Not all actions have the same consequences
Adjusting lighting and changing a critical process are different categories. Industrial architecture distinguishes them before discussing automation. A recommendation can be useful without automatic execution, and limited authorisation must not open access to other assets.
What the system contributes
The intelligent layer can recommend schedules, detect abnormal consumption, anticipate failure, prioritize maintenance and propose operational changes. Physical actions are enabled by policy and criticality. Lighting, auxiliary motors and process shutdowns do not share the same authority level.
Make decision authority visible
The organisation defines presented information, proposals needing review and actions that might be enabled after sufficient testing. Reviewers need the evidence and expected effect. A model should not expand its own powers because it finds a hypothesis convincing.
Check the effect, not only the command
Authorised actions within agreed scope need observable results and records. A sent request is not completed operation. The platform integrates with plant procedures to support explainable decisions, not create a second, hidden authority over the installation.