BASF Coatings Scales Governed Multi-Agent AI with MAUI on Databricks

BASF Coatings built MAUI, a governed multi-agent interface on Databricks, and its three first pilots are in production. According to the speakers, value grows from 0.5 million to 1.5 million annually as adoption scales, and more than 20 agents are in the pipeline. What used to take months now takes days, and the head of the division's data and AI office in central IT said, "Scaling no longer means complexity, it means leverage." BASF Coatings operates 70 sites in 50 countries with around 11,000 employees. It has worked with Databricks for six years and has data stewards in 16 domains who own data products. The problem was not missing data but the lack of an everyday access layer that turns governed data into usable answers. Knowledge was spread across tables, documents and expert knowledge, central teams stitched data and context together manually, and shadow AI risk grew. Marketing, supply chain and operations had each started building the same stack independently, with a front end, orchestration, data integration, monitoring and governance, which led to duplication, fragmented architectures and weak governance. BASF Coatings then moved from thinking in use cases to thinking in capabilities. MAUI uses a five-layer architecture. A single React application serves business users, data stewards and admins. A FastAPI backend runs chat synchronously with streaming and handles longer processes such as RAG indexing asynchronously through Databricks jobs. Agents are built dynamically from configuration at runtime as a graph that orchestrates LLM reasoning, tool use and retrieval, using Genie Spaces or SQL for structured data and knowledge bases for unstructured data. The native Databricks layer provides model serving, vector search, Unity Catalog and identity management, with requests executed on behalf of the user. Lakebase tracks conversations, feedback, cost and agent configurations. Governance is federated: each of the 16 domains has at least two data stewards who are responsible for data quality, availability and compliance, and together with domain experts they become agent owners. A self-improving cycle analyzes direct and indirect feedback from conversation traces and suggests improvements to agent owners. The roadmap includes document generation, proactive notifications and connecting SAP, Salesforce and Office through MCPs.