Session

Beyond Dashboards: How Grupo Bimbo Delivers Insights with Databricks AgentBricks

Overview

Experience	In Person
Track	Artificial Intelligence & Agents
Industry	Manufacturing, Retail & Consumer Goods
Technologies	Genie
Skill Level	Intermediate

Commercial operations at Grupo Bimbo generate thousands of transactions every day across sales routes. While dashboards highlight valued insights, commercial team are constantly on the move and don’t have time to explore BI tools.

Grupo Bimbo rethought insight delivery using Databricks Agent Bricks — moving from “look at dashboards” to “ask and receive answers.” For supervisors, we implemented AI/BI Genie as a conversational AI layer over sales data, enabling natural-language questions like “Which clients were missed this week?”— no dashboards required.

For executives, we built an AI query agent powered by Claude that automatically reads multiple KPIs and summarizes insights, generating reports delivered directly by email.

All agents run on Databricks with Unity Catalog ensuring secure data access.

We’ll share the real architecture, prompt design choices and lessons learned and how this approach accelerated decisions, expanded insight adoption and improved sales process efficiency.

Session Speakers

Esteban Ramirez

/Data Science Manager
Grupo Bimbo

Mauricio Imenez

/Global Director Data Analytics
Grupo Bimbo

Full Summary

How Grupo Bimbo rethought sales insights with machine learning, Genie Agents, and AI functions

Dashboards did not change behavior for Grupo Bimbo's field-led sales organization, despite accurate indicators and clear revenue opportunities. The problem was delivery, not data: insights were not reaching the right person, in the right format, at the right moment across 39 countries, 1,700 sales centers, and 54,000 delivery routes.

The team reframed the solution around prioritization and usability, combining machine learning, conversational access, and automated executive reporting on the Databricks Platform. The result is a system that ranks where to act, answers natural language questions in the field, and sends concise, personalized summaries to leaders. More context on the approach appears in Databricks AI/BI and agentic analytics capabilities, and the shift toward conversational, role-aware delivery reflects a broader move described in how leading companies are delivering trusted AI-powered self-service analytics.

Why dashboards missed the mark in the field

Supervisors spend most days on routes with customers, not at desks. Leaders also reported an overload of indicators with no clear starting point. Insights existed, but the format did not match how people work, so adoption lagged and opportunities went unaddressed.

The team shifted from a single dashboard destination to a set of delivery mechanisms: a ranked risk score that focuses attention, a Genie Agents experience for on-the-go questions, and an automated "intelligence report" that highlights the few metrics each region should act on each week.

Route-level anomaly scoring that respects context

Anomalies are contextual, so a single global model would blur urban and rural realities. The team trained hundreds of isolation forest models in parallel, segmented by sales center, pre-sales type, and channel. Each model learns what "normal" looks like in its own context, avoiding one-size-fits-all thresholds.

Risk rolls up from ticket to route. Aggregation prevents users from auditing every transaction, and a weighted approach separates anomalous from routine activity so a handful of extreme events are not hidden by many normal ones. The formula is intentionally intuitive, enabling sales center teams to understand and challenge the score rather than treat it as a black box.

Delivering insights through conversation and automation

For field use, a Genie Agents experience connects to the curated opportunities dataset. Supervisors ask natural language questions such as which clients to visit this week, which routes need supervision, or where sales were recorded far from client geolocations. Results return as tables, charts, or maps, aligning to how supervisors make decisions on the move.

Three ingredients underpin reliable answers: a cleaned, route-level dataset instead of raw tickets, complete table and column metadata to bridge internal terminology, and a custom prompt that encodes audit rules and indicator definitions. For executives, AI functions in SQL generate concise, personalized weekly summaries per region with indicators and practical suggestions. The pipeline converts Markdown to PDF with ReportLab and emails it automatically, with no manual steps.

Behavior change and measurable impact

Deployments show a 17 percent reduction in missing sales opportunities, and every manager in those regions receives the automated reports. Supervisors no longer need to wait for their single office day to investigate issues, since prioritized questions and answers are available in the field.

User requests for enhancements have increased, a practical signal of engagement that dashboards alone rarely produced. Lessons include designing for actual work patterns, personalizing outputs by role and region, investing in prompt design, and cleaning data before invoking an LLM. Generative interfaces amplify, rather than replace, classical machine learning.

Design analytics around moments of use

Value emerged when insights fit the rhythm of sales work. By pairing contextual ML scoring with conversational access and automated summaries, the program shifted from well-built but underused dashboards to a trusted, action-oriented system.

FAQ

Supervisors work primarily in the field and had limited time or habit for dashboard analysis. Leaders also faced too many indicators without clear prioritization. The delivery format did not align with day-to-day decision moments.

Anomalies depend on local context. Segmented models by sales center, pre-sales type, and channel learn region-specific norms, avoiding misleading global thresholds that treat dense urban routes and sparse rural routes the same way.

Accuracy depended on three elements: a curated route-level dataset, comprehensive metadata to resolve internal jargon, and a custom prompt that captures audit rules and indicator definitions. Together these guide the model to unambiguous SQL and trustworthy answers.

An AI function runs in SQL against the opportunities data, writes a concise Markdown summary per region with key indicators and suggested actions, then the pipeline converts it to PDF and emails it on a schedule. The process is fully automated.

No. Isolation forest models do the analytical work of detecting and ranking opportunities. Generative AI improves consumption and actionability through conversational access and personalized reporting. The two approaches complement each other.