Your SAP agent needs an approval inbox, not a chat window
The interface that makes AI safe in SAP isn't conversational — it's a queue of proposed postings a human can scan in seconds.
Chat is a great way to explore a system and a terrible way to operate one. When an agent proposes changing a delivery date on 40 sales orders, nobody wants to read that as a paragraph.
What works instead
An approval inbox: a Fiori list of proposed actions, each with the payload, the reason, and the evidence the agent used. Approve, reject, or edit. Bulk-approve the boring ones.
This is also how you earn trust. In the first weeks, every proposal is reviewed. As precision holds, teams widen the auto-approve rules — deliberately, with numbers.
Implementation notes
- Store proposals in a CAP entity with draft handling; the approval is the activation
- Keep the model's reasoning as a text field — auditors ask for it
- Fire the actual RAP action from the approval, never from the agent
The chat window can stay. Just don't let it hold the pen.