The Fragmentation Thesis
AI Isn't Killing SaaS. It's Splintering It Into a Thousand Pieces.
Discuss with AI
Why I wrote this
The AI conversation has split into two camps, and neither is useful if you’re actually building things.
On one side, unbounded hype. AI changes everything, disrupts all industries, renders all existing software obsolete by next Tuesday. On the other, existential doom. The robots are coming, jobs are over, the economy as we know it is finished. If you’re an engineer, an operator, or an investor trying to make real decisions, both narratives leave you stranded.
So I did the analytical work. I sat down with established economic frameworks (Coase, Christensen, Anderson, Perez), current industry data, and first-principles reasoning, and tried to understand what’s actually shifting in the economics of software. Not to predict the future, but to build a framework for thinking about it clearly.
This connects to something I wrote about previously: Abundance Engineering. The premise is simple. Instead of fixating on what AI destroys, understand where it creates abundance. This analysis is an attempt to do exactly that for the software industry.
What I found is that AI isn’t killing SaaS. It’s fragmenting it. The economics of who can build software, and for whom, are changing in ways that the hype-vs-doom binary completely misses. The result isn’t destruction. It’s restructuring, and restructuring creates opportunity.
Here’s the framework.
The foundation: why SaaS exists
Before understanding what AI changes, it helps to understand why things are the way they are.
The economist Ronald Coase explained in 1937 that organizations exist because coordinating activity internally is sometimes cheaper than coordinating through markets. Applied to software: when the total friction of building in-house exceeds the cost of subscribing to someone else’s product, buying is the rational choice.
SaaS vendors absorbed every category of that friction. Search costs (figuring out what to build) collapsed into product demos. Contracting costs (months of legal negotiation) collapsed into published pricing and self-service signup. Coordination costs (managing a dev team) disappeared entirely. Maintenance costs (infrastructure, security, compliance) were handled by dedicated teams no individual customer could justify. Integration costs dropped through standardized APIs.
The result was a powerful economic default: buying was almost always rational.
But this created a hidden consequence. The median SaaS company spends 88% of revenue on operating expenses. Customer acquisition takes 24 months to pay back. You need a multi-billion-dollar addressable market to make the unit economics work. Anything below that threshold, no matter how valuable to a specific audience, simply never gets built.
SaaS didn’t just determine which products won. It determined which products were allowed to exist.
The two shifts
AI disrupts this economic structure through two distinct mechanisms. The first reduces the cost of building software. The second changes what software is capable of doing once built. Together, they’re more consequential than either alone.
Building gets radically cheaper. The evidence on AI development productivity is more nuanced than headlines suggest. One of the most rigorous studies (METR, 2025) found experienced open-source developers were 19% slower with AI tools on complex, real-world tasks. Large-scale industry trials show 5-26% measurable gains, with less experienced developers benefiting most.
But these studies measured first-generation autocomplete tools. Current agentic tools (Cursor, Claude Code) operate at a qualitatively different level. And the real insight isn’t about coding speed. It’s that experienced engineers can now shift almost entirely into planning and specification, the work that actually determines whether software is good. Once a thorough plan is in place, AI implements it at a level that no longer bottlenecks on individual coding bandwidth.
The compounded effect of faster development, infrastructure cost reductions of 80-85%, eliminated boilerplate, AI-generated tests, and compressed time-to-market is significant even at the conservative end. What used to cost $500K-$2M and take 12-18 months can now cost $50K-$200K and take 4-8 weeks.
Software starts thinking, not just storing. The deeper shift changes what software can do. Consider Markel, a specialty insurer underwriting hard-to-place risks like equine mortality and commercial pollution liability. Before AI, their underwriters spent 30% of their time on low-skill triage: manually reviewing submissions, rekeying data, routing cases. After deploying an AI system, quote turnaround dropped from 24 hours to 2 hours. Underwriting productivity increased 113%.
