Session

Building Workforce Intelligence With Unity Catalog and Databricks Genie

Overview

Experience	In Person
Track	Analytics & BI
Industry	Retail & Consumer Goods
Technologies	Genie
Skill Level	Intermediate
DOWNLOAD SESSION SLIDES

Mars Inc., with more than 140,000 associates across its confectionery, pet care, and food businesses, set out to reimagine how workforce decisions are made. Leaders needed a single, trusted view of the associate journey to move beyond fragmented dashboards and enable faster and fairer decisions.

Partnering with Tiger Analytics, Mars established Associate 360, a unified workforce intelligence platform. A common data model brings together hire-to-retire data across segments, standardizes definitions, and ensures trust through Unity Catalog. Natural language querying enabled through Databricks Genie lets business users get insights and visualizations instantly.

Impact delivered:

One connected view of the end-to-end associate lifecycle
Consistent, trusted metrics across businesses
Time-based insights for better workforce planning
Faster onboarding of acquisitions without compromising data quality
A shift from static reports to decision-ready workforce intelligence

Session Speakers

Rachel Belino

/HR Data Officer
Mars, Inc.

shachin prabhat

/Vice President and Business Unit Head
Tiger Analytics

Full Summary

Building an AI-ready workforce intelligence foundation at Mars

Enterprises cannot get to trustworthy conversational analytics by adding a chatbot to fragmented HR data. Mars reframed people analytics around a governed, unified data foundation that can answer natural language questions across the associate lifecycle. The effort paired Unity Catalog, Genie, and a purpose-built common data model to support secure, accurate, role-aware experiences on the Databricks Platform.

The conversation reinforced a broader point: the visible AI layer only works when the underlying data model, governance, and business context are meticulously engineered on a modern Databricks Platform. Organizations pursuing this path can find a practical frame in how AI is transforming data analytics across industries, where the lesson is consistent: reliable intelligence requires an engineered foundation, not just a capable model on top.

The result is an Associate 360 architecture that normalizes core HR, recruiting, performance, compensation, learning, and survey data, enabling consistent metrics and scalable onboarding of new sources, including acquisitions. A synthetic demo showed how users can start with a vague prompt, then iterate to deeper insights, with the agent inferring defaults, maintaining context, and returning both narrative and tabular outputs.

From medallion pipelines to an Associate 360 model

Mars had already standardized ingestion with a bronze, silver, and gold medallion architecture and built reusable datasets. What was missing was a transversal view of each associate across domains, normalized into a single governed structure that agents and humans can reason over.

The team designed an Associate 360 common data model that abstracts upstream systems and stabilizes metric definitions. A key construct is a time-phased link table that stores surrogate keys down to the day, allowing agents to anchor questions to the right slice of an associate's journey. That design reduced hallucinations and improved query performance during conversational interactions.

Governance and context as first-class requirements

HR data sensitivity made governance a non-negotiable design constraint. Unity Catalog became the system of record for lineage, access controls, and discoverability, enforced by a simple rule: if it is not cataloged, it does not exist. Data quality checks were implemented to give stakeholders confidence in the outputs that inform decisions.

Context was engineered into the model through semantic views and context-centric tables that translate internal terminology, codes, and definitions into language agents can use. Instructions provided to Genie, combined with these curated views, helped prevent misinterpretation of business terms and ensured users only interact with data appropriate to their role.

Designing for agents, personas, and secure conversational analytics

The solution was built for agent consumption from day one. Persona-based Genie spaces were created for roles such as learning managers, recruiters, and HR business partners, each with tailored instructions and guardrails. Unity Catalog permissions, combined with these scoped agent spaces, restricted access to sensitive attributes like compensation and ratings.

A synthetic demo illustrated the interaction pattern rather than a single answer. A broad prompt about attrition led to sensible defaults, trend detection, clarifying follow-ups, and drill downs by country and performance segment. The experience shifted routine exploration from analyst ticket queues and siloed dashboards to governed conversational analysis, while still allowing lightweight visualizations and curated dashboards where needed.

Implementation lessons and what is next

Several practices stood out. Put AI on top of the data model as it is built to validate whether the structure answers real user questions, and adjust the design accordingly. Invest heavily in business context, semantic layers, and clear agent instructions, since context is what unlocks accuracy. Treat visualization quality, persona personalization, and change management as core workstreams, not add-ons.

Dashboards remain part of the toolkit, but the primary exploration surface is moving to conversational analytics over governed views. The roadmap focuses on moving beyond "what" to "why," which depends on richer ontology and multi-agent capabilities. For predictive and especially prescriptive HR insights, the team emphasized close partnership with legal.

Under the chatbot, the engineered data layer does the heavy lifting

Mars's journey shows that conversational workforce intelligence is only as reliable as the normalized, governed, context-rich foundation beneath it. The polished Genie exchange is the tip of the iceberg; consistent answers come from the common data model, Unity Catalog governance, and semantic context that sit below the surface.

FAQ

Reusable datasets were not stitched together across the associate lifecycle, risking duplication and inconsistent metric definitions. A common data model provides a stable abstraction that agents and users can query confidently, even if upstream systems, such as HRIS or survey tools, change later.

The link table is a time-phased structure that stores surrogate keys for each associate's journey at day-level granularity. Genie queries it first to anchor intent to the correct time slice, then fans out to supporting tables. The approach reduced hallucinations and improved performance in natural language interactions.

Unity Catalog enforces lineage, access controls, and governance for every asset. Persona-scoped Genie spaces add guardrails so users only see data and questions appropriate to their role, preventing exposure of sensitive attributes like compensation or ratings.

No. Many users still want dashboards, and curated views remain in place. The shift is toward offering conversational analytics as the primary exploration layer, with on-demand visualizations, while retaining dashboards where they add value.

The team is moving toward explaining "why," and eventually toward predictive and prescriptive insights. Progress depends on richer ontology, multi-agent capabilities, refined persona experiences, and, for HR use cases, close collaboration with legal before surfacing prescriptive recommendations.