What separates AI companies with real moats from the ones that just have wrappers

The "wrapper" critique has evolved. In 2023, it was a blunt instrument: any product built on top of an OpenAI API call got dismissed as defensible. That turned out to be wrong in interesting ways. Some wrappers became real businesses. Some "deep tech" AI plays with impressive model architectures had no distribution and died quietly. The original framing missed something important.

VCs have gotten sharper about this. The question is no longer "are you a wrapper?" It's "where does the defensibility actually live, and can it hold as the underlying models improve?"

The three moat types holding up under pressure

1. Proprietary data loops

The strongest AI moats in 2026 are built on data that improves with usage and that competitors can't easily replicate. This isn't about having a big dataset at training time. It's about a flywheel where the product gets meaningfully better as more customers use it, generating feedback signals that no one else has access to.

The test: if your underlying model provider released the same capability tomorrow, would your product still be better? If yes, you probably have a data moat. If no, you have a feature.

2. Embedded workflow switching costs

Several "wrapper" companies that looked fragile in 2023 turned out to be highly defensible because they embedded themselves into critical workflows. The AI capability was almost incidental. The moat was the switching cost of ripping out a tool that now lives at the centre of how a team works every day.

This is an underrated source of defensibility, especially in B2B. If your product becomes the system of record for something, or if it creates outputs that feed into other systems, you've built lock-in that doesn't depend on having better AI than your competitors.

3. Domain-specific fine-tuning that models can't replicate out of the box

General-purpose models are getting better fast. But there remain domains, highly specialised professional knowledge, regulatory context, proprietary terminology, niche workflows, where a fine-tuned model on curated domain data consistently outperforms a general-purpose frontier model, even with perfect prompting.

The key word is "consistently." If a general model can match your fine-tuned model 80% of the time, that's not a moat. If it fails in exactly the 20% of cases that matter most to your customers, you have something real.

What the best wrappers actually got right

The wrapper companies that built durable businesses didn't survive because the models stopped improving. They survived because they figured out distribution, built genuine user habits, and created switching costs before anyone noticed. They used the AI capability as a wedge to get in the door, then built the real moat while incumbents were busy dismissing them.

The lesson for founders and investors alike: the moat question matters, but it's a second-order question. The first-order question is whether you have a real problem, a real customer, and a real reason to exist. The moat protects value that has already been created. It can't create value on its own.