Technical RFC: Building an "Auditable Autopilot" for Material Shortages – Where is the real Excel He

[skalej]

Hi everyone,

I’m an AI developer in Germany architecting a "Human-in-the-loop" agent to automate exception handling for discrete manufacturers (Mittelstand).

The goal is to reduce the 4+ hours of manual coordination that happens when a supplier is late or a carrier delay occurs, leading to potential line stops. We are focusing on a safe, auditable "exception autopilot" model rather than a generic "digital brain."

I have a few "reality check" questions for the consultants and experts here. I need brutal honesty so I don't build a product that fails in a real German factory environment:

1. The 'Shadow' System: When a delay hits, does the rescheduling happen in SAP (e.g., MD04/CO06) immediately, or do planners usually run a 'Shadow Excel' for 1-2 days first to figure out the impact?
2. The Data Gap: What’s the one piece of information that is never in SAP (e.g., real-time carrier GPS, unstructured email updates) that forces a human to manually 'bridge' the gap?
3. The 'Green Button' Barrier: If an AI drafted the SAP update and the supplier email, what is the #1 reason a German Plant Manager would not press 'Approve'? (e.g., legal liability, lack of audit trail, trust issues).
4. On-Premise vs. Cloud: For a mid-sized manufacturer, would the IT department demand this system runs 100% on-premise for data sovereignty, or is a highly controlled cloud connection (like SAP BTP) acceptable now in 2026?

Looking forward to the insights from the community so we can build something that actually solves the real pain points.

Thanks in advance for your expertise!