AI

AI Core inside your SAP system

Clean architecture view: runtime, governance, masking, skills, memory, and model connectivity.

Mask before LLM 1234 tabs · ←→ steps
YOUR SAP SYSTEM / ABAP STACK Runtime runs here. Only the model HTTP call leaves this boundary. AI Core governed runtime orchestrator -> governance -> skills -> audit USERS + AI CHANNELS HUMAN CHANNEL OPTIONAL AI CHANNEL NETWORK / MODELS provider + key control Human Users SAP GUI / Web Mobile 1 GUI Dock classic transactions right-side assistant Web Chat SICF surface External AI MCP client only if allowed 2 AI Core Hub control plane - config only Zones Apps Personas Skills / Keys / Policies / Masking / KB configuration only Zones scope answers + token spend GUARD MDM OPS CENT one engine, isolated team knowledge team ownership + isolation Screen Snapshot tcode, fields, message F13 session capture Async Runtime no GUI freeze aRFC + push fallback MCP Gateway SICF + SAP auth not standalone / zone+app ICF Auth SAP user Orchestrator Sense -> Ground -> Act -> Learn ROUTE GROUND TEMPLATE BUDGET PLAN/ACT LEARN bounded model/tool loop reasoning + tool calls traced Skills only SAP access GUI SQL/BAPI OPS ACT Governance Gate inside dispatch_skill before any execution Policy Approval W/X Trace Audit mandatory SAP-access gate Memory KB_SEARCH / RECALL / PTMPL / QTMPL / CASE docs + cases + templates; smarter without retraining Mask before LLM [[USER_1]] External LLM OpenAI-compatible masked HTTPS only 3 On-Prem LLM inside network masking can be off Key Resolver user -> team -> global SM59 / provider config HUMAN PATH OPTIONAL AI PATH SAP USER CONTEXT HTTPS UNMASK only here data leaves SAP inside SAP: real values wire: placeholders AI Core MCP surface not a bypass governed skills MCP calls enter governance first green = real SAP-side runtime purple = protected boundary grey dashed = Hub configuration

AI Help cockpit: user view + behind-the-scenes runtime

The user only sees SAP GUI and a chat dock. The right side shows what AI Core is doing in the ABAP stack while the conversation happens.

runtime surface
What the user sees
A user at SAP with the AI Help chat docked beside the screen
SAP GUI · VA03AI Help
AI Help
AI Core Chat
Why is this order blocked?
I can see VA03 and the current sales order. Checking status and document flow…
What AI Core does behind the scenes
SAP GUI DIALOG SESSION PARALLEL aRFC WORKER ISOLATED SKILL + DELIVERY AI Help / send sapevent:send run_turn record msg + snapshot GUI released typing dots remain Z_AICORE_RUN_TURN new work process Orchestrator loop factory -> run_chat Model decides masked LLM call which skills to call dispatch_skill policy + approval gate Skill session Z_AICORE_RUN_SKILL DESTINATION NONE Tool result back into model loop Final answer ARESULT + AMC send APC Push WebSocket paints Fallback next PAI catch-up LOOP isolated
Left: dialog session releasesMiddle: AI loop in workerRight: push or fallback delivery

Drill-down: how AI Core Hub configures the platform

The Hub is the control plane. It does not execute user turns or own SM36 schedules. It writes governed configuration, versioned templates, and data-source choices that runtime resolves when a dock, web chat, MCP call, or background agent starts.

control plane
What the admin sees
An admin configuring a governed assistant in AI Core Hub
AI Core HubGUARD / Sales Support
ZonesAppsPersonasSkillsPoliciesKnowledgeDatasetsRoutes

Configure a governed assistant

GUARD · Sales
Support chat
Sales expert
GENAI · GPT-4.1-mini
Mask before LLM
Approve writes
Live data read
Allowed skillAccess
READ_TABLEread
JOB_LOGread
USER_UNLOCKapproval
How Hub configuration becomes runtime behavior
HUB CONTROL PLANE CONFIGURATION TABLES RUNTIME RESOLUTION Zone owner + scope App chat / jobs / MCP Persona role + model style Skills allowed set Policy mask + W/X Knowledge KB/PTMPL QTMPL/CASE Tables ZAICORE_* active rows 1 active ver. logical IDs Dataset MV or re-run no SM36 control Factory resolve config Runtime execute turns
Hub: writes configFactory: resolves active setupRuntime: executes and learns

AI-to-SAP architecture: where the agent loop happens

Two ingress paths, one governed tool gate. A native SAP trigger drives the in-SAP agent loop: the masked model decides, dispatch_skill governs, isolated skills run, tool results return to the model, then the answer is delivered. An outside AI agent arrives over MCP already as its own agent — it calls a specific tool, so it goes straight through the same governed gate (auth → policy → approval → isolated skill) and the result is returned to it. It never runs our model.

governed ingress
Outside AI and native SAP triggers meet in the same ABAP model/tool loop
ENTRY PATHS AI CORE ABAP RUNTIME: MODEL / TOOL LOOP BOUNDARIES ingress boundaries direct RFC / ADT — ungoverned (no policy, no audit) agent → tool · no model Outside AI Agent MCP into SAP /sap/zaicore_mcp Native SAP Trigger GUI / job / event Direct access RFC / ADT / DB no AI Core SAP ICF auth + service user AI Core Gateway tools/list Factory zone/app/persona Trace + Audit RUNSTEP / usage Model decides answer or tool dispatch_skill policy / approval audit gate Skill session RUN_SKILL DESTINATION NONE Tool result back to model Final answer push / MCP reply External LLM masked HTTPS only model call SAP Data only via skills ABAP WORKER LOOP model/tool loop, not GUI session
Native loop: model decides → skill → result → modelMCP: tool call → gate → skill → reply (no model)Shared: one governed tool gate + audit
Reference-style board. Click boxes for detail. Purple path = masked model wire.