Workflows are the new IP

The attention in AI goes to model training - benchmark scores, architecture debates, whether open weights catch up - but the firms that drive the economy do not train models and never will. Their competitive asset is not in the weights but in the workflow: the sequence that routes a task through business logic, pulls proprietary data, applies institutional know-how, and produces output a competitor cannot replicate even with the same model.

Most firms are incapable of training a frontier model because the cost is measured in hundreds of millions of dollars and the talent pool is a few thousand people. They will call a model through an API, and the API will deliver roughly equivalent capability regardless of provider. The model itself is interchangeable in a way the workflow is not.

What makes a workflow hard to copy is the business logic the calls are embedded in - the rules, checks, and decision paths accumulated over years in a domain. It is also the proprietary data the workflow draws from - customer histories, internal documents, error logs, the record of past decisions that exists nowhere else.

A firm that builds workflows around its domain expertise is building the layer that makes the model useful and that a competitor will struggle hardest to reconstruct.

The model will improve whether you optimize for it or not, and that improvement belongs to every customer equally. The workflow improves only when you feed it your logic and your data, and that improvement belongs to you alone. A moat is just an asset a competitor cannot buy, and no model vendor sells one.