Computer vision on the shopfloor: copy the smart city privacy playbook
Plants across Europe are testing AI vision for safety, quality, and logistics. The models are good enough. The bottleneck is trust, governance, and who holds the data.
Smart city pilots offer a clear lesson. The successful ones process video on site, do not retain raw footage, and prevent vendors from touching the data. That design cut privacy pushback and let them focus on outcomes like safer intersections and cleaner streets.
Factories can do the same. You protect workers, cut legal friction, and keep speed high.
What the cities got right
Recent projects in the US point to a pattern worth copying.
One city kept all equipment and data management inside its own data center. The technology supplier delivered hardware and integration, but had no access to the data. Another city used video only to extract counts and classifications for traffic engineers. The system did not retain video, only the derived data.
Two ideas stand out. Process at the edge or on site. Keep vendors out of your data path. Purpose limit what you collect and how long you keep it. That combination turns a surveillance concern into an operational tool.
Why this matters even more in European manufacturing
You operate under GDPR, works councils, and high workforce expectations. Trust is part of your license to operate. If people feel watched, your improvement program will stall.
There is also jurisdictional risk. A provider that is subject to non‑EU law can be compelled to hand over data, even if the servers sit in the EU. Physical location is not the same as legal sovereignty. Real sovereignty means the company controlling the data is under EU law, with EU ownership and governance.
Designing for privacy and sovereignty is not a nice to have. It is the fastest path from pilot to scaled use, because you remove the debates that slow decisions for months.
A privacy‑by‑design checklist for vision projects
Use this to scope your next pilot. If an item fails, fix it before you start.
- Define one narrow purpose. Example: detect pedestrian and forklift near‑misses. No open‑ended monitoring.
- Data minimization. Default to processing on device or on premises. Export only the signals you need, not video.
- No raw video retention unless there is a defined incident workflow. If you must retain, set short, automated deletion windows.
- Role‑based access. Named users. Least privilege. Every access is logged and reviewed.
- Vendor boundary. Supplier may design and integrate. They do not access, collect, or analyze your data. Put it in the contract.
- Sovereignty by architecture and contract. EU‑owned provider. EU jurisdiction only. No parent company under extra‑EU laws. Servers in the EU are not enough.
- DPIA ready. Document purpose, lawful basis, risks, mitigations, and retention. Involve the DPO early.
- Worker transparency. Clear signage where cameras operate. Briefings with Q&A. Share what is collected, what is not, and who can see it.
- De‑identify by default. Blur faces and badges when the use case does not require identification. Keep outputs at the object or event level.
- Incident and audit drills. Prove you can answer who accessed what, when, and why. Test deletion works.
Start small and time‑box it
Pick one use case with obvious value.
Examples:
- Near‑miss detection in mixed traffic zones.
- Quality anomaly detection at one critical workstation.
- Illegal dumping around your perimeter or laydown areas.
Run a six‑week sprint.
Week 1. Install edge devices. Write the purpose, retention, and access rules.
Weeks 2–4. Tune models on site. Produce only event counts, heatmaps, and short incident clips with automatic deletion.
Week 5. Review with HSE, works council, and the DPO. Walk through access logs and deletion tests.
Week 6. Decide. Scale, adjust, or stop. If you scale, the rules follow the system.
This cadence keeps execution flow high. You reduce decision latency because the governance is built in, not bolted on.
The strategic choice
As software costs drop, advantage shifts back to execution and flow. Vision is just another sensor. The edge comes first. The plant owns the data. Vendors ship capability, not custody.
If you hold that line, you move faster and keep control where it belongs. Inside your walls. Under EU law.
Want a one‑page privacy and sovereignty spec for your next vision pilot? We share our template with operators who plan to run a six‑week sprint. Get in touch.