Disrupted or dead — did you sell the thing AI now gives away?

A grim number from this month: more than 220 companies that once crossed a billion-dollar valuation are now "fallen unicorns," worth less than half their peak — names like Glossier and The Farmer's Dog. Startups that last raised in 2021 are down 68% on average. A former DoorDash engineering leader summed up the mood: workflow-driven enterprise SaaS will be "disrupted or dead in the next decade."

Now hold that next to a second fact that sounds like its opposite: roughly 80% of the new AI startups — the "wrappers" — are expected to fail too. So the pre-AI companies are dying and the AI companies are dying. That looks contradictory until you see what kills both — because it's the same thing, and it's a test you can apply to whatever you're building right now.

Two opposite companies, one fatal mistake

The dying pre-ChatGPT startups mostly sold labor and workflow: a tool plus a team to run a process, often priced per user. Their advantage was that doing this used to take a lot of people. AI erased that advantage — 50 engineers now do what took 500, and a process you could automate is a process someone will. The moat was the difficulty, and AI made it easy.

The dying AI wrappers made the mirror-image bet. Their product was a thin layer over a model — a nice interface on top of ChatGPT or Claude. The problem: that's a feature, not a company, and the model provider can add it next month, for free, to everyone. Worse, the wrapper pays the token cost on every call, so its margins sit near 40% while real software runs at 70–80%. One sold the labor AI replaced; the other sold the model AI commoditized. Opposite directions, identical mistake: both built their whole value on the thing AI now provides.

The survival test is one honest question

Strip away the jargon and there's a single question that sorts the survivors from the casualties:

If a foundation model added your main feature next month, or if five engineers with Claude could rebuild your product over a weekend — what would be left?

If the honest answer is "not much," you didn't have a company; you had a head start, and AI just ate it. If the answer is "the part that's actually hard to copy" — that's your real business, and everything else was scaffolding.

This is the model-was-never-the-moat argument pointed at your own survival. The defensible things are the ones AI doesn't hand to your competitor: proprietary or regulated data nobody else has, deep integration into a workflow that's painful to rip out, distribution and trust you've earned, a system that takes real action in a specific domain. The fragile things are the ones AI now provides on demand — generic features, undifferentiated labor, a wrapper over a public model.

Why per-seat SaaS specifically is exposed

There's a sharper edge worth naming. A lot of the dying companies charged per human seat — and that model is structurally threatened by agents. If the "user" of your software is increasingly an agent acting for a person, "charge per employee who logs in" stops describing reality. This is the same shift as when your customer becomes a bot: the assumptions your pricing and product were built on quietly stop holding. If your revenue scales with the number of humans doing a task, and AI is shrinking the number of humans doing that task, you've tied your growth to a number going down.

What to actually do

You don't need to be a founder for this to matter — it applies to a product, a feature, even a career:

• Find the part AI can't hand to a competitor, and double down on it. Proprietary data, hard-won workflow integration, real distribution, a regulated domain. If you don't have one yet, getting one is the work.
• Stop defending the commoditized layer. The generic feature, the thin wrapper, the manual process — don't pour effort into a wall AI already walked through. Move up to the thing around it.
• Re-examine anything priced per human. If agents shrink the headcount doing the work, per-seat revenue shrinks with it. Price the outcome or the value, not the number of logins.
• Assume your current edge has a short shelf life. "Hard to build" is no longer a moat when building got cheap. Ask what stays valuable after everyone can build it.

The bottom line

"Disrupted or dead" sounds like a story about other people's failed startups, but it's really a diagnostic. The companies collapsing this year — old-guard and AI-native alike — share one trait: their value was the thing AI now provides, whether that was human labor or a raw model. The survivors are the ones whose value lives in what AI can't just hand out: the data, the integration, the distribution, the domain.

So ask the uncomfortable question about your own work, honestly: if the model got your feature for free next month, what would you still be worth? Answer that, build that, and the disruption that's killing everyone else becomes the tailwind that clears your competition. Dodge the question, and you're just a head start waiting to be erased.