The Wrong Question
A memo on AI, SaaS, and second-level thinking
There’s a question that has consumed a remarkable amount of investor attention over the past two years:
Will AI kill SaaS?
I want to suggest that this is the wrong question. And in investing, asking the wrong question is often more dangerous than having the wrong answer.
First-Level Thinking About AI and Software
First-level thinking on this subject goes something like this:
AI makes it cheaper to build software. Therefore companies will build more in-house. Therefore they’ll buy less from vendors. Therefore SaaS is in trouble.
This chain of logic isn’t wrong exactly. It’s just incomplete. The investors I most respect insist on asking not just “is this true?” but “is this true in the way the market assumes, and for whom, and by how much?” Those qualifications are where the real work happens.
The Test That Actually Matters
For any given SaaS product, ask one question:
Can a customer achieve better total economics by building and maintaining this capability in-house — with AI assistance — than by paying the vendor?
If yes, the vendor is in trouble. If no, the vendor is fine. Call the first category Pseudo-SaaS and the second Real SaaS.
This framework has a virtue most frameworks lack: it’s testable. You’re not asking “does this product have a moat?” — a question that invites motivated reasoning. You’re asking whether the all-in cost of building, maintaining, and operating internally beats the subscription price. That’s a question you can actually try to answer.
When you do, an important insight emerges.
The relevant customers for most enterprise SaaS revenue are large organizations with procurement processes, compliance requirements, and engineering teams competed for by dozens of priorities. For them, “you could theoretically rebuild this internally” has always been true. The reason they didn’t was that the full cost of doing so exceeded the subscription. AI compresses build costs at the margin, but it doesn’t change the fundamental nature of large organizations, which is that they resist taking on new operational responsibility unless the economics are overwhelming.
There is still a real category of Pseudo-SaaS where the economics have genuinely shifted and the market’s concern about those products is very much. But it is not most of the enterprise SaaS market by revenue.
Consolidation, Not Extinction
The narrative’s second error is assuming that (Pseudo-)SaaS vendors will simply die. The more probable outcome is consolidation. Larger platforms will acquire the thin-layer tools, bundle them into existing contracts, and use AI to maintain them cheaply. This is how commoditization usually resolves. The value accrues upward to platforms with distribution, the middle hollows out, but the functionality survives inside something larger.
AI may actually accelerate this dynamic. Maintenance costs on acquired products fall. Integration work gets easier. The case for the roll-up improves. The outcome for standalone Pseudo-SaaS vendors is ugly, but it’s more likely to be an M&A outcome than an extinction event.
The AI-Native Challenger
There is a second threat, that of a the well-funded AI-native startup proposing to rebuild a Real SaaS category from scratch, which deserves separate treatment.
The pitch is compelling — start clean, move fast, undercut the incumbent on both price and capability.
I believe this narrative gets the economics backwards though. The AI-native challenger’s product is, at its foundation, a wrapper around model capabilities. Those capabilities are not free. They are a variable cost that scales with usage. Every customer interaction, every automated workflow carries a real inference cost that a traditional SaaS vendor, whose marginal cost of serving an additional user is near zero, does not face in the same way. Furthermore, this AI-startup needs to be on the latest models to keep the edge which is also the costliest and without necessarily proven ROI.
The challenger therefore needs to price high enough to cover model costs plus the full overhead of building enterprise-grade integrations, compliance, security, and support from scratch. The incumbent, meanwhile, can add AI features on top of a cost structure already absorbed across its customer base, and can afford to subsidize them for years.
The incumbent also has something the challenger cannot easily replicate: data. Years of customer workflows, transaction histories, and domain-specific signals are precisely the inputs that make AI genuinely useful in enterprise contexts. A generic model is impressive. A model informed by a decade of your industry’s data is a different product entirely.
Incumbents have failed to exploit data advantages before, through inertia or complacency. That risk is real. But the casual assumption that the AI-native challenger holds the structural high ground deserves more skepticism than it usually receives.
What This Means for Valuations
Saying the existential risk to SaaS is lower than consensus suggests is not the same as saying valuations are attractive. These are separate questions, and conflating them is a common error.
Multiple compression on standalone Pseudo-SaaS vendors can be entirely rational even if those vendors don’t disappear — lower growth ceilings, eroding pricing power, and acquisition at compressed multiples are all legitimate reasons to pay less. The market can be right about the price and wrong about the narrative simultaneously.
The more interesting opportunity may lie with Real SaaS platforms caught in the same narrative downdraft despite fundamentally different risk profiles. If the market is applying a broad “AI kills SaaS” discount indiscriminately, platforms with genuine moats — complex workflows, deep integrations, compliance requirements, real switching costs — may be mispriced. Not because AI isn’t a risk, but because the market is asking the wrong question.
The Hardest Part
The build-vs-buy TCO framework sounds clean but is hard to apply with confidence. Costs are uncertain, organizational willingness to take on complexity is hard to measure, and the boundary between Pseudo and Real SaaS will shift as AI improves. A product that is firmly Real SaaS today might migrate in three years.
What I am most confident of is the narrower claim: “will AI kill SaaS?” is too blunt to be useful. The better questions are which products, for which customers, on which timeline, and with what structural outcome. Those don’t have easy answers.
But they are at least the right questions to be asking.