01
What one question touches
02
What "governed" means here
One definition per term.What "available capacity" means is agreed with operations and written once, not in each prompt.
One owner per source,who decides which version is right when two systems disagree.
Lineage.Every number the agent uses traces back to a system and a timestamp.
Access.The layer applies the same permissions as the people asking.
Freshness targetsper source, matched to how often the decision is made.
03
The build order
Pick one decision,such as the delivery date promise. Not "all plant data".
Map the inputsit needs, and name an owner for each.
Reconcile the codesfor plants, lines and products into one model, in the warehouse or lakehouse you already run.
Compute in code.Capacity, material and schedule checks run as planning and optimization logic, not inside the language model.
Expose a few governed views or APIs.
Only then add the agent.It asks, calls the views, and explains the answer.
04
Shop-floor data comes in carefully
Historian and SCADA data is read-only to this layer. Following the Purdue model, data flows up from the plant network through a demilitarized zone, and nothing flows back to the line. Where data has to stay on site, the processing runs at the edge.
05
Start small, let it grow
The layer is scoped to the first decision. Each next use case reuses what is already reconciled, so it costs less, which is how one system makes the next one cheaper. For a first decision, that is a matter of weeks, not a data programme.
The language model is the interface. Planning, forecasting and optimization compute the answer.
Related use case
Decision intelligenceOne answer across systems that disagree.