ERP, PLM, drawing libraries — the integration gap that sinks AI projects

There was a number at the VDMA AI summit in Frankfurt today that should stop every project sponsor in the room: as one speaker put it, 70% of AI projects fail due to poor data quality and integration. Not bad models. Not wrong use cases. The pipes.

I work at the point where an AI idea meets the systems that hold the actual data — ERP, PDM, PLM, drawing libraries that were built before the word "API" was common vocabulary in Maschinen- und Anlagenbau. The job is eliminate, simplify, automate. In that order. The automation only works if the simplification happened. The simplification only works if the unnecessary step was eliminated first. Bolting an LLM onto a broken process does not fix the process. As the same speaker noted, it produces 0% productivity gain and a team that now distrusts AI.

What the integration layer actually requires: a clean semantic map of what each source system means by the same word. A "part" in ERP is not the same "part" in PLM. A "revision" in the drawing library is not the revision the machine tool tracks. Until someone has drawn that map — and validated it with the people who use the systems daily — the model is translating between languages it hasn't been taught.

This is not an AI problem. It is a data architecture problem that predates AI. AI just makes the gap visible faster and more expensively than it used to be.

The good news: the gap is fixable. We fix it before the model touches the data. That is the right order of operations.