What clean core means
Clean core is SAP’s architectural principle for S/4HANA: keep the SAP system core unmodified and move all custom logic to SAP BTP, to SAP Build, or to supported SAP extension points — where it does not interfere with upgrades and where it can run alongside SAP’s AI agents.
Why in-stack modifications are the problem
An in-stack modification is any change to SAP-delivered repository objects requiring the modification key, or any custom ABAP program accessing SAP internal tables directly. These modifications break every time SAP delivers a significant update. Regression testing grows proportionally. Upgrade timelines stretch from 3 months to 18. For a large ECC estate accumulated over 15 years, modification inventories routinely run to thousands of objects.
The three clean-core extension patterns
- In-app extensibility: SAP-delivered extension points — BAdIs, custom fields via Key User Extensibility, custom logic via the ABAP Cloud development model. Upgrade-stable because SAP commits to not changing the interface.
- Side-by-side on BTP: Custom applications and workflows built on BTP, communicating with S/4HANA via OData and SOAP APIs. Custom logic runs on BTP; the S/4HANA core is untouched. This is the pattern for complex apps and AI agents.
- Developer extensibility: The ABAP Cloud development model — programming in S/4HANA using only released, stable APIs and CDS views. Allowed in RISE Private Edition; not in GROW Public Edition.
Why clean core is the precondition for AI agents
SAP Business AI agents run on BTP and consume SAP data via SAP’s published APIs. They do not read SAP internal tables directly. If your S/4HANA core is polluted with in-stack modifications that alter how financial data is stored or accessed, agents cannot reliably read that data. Every SAP update also requires expensive regression testing, slowing adoption of new Joule capabilities.
Clean core is not an abstract principle. It determines whether your S/4HANA upgrade next year takes 3 months or 18, and whether your SAP AI agents work on your actual data or produce unreliable output.