frank-luongt name: sap-ai-core description: SAP AI Core and Generative AI Hub integration patterns. Use when connecting AI agents to SAP S/4HANA, SuccessFactors, or Ariba via SAP BTP, or when using SAP's multi-LLM orchestration layer (Claude, GPT-4, Gemini). tags: [sap, erp, joule, btp]
SAP AI Core & Generative AI Hub
Build AI agents that connect to SAP enterprise systems using SAP AI Core's Generative AI Hub (multi-LLM), SAP BTP integration, and the ABAP SDK for Google Cloud.
When to Use
- Connecting AI agents to SAP S/4HANA, SuccessFactors, Ariba, or Concur data
- Using SAP's Generative AI Hub as a multi-LLM orchestration layer (GPT-4, Claude, Gemini, Llama)
- Building agents that query SAP OData/REST APIs for ERP data
- Integrating with SAP Joule copilot capabilities
- Using the ABAP SDK to call AI services from within SAP systems
SAP AI Core -- LLM Support (Most Model-Agnostic Enterprise Platform)
| Provider | Models | Access |
|---|---|---|
| OpenAI | GPT-4, GPT-4 Turbo, GPT-4o | Via Azure OpenAI |
| Anthropic | Claude 3 Haiku, Sonnet, Opus; Claude 3.5 Sonnet | Direct + Bedrock |
| Gemini 1.0 Pro, 1.5 Pro, 1.5 Flash | Direct | |
| Meta | Llama 2, Llama 3 | Via AI Core |
| Mistral | Mistral Large, Mixtral | Via AI Core |
| Aleph Alpha | Luminous | For European/German use cases |
Patterns
1. SAP Generative AI Hub SDK
from gen_ai_hub.proxy.core.proxy_clients import get_proxy_client
from gen_ai_hub.proxy.langchain.openai import ChatOpenAI
# Initialize proxy client (authenticates via SAP AI Core)
proxy_client = get_proxy_client("gen-ai-hub")
# Use with LangChain (model-agnostic)
llm = ChatOpenAI(
proxy_model_name="gpt-4o", # or "claude-3-5-sonnet", "gemini-1.5-pro"
proxy_client=proxy_client,
temperature=0.1,
)
response = llm.invoke("Summarize the purchase order approval process in SAP S/4HANA")
2. SAP OData API Integration
import requests
from urllib.parse import quote
class SAPS4Client:
def __init__(self, base_url: str, username: str, password: str):
self.base_url = base_url.rstrip("/")
self.auth = (username, password)
self.headers = {
"Accept": "application/json",
"Content-Type": "application/json",
}
def _get(self, path: str, params: dict = None) -> dict:
response = requests.get(
f"{self.base_url}{path}",
auth=self.auth, headers=self.headers, params=params,
)
response.raise_for_status()
return response.json()
# --- Sales Orders ---
def get_sales_orders(self, top: int = 20, filter_expr: str = "") -> list[dict]:
params = {"$top": top, "$format": "json"}
if filter_expr:
params["$filter"] = filter_expr
result = self._get("/sap/opu/odata/sap/API_SALES_ORDER_SRV/A_SalesOrder", params)
return result.get("d", {}).get("results", [])
def get_sales_order(self, order_id: str) -> dict:
return self._get(f"/sap/opu/odata/sap/API_SALES_ORDER_SRV/A_SalesOrder('{order_id}')")["d"]
# --- Purchase Orders ---
def get_purchase_orders(self, top: int = 20) -> list[dict]:
result = self._get("/sap/opu/odata/sap/API_PURCHASEORDER_PROCESS_SRV/A_PurchaseOrder",
{"$top": top, "$format": "json"})
return result.get("d", {}).get("results", [])
# --- Business Partners ---
def search_business_partners(self, name: str) -> list[dict]:
filter_expr = f"substringof('{name}',BusinessPartnerFullName)"
result = self._get("/sap/opu/odata/sap/API_BUSINESS_PARTNER/A_BusinessPartner",
{"$filter": filter_expr, "$top": 10, "$format": "json"})
return result.get("d", {}).get("results", [])
# --- Material ---
def get_material(self, material_id: str) -> dict:
return self._get(f"/sap/opu/odata/sap/API_PRODUCT_SRV/A_Product('{material_id}')")["d"]
3. SAP Joule Copilot Capabilities
SAP Joule provides pre-built AI skills across SAP applications:
S/4HANA:
- Natural language queries across ERP data
- Transaction execution (create PO, approve requests)
- Intelligent Situation Handling (automated issue detection)
SuccessFactors:
- Job description generation
- Interview question suggestions
- Learning content recommendations
Ariba:
- Supplier risk scoring
- Contract clause extraction
- Guided buying recommendations
4. ABAP SDK for Google Cloud (Call AI from SAP)
" Call Vertex AI Gemini from within SAP ABAP
DATA(lo_client) = NEW /goog/cl_generative_model(
iv_model_key = 'GEMINI_PRO' ).
DATA(lv_response) = lo_client->generate_content(
iv_prompt = 'Analyze this purchase order for compliance issues...'
)->get_text( ).
WRITE: / lv_response.
5. Google Cortex Framework for SAP Data
-- Pre-built BigQuery views over SAP data (after Cortex deployment)
-- Finance: GL accounts, AP/AR, trial balance
SELECT * FROM `cortex_sap.AccountingDocuments`
WHERE CompanyCode = '1000' AND FiscalYear = '2025';
-- Supply Chain: Purchase orders, inventory, deliveries
SELECT * FROM `cortex_sap.PurchaseOrders`
WHERE PurchasingOrganization = '1000' AND OrderStatus = 'Open';
-- Order to Cash: Sales orders, deliveries, billing
SELECT * FROM `cortex_sap.SalesOrders`
WHERE SoldToParty = 'CUST001' AND NetValue > 10000;
SAP BTP Integration Architecture
External AI Agent
--> SAP BTP Destination Service (manages connections)
--> SAP Cloud Connector (for on-premise S/4HANA)
--> SAP S/4HANA OData API
--> SAP Integration Suite (for complex orchestration)
--> Multiple SAP systems (S/4, SF, Ariba)
--> SAP AI Core / Generative AI Hub
--> LLM Provider (Claude, GPT, Gemini)
Anti-Patterns
- Calling SAP APIs without CSRF token handling -- OData write operations require
x-csrf-token - Ignoring SAP authorization objects -- always map agent actions to SAP authorization roles
- Embedding SAP credentials -- use SAP BTP Destination Service for credential management
- Skipping pagination -- SAP OData APIs use
$skip/$topfor large result sets - Using generic REST when SAP SDK exists -- use official SAP SDKs for type safety