I tend to think of it as a representation of feature sets in products. I know I might want feature x or y, but I might not know how frequent the use case is or how it might fit into the scalability aspect of my schema and architecture. So the ability to slow roll something new into a well defined (SQL) schema by amending it with a JSONB field property is the perfect time/complexity trade off. Once I have a clear understanding of how a user might actually use it, and if it offers value, it becomes an official part of the schema.

There are other parts of the process to determine where a new property fits for me( for example it might actually need an entire state store if its own), but in general this is where I begin.