Every established manufacturer runs two data estates. The first is structured and well tended: ERP records, machine telemetry, quality measurements, the numbers that feed the dashboards. The second is far larger, barely governed, and until recently unusable - service manuals, engineering notes, process documentation, decades of correspondence about what went wrong and how it was fixed.
Bystronic, the Niederönz-based sheet-metal-processing manufacturer, has deployed agentic AI across that second estate. The system is targeting a 50 per cent reduction in customer-service ticket volume and a 40 per cent shorter onboarding cycle for new employees, with more than 200 daily active users and growing.
Why unstructured data was left alone
The neglect was rational. Conventional analytics needs structure, and this material has almost none. A field engineer’s note from 2014 explaining why a particular alloy behaved unexpectedly is not a row in a table. It cannot be joined, aggregated, or charted. For most of the history of enterprise software, the only way to extract its value was for a person who remembered it to still be employed.
That is what changes. A system that can reason over documents, answer a question posed in ordinary language, and surface the relevant precedent turns an archive into something closer to institutional memory that does not resign.
The two numbers point the same way
Fewer support tickets and faster onboarding sound like separate wins. They are the same one seen from two directions.
Both are bottlenecks caused by knowledge existing but not being reachable. A support ticket is often a question whose answer sits in a manual nobody can find quickly. A long onboarding is a new engineer slowly rebuilding, from conversation and trial, a body of knowledge already written down somewhere. Making the archive queryable addresses both without adding a single new document.
For a Swiss manufacturer facing a retirement wave among its most experienced staff, this is a workforce problem as much as a software one.
The part that decides the architecture
Here the sovereignty question stops being theoretical.
The material that makes such a system valuable is the material a precision-manufacturing business least wants to disclose. In optics, photonics, and micro-fabrication, the difference between a firm and its competitors is rarely the equipment on the floor - that can be bought. It is the accumulated knowledge of how to run it: the process recipes, the tolerance decisions, the failure modes learned expensively over years. That knowledge is exactly what sits in the unstructured archive.
Feeding it to a general-purpose external service means handing over the competitive position in the same movement that makes it searchable. The dependency and the disclosure are one act. And unlike a data breach, it is not an incident anyone would notice.
There is a related consideration for firms whose products or process knowledge fall under dual-use export controls: transferring technical data across a border can be a controlled event in itself, independent of whether anything physical is shipped. An architecture where the archive never crosses the boundary avoids the question rather than having to answer it per query.
The practical shape
None of this argues against agentic AI on the shop floor. Bystronic’s results argue firmly for it. It argues about where the system runs.
Keeping the model and the archive inside one controlled boundary - dedicated hardware, a named jurisdiction, an auditable perimeter - preserves the whole benefit while removing the disclosure. The system still reads every document, still answers every question, still shortens onboarding. It simply does so somewhere the company can point to on a map and describe to a customer who asks where their drawings are held.
For the Swiss deep-tech corridor, that is not a compliance detail bolted on afterwards. It is the condition that lets the archive be opened at all.