From SAP as System of Record to SAP as System of Action: The Agentic AI Opportunity for Manufacturers

SAP is, in most manufacturing businesses, the most complete and best-governed data asset the company owns. It is also, in most of those businesses, almost entirely passive. It records what happened, produces reports about what happened, and waits for a human to decide what to do about it.

The instinct when adding AI is to build somewhere else, a lake, a platform, a separate analytics estate, and leave SAP as the source. That instinct costs you the three things that make SAP valuable: the master data, the authorisation model, and the audit trail.

What a parallel platform gives up

• Master data integrity. Extract to a lake and you now maintain a second version of your material, vendor and customer masters, which will diverge. It always diverges.
• Authorisation. SAP already knows who may approve what, up to which value, in which company code. Rebuilding that elsewhere is both expensive and a genuine audit exposure.
• Audit trail. Change documents in SAP are a regulatory asset. A decision taken outside and pushed back in as a bulk update has no equivalent lineage.
• Actionability. This is the big one. A recommendation produced outside the transactional system still needs a human to re-key it inside, and that step is where most recommendations quietly die.

The value isn't in analysing SAP data somewhere else. It's in closing the loop where the transaction actually lives.

Four use cases that make the case

1. 1Finance, invoice and payment exceptions. Three-way match failures currently route to a human who investigates across the PO, the goods receipt, the invoice document and often an email thread. An agent assembles all four, identifies the specific discrepancy, proposes the resolution, and drafts the correction, inside SAP, subject to the existing approval hierarchy.
2. 2Procurement, supply risk and commitment tracking. Continuously reconcile open POs against actual delivery performance, supplier correspondence and lead time drift. Flag commitments at risk before they breach, with a recommended action and a pre-drafted expedite or re-source.
3. 3Planning, demand and supply exceptions. Rather than a planner reviewing every exception message, the agent triages them by materiality, assembles the context behind each one, and proposes plan changes for approval. Planner time moves from triage to judgement.
4. 4Maintenance, from notification to work order. Condition signals and maintenance history become a proposed intervention with parts availability, a costed production window and a drafted work order, ready for release.

Clean Core is the constraint that makes this work

The obvious objection is that this means heavy modification, and heavy modification is exactly what a Clean Core strategy exists to prevent. It doesn't have to.

The agentic layer should sit on released APIs and extension points, side-by-side extensibility rather than core modification. It reads through published interfaces, reasons outside the core, and writes back through standard APIs subject to standard authorisation. The core stays upgradeable. The intelligence is versioned separately from the ERP.

Sequencing

Start with finance exceptions. The rules are explicit, the data is entirely inside SAP, the volume is high, and the benefit is measurable in days-payable and processing cost within a quarter. It also builds the pattern, read via API, reason outside, approve via standard workflow, write back through standard interfaces, that every subsequent use case reuses.

Then extend to procurement, planning and maintenance in whatever order your pain dictates. The architecture doesn't change; only the domain logic does.

You've spent years and a great deal of money making SAP the authoritative record of your business. The opportunity now is to make it the place decisions get taken, not just the place they get recorded afterwards.