Found this video interesting (especially as someone who hasn't been a user of BI tools for 5+ years). Some interesting implications from this Rippling case study (all of these are my opinions):

- For agents, choosing a tool / MCP and choosing what data to retrieve start to look like the same problem: both are pieces of context needed to answer a question. The old boundary between "business logic" and “data” gets blurrier.
- Also interesting: Rippling’s data team still manually inspects agent queries and creates new rollup tables when patterns emerge. Where are the materialized view database fiends -- they would find this very fascinating!!
- I could be wrong but: it feels like trad ELT is mostly push based: i.e., humans anticipate questions and build tables / dashboards. Agents make BI much more pull based than in the past: the question arrives first, then the system figures out what data, transformations, and tools it needs.
- And perhaps one reason why existing semantic layer tools aren't great yet: not all past queries are equally useful signals. Perhaps we cares more about which queries led to useful business actions, not just query history in aggregate.