DATA LAYER · WHY PILOTS FAIL
88% of pilots fail·58% of budget on integration·40% cancelled by 2027
The model you piloted was fine. It was handed inconsistent names, mixed units and unvalidated readings, and asked to be smart about them. That is a data layer problem, and no amount of model upgrade fixes it.
The uncomfortable version: swapping models does not move the number. Fixing the input does.
This is the thesis the rest of FoundryNet is built on. A Fanuc CNC speaks FOCAS. A Siemens PLC speaks PROFINET. A KUKA robot speaks KRL. A 20-year-old press speaks Modbus over a serial cable. Your agent sees five dialects with no shared vocabulary and produces confident nonsense or nothing at all.
The industry response has been to buy a better model. The numbers do not support that response. 88% of AI pilots never reach production, 40% of agentic AI projects will be cancelled by 2027, and only 15% of enterprises have AI-ready industrial data. Those three statistics describe one failure, counted three ways.
Two vendors, two names, one physical quantity. The agent treats them as unrelated series and reports no correlation where there is a direct causal link.
A raw value of 4.2 is millimetres per second or inches per hour depending on a vendor decision nobody wrote down. A PWM duty cycle on a 0-127 scale read as a percentage is 25 points wrong, silently.
A sensor stuck at a plausible value looks like a healthy machine. Nothing in the model layer catches this. It has to be caught at read time, before inference.
Every new machine, vendor or protocol restarts the mapping work. Cost scales with fleet size, benefit does not, and the project is cancelled before it clears the second vendor.
Forge sits where the mapping work would otherwise live. Five layers resolve a tag; six validators police the result; the corpus remembers what it learned.
16,908 confirmed mappings across 18 OEM families resolve the common case with no model call at all.
A mapping that implies an impossible value is rejected regardless of how confident the match looked. Physics is a cheaper referee than a production incident.
LLM resolution → cache → confirmed at 5 uses → auto-packed at 10 → Layer 1 forever. The cost of a novel tag is paid once per tag, not once per call.
Below the confidence floor, Forge returns the field unmapped. A missing field is recoverable. A confidently wrong field is what your incident review will be about.
Longer-form writing on industrial AI infrastructure by Foundry Labs.