Models lived offline, not on the floor

Before continuous live positioning became practical in most plants, industrial AI conversations stayed anchored to models that never left the conference room or the simulation server. Static digital twins came from CAD files, scans, and BIM. Simulation twins dropped those models into physics engines to run what-if layout changes. Data platforms tried to organise MES, ERP and sensor streams into digital threads. Predictive claims sat on top of those layers.

What was missing was the continuous, spatially accurate location of the things that actually moved product: forklifts, people, materials. Without that stream, the AI had no persistent ground truth to compare against the model. A platform that began development in 2014 had to work inside that reality for years. Live RTLS of the kind that later became central to operational twins was still maturing. BLE and UWB were not yet everyday fixtures on most factory floors.

Simulation could plan, but not watch the shift

Simulation tools answered useful questions before a change was made. They could not show the actual spaghetti of routes taken on a given shift. Without position updates arriving every few seconds, there was no continuous record of where a pallet sat, how long a driver waited, or which zone three forklifts occupied at the same moment a line stopped.

Plant teams evaluating those earlier systems usually received dashboards of aggregated KPIs and OEE values fed from the MES. The spatial context that would later let someone see a micro-stoppage in three-dimensional relation to a forklift path was still years from being common. Causality talk between logistics gates, HVAC loads and quality remained conceptual. Correlation required the where and the when of every movement, and that data simply was not present at scale.

Digital threads stopped at the abstract

Many solutions organised heterogeneous streams into digital threads. They ingested PLC data, production counts and quality events. Yet without a live map of physical positions the thread stayed abstract. An engineer could see that a line had stopped. They could not immediately see that the stop coincided with three forklifts idling in the same zone or a material buffer that had emptied two aisles over.

The long build that started in 2014 unfolded against this background. Positioning technologies had to become reliable enough and inexpensive enough for everyday floor use before an operational layer could close the gap between model and reality. Until then the AI pitch stayed one step removed from the physical movement of goods and people.

How the missing positions shaped the language

Sales language in that earlier period leaned on future closed-loop architectures and AI that would eventually detect root causes automatically. The concrete mechanism—continuous three-dimensional positions of every relevant object—was still emerging. Accuracy ranges that later became familiar (BLE typically 1-3 metres, UWB under 30 cm) were not yet baseline assumptions in most plants. The ability to replay a full year of positions was not a standard feature buyers expected.

As a result many AI-in-manufacturing presentations remained high-level: optimise flows, cut idle time, improve reaction. The data that would later turn those phrases into measurable route analytics, spaghetti diagrams and shift-level movement reports had to be built into the infrastructure first. The industry was selling the destination while the road surface was still being laid.