An agentic system with BMW: the code calculates, the model explains

The most important rule in our design fits into one sentence: the code calculates every number, the language model calculates none.

Since September I’ve been in the seminar Trends and Concepts, this year’s topic: Building Agentic Products. Our industry partner is BMW. We’re six people from HPI, and the final presentation is at the end of January. The project is under NDA. So this post only covers which problem we’re tackling, how, and with what. No names of sites or machines, no numbers, no process flows.

The problem. In highly automated production, machines report errors and states all the time. A few of those announce an unplanned stop, and exactly those early signs are easy to miss. On top of that, delays without an actual stop often don’t create a message of their own. This is about messages, not sensor data. Continuous measurements are out of scope on purpose.

The approach has four parts. Detect: deterministic statistics on patterns in the messages. Prioritise: how hard would what’s building up hit production? Notify: tell the right role, clearly and in time. And learn: feedback flows back in. On top of that sits, by design, an agent with a language model that works only through tools. It may look things up, investigate, explain what it finds and route it to the right people. It can’t make up numbers because it doesn’t calculate any. And anything that would be written back to the ERP system is a draft that a human approves. The data stays local.

The demo on this page is a concept illustration with invented, generic data; none of it comes from the partner. It shows the flow in five steps: messages come in, the code detects a pattern, the agent investigates it through its tools, writes a draft, and a human approves it or doesn’t.