The moat isn’t the model

There’s a quiet panic underneath a lot of AI products right now, and it goes: if the model gets good enough, doesn’t it just do what we do? Fair fear, wrong layer. The model was never the moat, and treating it like one is how you end up with a gorgeous demo and a business that evaporates the week a new checkpoint ships.

The thing to internalize early: everyone building on frontier models is renting the same engine. You, your competitor, and the well-funded team that hasn’t launched are all calling roughly the same three APIs. Whatever cleverness you wrap around a prompt, someone wraps a slightly different cleverness around the same prompt by Friday. The weights are a commodity input, and commodities defend nothing.

Where defensibility actually lives

If not the model, then what? The honest answer is boring, which is why it works. It lives in the layers a model can’t generate for itself:

• Proprietary data it has never seen and can’t buy: your users’ behavior, your measurements, the outcomes of actions taken inside your product.
• The action loop: not “here’s an answer,” but measure → decide → do → observe → adjust. The model gives you the middle. The loop is the product.
• Workflow gravity: the accumulated state and habits that make leaving expensive even when a shinier tool shows up.
• Trust: the unsexy compounding asset. People let you touch things that matter only after you’ve earned it, and that can’t be prompted into being.

Notice none of these get weaker as the model improves. They get stronger. That’s the whole tell: does a better model help you or threaten you? If a model upgrade is a threat, you’re selling the model. If it’s a gift, you’re selling something the model makes more valuable.

The right question isn’t “what can the model do?” It’s “what is true about my users that no model will ever know unless it goes through me?”

The wrapper insult, reconsidered

“It’s just a wrapper” became the favorite dismissal of the last two years, and like most dismissals it’s lazy. Stripe wraps card networks. Every SaaS wraps a database. The wrapper is usually where all the user-facing value lives. The only question is whether it’s thin or thick.

A thin wrapper passes the model’s output straight through and takes a margin. Genuinely fragile. A thick wrapper owns the problem end to end: it knows the user’s context, takes actions on their behalf, closes the loop and learns. Thin wrappers die when the model improves. Thick wrappers get better. The word isn’t the problem; thinness is.

What it means on a Tuesday

1. Stop competing on model quality. You’ll lose to the labs, and you’d be spending scarce effort on the one layer you don’t control.
2. Hoard the data only you can see. Every action inside your product is a proprietary signal. That’s the flywheel that outlasts any single model.
3. Own the outcome, not the answer. Anyone can answer. Very few products actually move the number the customer cares about. Be the one that does, and the model underneath becomes a detail.