The Approval Hub

AI agents are increasingly given write access to enterprise systems like SAP - or none at all while businesses watch how other businesses implement. Most implementations are either fully autonomous (dangerous — agents making material changes with no oversight) or fully manual (defeats the point). The Approval Hub is a governance layer that intercepts every agent action targeting SAP. Procurement emails arrive, agents parse them, enrich them from SAP data, evaluate against policy rules, and propose actions. The hub gates those actions: clean cases auto-execute; risky or ambiguous ones route to a human reviewer with the full context pre-assembled. Every write is logged with a content hash and correlation ID. HOW IT WORKS 1. Email arrives at the procurement inbox (B2B orders, supplier confirmations, customer enquiries). 2. Email Agent (Gemini) parses intent + entities — strict JSON, schema-validated. 3. SAP Agent runs read-only enrichment (vendor, material master, pricing, open POs, stock). 4. Policy Engine evaluates deterministic rules — sign-off thresholds, KYC, pricing alignment, fraud — and returns auto-approve, human review, or block. 5. For sales orders, the agent auto-creates the SO in SAP immediately. The customer ordered via our portal using our terms; we honor that. Hub becomes the fulfilment workspace. 6. Fulfilment plan computes per-line: in-stock, waiting on inbound PO, or needs new PO. Supplier grouping batches required POs. 7. Human reviewer sees the case pre-assembled. Approve → SAP writes execute → confirmation email auto-drafts from the SAP document. One click to send. 8. Audit trail captures everything — agent calls, policy decisions, SAP writes, human actions. PRINCIPLES Agents talk only to the hub. LLM extracts and plans; Python decides and executes. Read-only by default. Fail-to-human is a feature. Every write produces a SAP document ID.