Spatial data has to exist before the model does anything useful
Most conversations about AI in manufacturing jump straight to the model. What gets left out is the data the model actually needs to see: not just rows in a table, but positions in space and time. A forklift’s location three seconds ago, the dwell of a pallet at a gate, the path a material handler took during the last shift. Without that spatial layer, correlation between logistics, production state, and environmental conditions stays guesswork.
Building the layer is the unglamorous part. It means RTLS coverage dense enough to produce usable coordinates, feeds that arrive with consistent timestamps and clear ownership, and enough history stored so past shifts can be replayed. None of this is the AI model. It is the prerequisite spend that has to sit on the plant floor first.
Tag and anchor density is the first hard cost
Accuracy is technology-dependent. BLE typically delivers 1-3 meter accuracy; UWB can reach sub-meter. In a 30 000 sqm plant that difference decides whether a zone boundary is trustworthy or just noise. Enough fixed anchors have to be placed so tags report without dead zones, and every asset or person that matters needs a tag. Battery life runs 1-3 years on many BLE tags; UWB units often need more frequent charging.
Installation is not a software exercise. Fixed sensors or readers go in, assets get tagged, the system is calibrated. Typical full deployment runs one to three months. The hardware, mounting labour, and network load are real plant costs that appear long before any model training budget is opened. Cases that later showed fleet reductions and lower logistics OPEX still began with this physical layer.
Governed feeds turn raw positions into a usable dataset
Positions alone are not enough. The platform has to ingest them alongside data from MES, ERP, WMS or SCADA so each location point can be joined to order status, equipment state or material identity. Twinzo functions as the umbrella layer for that ingestion; the feeds themselves still need governance—consistent time stamps, role-controlled access, clear rules on what is tracked and why.
Without that discipline the spatial set fragments. One system updates every three seconds, another only on events, a third carries no location context at all. The joins fail and any later model inherits the mess. The cost here is less about licence fees and more about the engineering time spent defining ownership, cleaning the streams, and keeping them aligned after the first pilot ends.
Replay infrastructure is the quiet enabler
Once the live layer is running, history becomes the next requirement. The ability to replay positions a year back, generate spaghetti diagrams, heat maps and zone-activity reports turns a live map into a training and audit resource. Engineers can return to a shift that went wrong, examine the actual routes taken, and test whether a proposed change would have altered the outcome.
Storage, retention policy and the compute to serve those replays are again separate from model cost. They sit in the same category as the anchors and the governed feeds: necessary plumbing. Without them the spatial dataset stays ephemeral. Models can only learn from what is still retained and still linked to the original locations.
The spend stays unglamorous for a reason
None of this appears on the glossy AI slide. It shows up as tag purchases, anchor mounting, network upgrades, data-governance workshops and a year of position history sitting on disk. The plant still has to decide whether the density is high enough, the feeds clean enough, and the retention long enough before any model is asked to find patterns. That decision is the real starting point for AI that claims to understand manufacturing space and time.