Scale Trust, Not Headcount: AI-Powered Semantic Engineering at Southern Glazer's Wine & Spirits

Anoop Kuriakose, Sankar Chidambaram16 Jun 2026

Problem

Southern Glazer's Wine & Spirits, a beverage distributor in 47 U.S. states and Canada, grew through acquisitions. Each acquired company brought its own systems of record, reports and metric definitions. It has over 20,000 active reports across Tableau, Business Objects and several legacy platforms. Every BI tool governs itself, so the same question asked in two tools most likely gets two different answers. Trust in decisions erodes, and AI agents inherit the conflict.

Example from the article: the speakers describe "thousands of dialects" for basic questions: what is a customer, what is revenue, what is on-time delivery.

Solution in brief

The team defines each KPI once in a metrics bank, serves it through a semantic layer built on Databricks metric views and owned data products, and puts no calculation logic in the consumption layer. A business analyst agent and an engineer agent read this shared "enterprise context" and turn a plain-English request into a dashboard, an API or a chat answer.

Questions this case answers
  • How to converge thousands of duplicated reports with conflicting metric definitions into one governed set of definitions that every tool reads
  • How to let business users get a report from a plain-language request while an LLM can only use governed columns and joins
  • How to converge thousands of duplicated reports into one governed set of metric definitions
  • How to make every BI tool and AI agent return the same value for the same KPI

Tools and methods

  • Metrics bank — 600 KPIs with meaning, calculation and rationale
  • Semantic layer on Databricks metric views — the only place where metric logic lives
  • Data products with owners and SLAs — certified data by functional area
  • Analyst agent and engineer agent — write the requirement, then build, test, deploy
  • LLM as judge with confidence score — rates each generated report
  • RAG with FAISS, Unity Catalog memory, FinOps tracker — production fixes
Continue reading — it's free

Google or a one-time email code. No password.