Model for the token, not the table

Modeling Gong call transcripts directly in the warehouse with dbt, instead of hitting Gong's API on every AI call, cut token costs by 20x. The takeaway: for AI workloads, model the data at the token level, not just the table level.