This isn’t a story about building faster. It’s a story about software performing cognitive work it couldn’t do before. Traditional software was a system of record: it could store a completed review, an underwriting decision, a triage classification. But it couldn’t perform that judgment. An HR platform could hold a manager’s written feedback but couldn’t assess whether it contained actionable specifics or empty generalities.
AI changes this. Domain experts define criteria in natural language (“escalate if the customer seems frustrated,” “flag anything unusual for review”) and the system executes with contextual comprehension that deterministic code could never achieve. This is the shift from a system of record to a system of intelligence.
How they compound. The first mechanism builds the scaffolding cheaply: application framework, data models, APIs, UI, infrastructure. The second adds the intelligence layer that makes the software genuinely useful: domain experts define business logic, the system performs cognitive work. Together, they collapse the minimum viable cost of a purpose-built application by an order of magnitude.
The long tail comes to software
In 2006, Chris Anderson observed that when distribution costs approach zero, niche products in aggregate can rival blockbusters in economic value. Netflix and Amazon proved him right.
AI extends this insight from distribution to production. Anderson’s original point was about the cost of getting niche products to consumers. What AI collapses is the cost of creating niche products in the first place. That’s arguably more transformative: distribution cost collapse enabled consumption of existing niches; production cost collapse enables niches that never existed at all.
The US Bureau of Labor Statistics tracks dozens of specialized professions with fewer than 50,000 workers each. Podiatrists, forensic science technicians, marine engineers, audiologists, radiation therapists. Complex workflows, heavy documentation requirements, terrible existing tools. Each one a market that traditional software economics said wasn’t worth serving.
For every 47-minute patient session, genetic counselors spend 3 hours on surrounding paperwork.
The US has approximately 4,000 certified genetic counselors. At $200/month per user, that’s a $9.6M total addressable market. No VC would touch it. No SaaS company would build for it. Only 20% of their time is spent face-to-face with patients. The remaining 64% goes to case preparation, letter writing, insurance documentation, and follow-up. Existing tools are fragmented point solutions, some literally free, not because the market is well-served but because it’s too small to sustain real software economics.
A 2-person team, at roughly $75K in salary and overhead for 8 weeks, can reach profitability at 35 customers.
Under the new economics, a small team can build a purpose-built workflow tool covering case preparation through follow-up, with an AI intelligence layer that drafts letters, synthesizes case notes, and flags compliance issues. The market never existed before. It does now.
The Shopify parallel is instructive. Shopify reduced the barrier to e-commerce from $5,000+ to $29/month, enabling millions of small merchants. AI is doing something analogous for software creation: reducing the barrier to building a niche SaaS product from $500K+ to $50K-$150K, enabling thousands of small software businesses serving audiences of hundreds rather than thousands.
And this isn’t speculative. Vertical SaaS is projected to reach $369B by 2033, up from $106.5B in 2024. Samsara (logistics) is growing at 33% while Salesforce sits at 8%. The long tail is beginning to express itself economically.
What markets are pricing in
Over $1 trillion in software market cap erased in weeks.
In early 2026, the software sector experienced the “SaaSpocalypse.” Broad software indices fell roughly 15% in weeks, about 25% from 12-month highs. Salesforce declined approximately 26% year-to-date, ServiceNow 28%, Intuit over 34%.
The per-seat licensing model that powered two decades of SaaS growth is under direct pressure. Output is decoupling from headcount. Multiple vendors reported slowing seat growth as customers became more efficient with AI, buying fewer seats while maintaining output. Buyer budgets are shifting from incremental software purchases toward AI tooling. Gartner projects 40% of enterprise apps will feature AI agents by end of 2026, up from less than 5% in 2025, and that 40% of SaaS spend will shift from subscription-based to usage or outcome-based pricing by 2030.
Some analysts argue the sell-off reflects overextended valuations finding an excuse to correct, and incumbents with deep data moats retain significant resilience. That’s partly true. But the categories most exposed are exactly the ones the thesis predicts: point solutions with narrow functionality, high per-seat pricing, and limited proprietary data moats.
