Frontier models are a commodity input. What compounds — and what a competitor can't simply buy — is the workflow underneath: the accumulated context, the proprietary data, the traceable trust. A model is a capability. A workflow is a position.
Every company building an AI product right now is renting the same engines. The frontier models are extraordinary, and they are also a commodity input — a few labs make them, everyone calls the same APIs, and the gap between them narrows with each release. If your product is a thin wrapper around a model, you have built a feature, and features get absorbed. The model you're proud of today is the baseline everyone has by next quarter.
So the real question — the one I ask myself about every product I touch — isn't which model are you using. It's what gets better every time someone uses your product that a competitor can't simply buy.
The answer is almost never the model. It's the workflow: the accumulated context, the proprietary data that compounds, the way the work actually gets done, and the trust that builds up around all of it. A model is a capability. A workflow is a position.
I think about this most concretely through the customer intelligence platform I've been building. Underneath the surface feature — synthetic focus groups, persona-based research — there's a stack that gets harder to replicate at every layer. Ingestion is replicable; anyone can pull in reviews and survey data. Normalization and theme extraction take real engineering, but they're catchable. It starts to get interesting at persona enrichment, where the trick is translating signals into behavior rather than injecting raw facts — a methodology choice that competitors leaning on basic retrieval won't stumble into. And it becomes genuinely defensible at the traceability layer, where every claim carries its provenance, confidence, and freshness all the way through to the final report.
Stack those together and you get a flywheel, not a feature. More connected signals make the personas more realistic. More realistic personas produce better sessions. Better sessions produce reports with traceable evidence. That evidence builds buyer trust, trust drives more usage, and usage pulls in more signal. None of those loops is the model. All of them are workflow. And the whole thing compounds in a direction a well-funded competitor can't shortcut, because they'd have to rebuild not just the software but the accumulated, brand-specific intelligence that the software has been quietly collecting.
This is the same logic that made the last generation of software durable. Salesforce and HubSpot didn't win because their databases were technically superior. They won because the CRM got more valuable the more you put into it, which made leaving expensive. The data was the moat. AI doesn't change that pattern — it raises the stakes, because now the accumulated data doesn't just sit there, it actively makes the product smarter.
There's a hard corollary I've had to internalize: a workflow moat only compounds if the workflow is actually being used the way you designed it. If users connect one data source and never the second, the intelligence layer stays thin and the switching cost stays low. If the evidence isn't persisted, if the value doesn't visibly grow over time, the moat exists on a whiteboard and nowhere else. The architecture being right is not the same as the moat being real. Execution is where defensibility is won or lost, and it's won in the unglamorous places — onboarding that makes the second integration feel obvious, an interface that lets people feel the system getting smarter.
This is also why I've stopped worrying about showing the product. Founders sometimes hide their UI out of fear that someone will copy the magic. But if the magic is a clever prompt, it was never a moat to begin with. If the magic is four layers of compounding, brand-specific intelligence with provenance baked in, you can demo the whole thing on stage and your competitor still leaves with nothing they can use. The visible part is the use case. The platform is everything underneath that gets better with every session.
So when I evaluate what I'm building, or what someone else has built, I've trained myself to look past the model entirely. Show me the loop. Show me what accumulates. Show me what a customer would lose by leaving. That's the moat — and it has almost nothing to do with which engine is under the hood.