Using probabilistic LLMs for deterministic data queries creates a high risk of hallucinated answers, which is unacceptable for critical business metrics.
While feeding raw schemas to an LLM fails in complex environments, the semantic layer approach introduces the problem of BI lock-in, trapping logic within specific tools.
Even with a universal semantic layer, AI agents are often constrained to predefined queries, failing to handle complex, ad-hoc investigations known as 'rabbit hole' questions.
The proposed solution allows an AI agent to first query the semantic layer API and then fall back to using it as a blueprint to write custom SQL for complex questions.
To enable powerful and reliable AI agents for the entire organization, data teams must first invest in building and exposing a robust semantic context.
Jobs that call for the skills explored in this talk.