BLOG

Every quarter, the foundation models get better, and a familiar anxiety ripples through vertical AI startups: if GPT-class models can now do what we do, what’s left?

It’s the wrong question. The capability of the underlying model was never the moat. As LLMs commoditize raw intelligence, defensibility moves to where it always lived in durable software businesses—into the workflow, the data, and the depth of the customer relationship.

Here’s the thesis we share with the founders we back.

Build moats on moats. A single defensible feature is a feature, not a company. The strongest vertical AI businesses stack advantages: proprietary data that compounds, regulatory expertise that’s expensive to acquire, switching costs that grow with usage, and distribution that’s hard to replicate. Any one of these is contestable. Layered together, they become a fortress. And don’t underestimate the power of great UI/UX.

Bundle the platform; don’t sell the modules. The moment you price value module-by-module, you invite a race to the bottom. Each module becomes a discrete line item a well-capitalized competitor can underprice to win the logo. Bundle multiple value propositions into one platform and you sell an outcome, not a SKU—and you make unit-economics warfare impossible to wage against you piece by piece.

Climb the ladder: cognition, then action, then record. Most vertical AI companies don’t start as the system of record. They start as a system of cognition—taking a customer’s proprietary data, running it through a mix of their own small models and external LLMs, and producing recommendations, savings, or revenue the customer can actually measure. Once you’re generating provable outcomes, you’ve earned the right to act on them. That’s the system of action: proprietary automations that execute on the recommendations produced by the system of cognition. And once you both think and act on a customer’s reality, you become the system of record—the source of truth they can’t rip out.