There is a popular narrative that the AI startup race is decided by who builds the best model. That narrative has rarely been true in technology and is becoming less true now. Model performance is converging. APIs are commoditising. The most durable AI businesses are not the ones with marginally better models. They are the ones with privileged access to customers, data, and workflows.
Distribution is the first-order question
Most enterprise buyers will not switch from incumbent vendors for a 5% performance improvement. They will switch for a 10x cost improvement, a clearly superior experience, or because the new offering is bundled into a system they already trust. Each of these is a distribution question, not a model question.
Workflow ownership is the second
Owning a workflow means the customer's daily work cannot be done without you. AI applications that own workflows accumulate proprietary signal, become harder to displace, and earn pricing power. This is why the most valuable AI businesses tend to be vertical: legal, healthcare, security operations, financial close, customer service — domains where the workflow is deep and specific.
Proprietary data is the third
Data moats compound when they are generated by the product itself. A startup whose customers continuously contribute labelled examples, corrections, and feedback ends up with a dataset its competitors cannot buy. The moat is not the dataset itself — it is the loop that produces it.
Implications for founders
Build the model that gets you to the customer. Then build the customer relationship that gets you to the workflow. Then build the workflow ownership that gets you to the data. Each stage funds the next. Founders who skip stages — and try to build defensible data moats with no workflow, or workflows with no distribution — burn cash without compounding.
Implications for investors
The most under-appreciated AI investments are vertical applications with deep workflow ownership in regulated industries. They look unglamorous, but they have the moat structure that compounds. Pure model bets, in contrast, increasingly look like infrastructure plays that require infrastructure-scale capital to win.
Key takeaways
- Model capability is converging; distribution and workflow are diverging.
- Vertical AI in regulated industries has the strongest compounding moats.
- Build the loop that generates proprietary data, not the dataset itself.
- Founders must sequence moats: distribution → workflow → data.