SAP

Langfuse 是专为 LLM 应用设计的开源可观测性平台，能追踪模型调用的输入输出、Token 消耗、延迟和工具调用路径，并支持 Prompt 版本 A/B 评估

在SAP、Oracle等传统ERP系统中，采购订单、收货单和发票三者必须在数量和金额上保持一致才能触发付款流程，这被称为"三单匹配"

SAP AI Core是SAP Business Technology Platform（BTP）中的AI基础设施服务，基于Kubernetes构建

BAPI (Business API) is a standard interface officially released by SAP that is stable across versions, with complete documentation, version management, and backward-compatibility commitments; a Z-function is an RFC function custom-developed by the customer under the SAP namespace Z*, lacking official documentation and upgrade guarantees.

During SAP upgrades, BAPI compatibility is SAP's responsibility, whereas Z-functions may silently fail due to changes in underlying table structures or function signatures.

Clean Core is an architectural principle that SAP has heavily promoted in recent years, whose core requirement is to keep the S/4HANA standard core unmodified by customers (no modifying standard code, no directly writing to standard tables), with all extensions implemented through released APIs, the BTP extension framework, or side-by-side extensions.

The Power Platform ecosystem has more than 1,500 official and custom connectors, covering systems such as SharePoint, Dataverse, SAP, and Salesforce.

The SAP clean core principle requires implementing custom logic through parallel extensions (deployed on BTP) or released APIs rather than directly modifying SAP standard code, to ensure the system can be upgraded smoothly and reduce technical debt.

In SAP scenarios, agent sprawl manifests as each workflow establishing its own MCP endpoints, prompt packages, and shadow integrations; architects should prefer parallel extensions and released APIs, require tool whitelisting and non-production-first MCP, ensure agent usage is consistent with the clean core principle, and track AI Units consumption as an architectural non-functional requirement.

For SAP architects, the article proposes a 12-month skills plan: in the first quarter, learn the MCP threat model and build a non-production agent with read-only tools; in the second quarter, add observability and write a permission matrix; in the third quarter, practice cost routing patterns and document AI Units forecasts; in the fourth quarter, lead a governed production use case end to end.

AI Units are the metering unit SAP uses for its AI services (such as Joule and model calls in AI Core), similar to Tokens or API call counts in cloud services; SAP bills customers based on consumption.

In the 2021+ version line, HANA Cloud native capabilities have enhanced columnar storage compression and encoding, multi-temperature data management, parallel processing, advanced SQL window functions and approximate aggregation, as well as JSON document storage and spatial/graph computation support.

The promise of Clean Core is that upgrades do not break standard functionality, and combined with the continuous improvement delivery model, this ensures that once AI extensions are built, they will not have to be torn down and rebuilt because of the next upgrade.

On the 2021 version line, the ABAP RESTful Application Programming Model (RAP) has implemented key capabilities such as automatic generation of OData V2/V4 services, decoupling of behavior definitions, a unified development model, and side-by-side extensibility.

On the 2021+ version line of SAP S/4HANA Cloud Private Edition, SAP backported ABAP AI and Joule-class assistant capabilities to private cloud versions from 2021 onward.

Modification refers to directly modifying SAP standard program code, which creates conflicts when SAP releases a Support Package or upgrade, causing standard updates to be locked out.

FPS (Feature Pack Stack) is a feature delivery unit in the SAP S/4HANA version system, sitting between major versions and Support Packages; some AI components were previously available only for the latest FPS.

In SAP AI scenarios, MCP allows Joule or the ABAP AI assistant to securely access enterprise systems, call APIs, and obtain business context, but it introduces a new access control plane that needs to be brought into the governance scope.

In predictive maintenance scenarios for production equipment on SAP systems before 2021, sensor data integration is fragile, and an integration project goes from an expected 6-8 weeks to 6-12 months.

Decision rules for enabling SAP AI capabilities: enabling in place applies to situations where you are already running on a supported 2021+ PCE version line, the Support Package is new enough to meet the required AI Notes, and the security team accepts the agent/MCP security posture; upgrading first applies to situations where you are below the documented minimum version requirement, lack prerequisite components, or the Support Package is too far behind.