Why incumbents can’t just “add AI”
Salesforce charges $175/seat/month. If value migrates to an AI intelligence layer sitting on top of the data, what’s left underneath is essentially a managed database. But Salesforce can’t launch a $30/seat alternative without dismantling the economics that fund its enterprise sales teams, premium support, and thousands of engineers. That’s not a strategic choice. It’s a structural impossibility.
Three barriers compound. First, margin cannibalization: the entire cost structure requires high-margin revenue that a commoditized data store can’t generate. Second, technical debt: the platform was architected decades ago, and building a genuinely AI-native layer on top of it is a replatforming effort, not an incremental feature. A startup building from scratch faces none of these constraints. Third, incentive misalignment: a sales rep compensated on per-seat deals won’t advocate for outcome-based pricing, and a product manager measured on feature adoption won’t recommend replacing features with AI automation. Thousands of micro-decisions across the organization collectively resist transformation.
This pattern is consistent. Kodak invented digital photography in 1975 but couldn’t abandon film margins. Blockbuster understood streaming but couldn’t restructure its real estate model. Nokia recognized the smartphone threat but couldn’t pivot from hardware to software ecosystems. None were blind. All were structurally trapped. Microsoft’s cloud transition is the rare counterexample, but it required a CEO change and nearly a decade of transformation. Not every incumbent gets that runway.
Where value accrues next
If fragmentation is the structural trend, value reorganizes into four layers.
Data portability. As systems of record lose intelligence to the AI layer above them, what remains is structured storage. The opportunity isn’t in storage itself but in defining the standard: industry-wide reference schemas that make data portable, the way FHIR standardized healthcare records. Whoever establishes that standard captures coordination value even as the layer beneath it commoditizes.
Vertical intelligence. This is where value migrates from traditional SaaS. AI orchestration layers that handle workflows, decisions, and automation for specific industries. The highest-value position in the new architecture. Whether it consolidates around a few platforms or stays fragmented across domains is an open question.
Builder infrastructure. Small teams can build niche applications with AI but can’t necessarily operate them. Uptime, security, compliance, billing. The opportunity is an opinionated platform that bundles all of this for AI-native micro-SaaS builders. A SaaS company that profits from the fragmentation of SaaS. This layer is especially interesting because it doesn’t depend on any single niche succeeding, only on the aggregate trend continuing.
Intelligent discovery. Thousands of niche tools are only valuable if buyers can find them. Traditional discovery (Gartner, G2, enterprise sales) was designed for hundreds of products, not thousands. Over 83% of App Store apps are essentially invisible. ClawHub, the public skill registry for the open-source AI agent OpenClaw, grew to nearly 14,000 community-contributed skills in weeks, requiring semantic search just to remain navigable. Discovery is already the bottleneck. Whoever solves it captures a toll position on the entire fragmentation trend. The recursive structure is worth noting: AI enables the creation of niche tools, and AI may be the only mechanism that can solve discovery for those tools.
What this is, and what it isn’t
This thesis does not predict the death of SaaS. It predicts a restructuring.
At the head, large platforms consolidate and defend through data advantages and switching costs, but face pricing pressure as value migrates upward. At the tail, thousands of niche products emerge, serving audiences that were never economically viable to address. In the middle, generic horizontal SaaS without deep data moats faces the highest risk.
The goal of this analysis was never to predict doom or generate hype. It was to reason through a possible future from first principles, using established economic frameworks, so that anyone building or investing in software can think about it more clearly.
The thesis predicts restructuring, not destruction. Software doesn’t go away. It becomes accessible to markets that were never served before. That’s abundance, not scarcity.
If you’d like the full thesis with all citations and supporting evidence, reach out. I’m happy to share it.
What resonated? What did I miss? I’d welcome the conversation.