Bayernwerk AI Agents, a generative AI field operations assistant built on the Databricks platform, is in a live pilot with 60 users (for example from field operations, HSE and technical policy). After the pilot, Bayernwerk plans to scale it to approximately 4,000 business users, and the story says the goal is to reduce time spent searching for documents, saving potentially thousands of hours collectively each year.
Bayernwerk Netz, a German grid operator and E.ON subsidiary that supplies roughly seven million people and manages nearly half a million decentralized generation plants, partnered with E.ON Digital Technology on the project. Technical specifications, HSE policies and standard operating procedures were spread across several systems, so staff spent excessive time searching for documents or verifying versions. "We wanted a way to quickly surface new policy revisions and link back to the official documentation without making people dig through folders or emails," said Iman Soltanzadeh, Data Science Manager at E.ON Digital Technology. The assistant had to understand everything from voltage procedures to PPE requirements in the German regulatory context, and it supports the company's Vision Zero goal of eliminating serious work-related accidents.
The team uses Delta Lake to store structured and unstructured data ingested from SharePoint, which holds SOPs, policies and manuals, and from MAQSIMA, which holds equipment specifications and worker safety manuals. Databricks Pipelines extract, clean and normalize the documents and trigger refresh jobs to keep the retrieval augmented generation pipeline current. Unity Catalog enforces source-level permissions, so a field technician's query only draws on documents that person can access. Databricks AI Search retrieves relevant chunks for field manager queries, and the assistant has begun to break down complex questions such as the safe procedure to replace a 20kV cable and to gather inputs such as cable specification PDFs. Agent Bricks AI Gateway routes queries to lightweight models for routine retrieval or to more advanced LLMs for multi-step safety reasoning. Developers test new agents with Databricks Notebooks and Databricks Connect.
Soltanzadeh said the validated pilot will be followed by a first rollout to around 600 field workers before wider groups across the company are added. The team is exploring Model Context Protocol (MCP) with Databricks to extend the agents into action-taking workflows, and plans to add environmental signals, incident patterns, multi-modal responses and tool-orchestrated workflows. Interest is growing in other E.ON business units, and proactive daily briefs, policy tracking and agentic workflows are next on the roadmap.
Details