Problem
Mars (170K+ associates, as it calls employees) kept its people data in fragmented, domain-specific dashboards. The same source data was processed again and again in reusable datasets, and nobody had joined them into one model. One metric could come out as three numbers, every leader question went through an analyst, and acquisitions became multi-quarter integration projects. The team wanted a conversational assistant for business users. An agent over unstitched tables answers inconsistently. HR data is also sensitive: compensation and ratings must stay hidden from the wrong people.
Example from the article: when the team first put Genie on its cleaned-up data model (last summer), the results were not what they wanted, until they added a link table.
Solution in brief
Mars and Tiger Analytics built "Associate 360": a common data model (CDM) over all HR sources, a day-level link table that Genie queries first, semantic views with instructions for business context, and one Genie space per persona. Unity Catalog gives lineage and access control. The assistant queries governed views, not raw tables.
Questions this case answers
- How to make a natural-language data assistant return consistent answers over a common data model built from many sources, using a link table that routes each question to the right tables
- How to onboard data from acquisitions into a governed model without a multi-quarter integration project, while hiding sensitive records from the wrong users
- How to make a natural-language data assistant return consistent, non-hallucinated answers on a warehouse
- How to design a common data model and semantic layer that an LLM agent can query reliably
Tools and methods
- Common data model — source-agnostic tables, standardized definitions
- Link table — time-phased spine of the associate journey at day grain
- Semantic views and Genie instructions — business context
- Genie spaces per persona — for example learning manager, recruiter
- Unity Catalog — lineage and access control
Google or a one-time email code. No password.