.
There are thousands of industries like these. Most people outside them barely know they exist, yet many of them are economically significant. And many still operate through a haphazard combination of legacy software, spreadsheets, documents, phone calls, and institutional knowledge.
The farther down the curve you travel, the less venture-backed software you tend to find.
Historically, there was a fuzzy cutoff point. To the left of that line, the market was large enough to justify attempting to build a software company, and to the right, it wasn’t.
AI may move that line much farther down the long tail.
The aggregate economic activity represented by these ignored markets is enormous. Individually, many may look trivial, but collectively, they constitute large portions of the physical and service economy. I’m increasingly less interested in another generation of AI applications competing in markets that software already serves well. I’m much more interested in enabling thousands of markets that have barely been software markets at all.
The Best Opportunities May Initially Look Bad
If this thesis is correct, some of the most attractive AI-native businesses of the next decade may look unimpressive through the traditional venture-capital lens. Small TAMs, obscure customers, slower growth, ugly workflows.
In a situation where building a software company requires tens or hundreds of millions of dollars of capital and hundreds of employees, those problems matter. However, in a world where a small, elite team can build, sell, and operate quality software at meaningfully lower cost, those factors matter much less.
A $50 million topline company with good margins and modest capital requirements can be a phenomenal outcome even if it never becomes a $10 billion public company. And there may be far more opportunities of the former size than there are giant horizontal software categories.
The tech industry is obsessed with building ever-larger companies, but AI’s killer ability may be making smaller markets more valuable.
From Software That Organizes Work to Software That Does Work
There is another reason the traditional TAM calculation may break down. The first generation of vertical software largely organized human activity. It stored customer information, scheduled workers, created invoices, tracked inventory, etc. The point is that people still performed most of the underlying work; software just digitized and organized large chunks of it.
AI, on the other hand, allows software to move deeper. It can read the document and prepare the response, reconcile the records, prepare the report, and communicate with the customer without any human involvement.
That changes the economic pool software can attack. A niche industry might spend only $15 million per year on software, but $100+ million on people performing the information-intensive work. Once software starts doing the actual work rather than organizing it, existing software spend becomes a poor proxy for market size.
This is a reason some seemingly small software markets may turn out not to be tiny at all. Measure them by the total cost of the work being performed, not what customers historically spent on software.
A Different Kind of Software Economy
If this thesis plays out, the implications could go beyond which startups get built. For the last twenty years, the “ideal” model was:
That model makes sense when building and execution is expensive, but the org structure that worked best when software was expensive probably isn’t the one that works best when software is significantly cheaper to create and operate. If the economy contains thousands of newly viable microverticals, we may need different institutions to find, build, and own them.
So while others are primarily focused on which new applications AI will enable, an equally interesting question is what kinds of businesses become newly possible when the fixed cost of specialization collapses?
Fewer Salesforces, ServiceNows, and Workdays, and more (i.e., hundreds or even thousands) highly specialized businesses that would never have made sense to build before.
For most of the software era, founders competed to capture the largest markets. AI may open another frontier entirely, not by making the biggest markets bigger, but by finally making thousands of markets that were previously too small worth serving.
As the cost of specialization falls, the minimum viable market falls with it.