How do you weight the value of semantic layers when all of the labs are chasing shell usage benchmarks like TerminalBench?

Aside from the enterprise stuff like consistent metrics I’m not convinced semantic layers improve agent performance. Case in point is Snowflake Analyst has been routing 90%+ of queries to traditional SQL as opposed to their own semantic SQL dialect.

A semantic layer is a concept from 2018 in the BI world. Metadata is one thing but semantic layer implies an abstraction from physical data with lossy translations.