Ecolab says a compliance report that once took two weeks of pulling data from nine siloed systems for a single customer location now completes in under two minutes with its Retail Intelligence application, which launched in mid-April 2026 and served hundreds of North American locations at launch. An FDA food code question that sent managers digging through a 700-page PDF for hours now returns a cited, plain-language answer in seconds, and the conversational agent supports approximately twelve languages at about 98% accuracy.
Josh McCoy, Product Manager for Retail Intelligence at Ecolab, said: “Our frontline staff across 600 locations get cited, plain-language answers from a 700-page FDA food code in seconds, in whatever language they're most comfortable with.” Nicholas Dylla, Technical Lead at Ecolab, said: “We had nine different data sources, nine different intelligences, and no way to see the full picture for a single location.”
The nine sources are audits, health inspections, pest IoT telemetry, checklists, chemical usage logs, weather feeds, Yelp reviews, CDC neighborhood data and the FDA food code. The application is a native Databricks App with Lakebase Postgres as its transactional backbone. All sources flow through Lakeflow and Spark Declarative Pipelines into a governed lakehouse under Unity Catalog. Through Databricks Foundation Model APIs, Ecolab serves Claude Sonnet for complex reasoning, Claude Haiku for summarization and Gemini for image analysis.
The system follows a Multi-Agent-Supervisor pattern orchestrated through Databricks Workflows. A Coordinator Agent breaks each question into subtasks for specialised sub-agents: one retrieves FDA passages via Vector Search, another queries structured compliance data through SQL and Unity Catalog Functions, and a third pulls pest telemetry from an external MCP server. A Response Agent assembles a cited answer and persists the interaction to Lakebase. Short-term memory carries the last ten conversation turns in the prompt, with Claude Haiku 4.5 summarizing earlier turns every three turns, while Claude Sonnet 4.6 maintains a long-term per-user profile. Five Judge LLMs evaluate every interaction, MLflow traces every execution path, and ai_query() applies Claude across thousands of records for offline batch workloads.
Next, Ecolab plans to add MCP-powered automated actions such as pest inspections, chemical reorders, food safety norms and work orders triggered from the chat interface.