AI Won't Eat SaaS

Nathan Gunn, MD‍

‍August 27, 2026

‍

I hear some version of "AI will eat SaaS" a lot. It comes up in nearly every enterprise procurement review. There are two issues in this assertion. First, it conflates four different arguments. Second, only one of those arguments threatens whether a software company survives, and that one fails. Why? Because making an LLM do the job of a SaaS company means building the software you were trying to skip.

Here are the four arguments, in order of how much damage each would do to a SaaS business if true.

• One: AI is a feature inside your software product. True, and true of every software company. Every serious platform has LLM calls running inside it now. Having AI in your product is necessary but nowhere near sufficient on its own.
• Two: AI makes your product cheaper to build, so new entrants undercut your price. Also true, but every software company gets the same benefit at the same time, so lower costs don't decide who wins, for two reasons.
  
  • First, price and margin are different: the cheaper tooling that lets a new entrant undercut your price has also lowered your costs; so, price may come down, but cost comes down with it and margin can hold.
  • Second, price does not determine which vendor wins in enterprise healthcare software. For example, hospitals pay a premium to implement Epic and then keep paying, because the software does what they need AND because leaving costs them more than staying does. That pattern repeats across enterprise software generally.
• Three: AI eliminates the unit you bill for, because agents do the work that per-user licenses used to cover. A very SaaS’y argument, and perhaps worrying wherever seats drive material revenue.
  
  • Healthcare software mostly isn't sold by the seat. The unit of economic value is the patient/case, and that's what gets priced. Value-based pricing exists. Seats were a choice, not a law, and if you're creating value, you'll find the unit that carries it.
• Four: a client substitutes an LLM for the product entirely. Feed the records to the LLM and skip the software. This is the version that could end a software business and the one that people usually mean.

Why doesn’t argument four hold up?

Ask what an enterprise in healthcare or finance actually requires of the software. Part of it is what the answer has to be: identical on every run, traceable back to the source document, right at volume. The rest is where the answer has to live: inside a workflow people already use, behind security and compliance review, in reporting somebody can audit. AN LLM on its own doesn't produce the first and doesn't provide the second at all.

(As an aside, anyone who has built an analytic pipeline knows this; but it's less visible if your AI experience is a chat window that answers questions quite well).

Moreover, on the first requirements list, what the answer has to be, capability and cost aren't separate objections, they're the same one. The only way to push an LLM toward that deterministic standard is to spend more tokens on every record; moving from “usually right” to “six-sigma right” to “always right” drives the spending vertical before the reliability arrives. And token spend and reliability are asymptotic: each increment of reliability costs more than the one before, and the curve approaches the enterprise standard without ever reaching it. Arriving costs infinity, and a discount on infinity is still infinity. That holds for the inexpensive Chinese LLMs too.

Ordinary software has no such curve. Code written to produce a given output produces it every time and running it again costs nothing at the margin.

The second requirements list, where the answer has to live, isn't a cost problem at all: an LLM doesn't do those things at any price. So, you reach for databases, infrastructure, middleware and human review to hold the LLM to the standard, and at that point you have built a software product with AI inside it. Which is our very first argument: AI is a feature inside your software product.

‍

Here is a tangible example: Claude is not going to eat Salesforce. Salesforce is going to put (commoditized, great) AI inside its product and be just fine.

This isn’t my first SaaSpocalypse

-Marc Benioff, All-In Podcast (I think), a few months ago

If you've tried to push an LLM to production reliability, what did it cost you to get from mostly working to reliable enough to ship?