Anthropic + Adidas + Databricks: Unlocking 400 Hours of Productivity Weekly with Conversational AI

Hosea Kidane, Vikalp Yadav17 Jun 2026

Problem

Marketing and CRM stakeholders keep asking analysts for "one more cut" of the data: why net sales dropped in Germany, why conversion is not picking up. Warehouses, self-service dashboards and a governed semantic layer did not end this loop. At adidas, more than 50% of the topics analysts worked on in the past year were such methodical, contextual questions, which kept analysts from harder problems. An LLM answering them must earn trust.

Example from the article: a marketer asks for conversion and open rates of all running North America campaigns, then a daily trend, then asks to remove the impact of the football World Cup.

Solution in brief

adidas built Diva, a conversational analyst on Databricks with Claude, on top of a governed data stack. Each question passes authentication, context, planning, parallel Genie queries and validation before an answer appears. Anthropic adds how to pick and tune the model: an eval loop re-run at every release.

Questions this case answers
  • How to build a conversational analytics agent that answers why-questions with governed and validated numbers instead of routine follow-up cuts of the data
  • How to tune the model and the surrounding harness of an analytics agent with an eval loop and gate it in CI with a question benchmark
  • How to build a conversational analytics agent that answers why-questions with governed, validated answers
  • How to stop analysts from spending most of their time on routine follow-up cuts of the data

Tools and methods

  • Claude (Opus 4.6 planner) — splits a why-question and writes the answer.
  • Databricks Genie, Unity Catalog — parallel queries over certified metric views with row- and column-level security.
  • LangGraph orchestrator — plans before executing.
  • Semantic layer and KPI governance — owned, certified metric definitions.
  • Anti-hallucination engine — checks schema, grain and time logic.
  • MLflow and RAGAS — tracing of every node.
  • 131-question benchmark in CI — merge gate at 95% accuracy.
Continue reading — it's free

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