At Sapphire 2026 in May, SAP announced something that will reshape how every SAP customer thinks about AI integration: the SAP Business AI Platform. It consolidates SAP BTP, Business Data Cloud, and SAP Business AI into a single, unified stack — and it changes the answer to the question every enterprise is currently asking: “Where do I even start with AI on SAP?”
This article breaks down what changed, what the new architecture actually looks like, and what it means practically for teams planning SAP AI integration projects right now.
What SAP announced
Before Sapphire, the SAP AI landscape was fragmented: BTP for development, Business Data Cloud for data, SAP AI Core for model hosting, and Joule sitting on top. Different licensing, different teams, different documentation.
SAP Business AI Platform merges all of this into one product with one architecture. The goal — stated explicitly in the keynote — is to solve enterprise AI fragmentation. The underlying components have not disappeared. They have been reorganised into a coherent stack with a single entry point for AI development and deployment.
The six components you need to understand
SAP AI Core
What it is: The model hosting layer. Runs LLMs on BTP infrastructure — both SAP-provided and third-party models.
Why it matters: You do not send your SAP data to a public API. Models run inside SAP's governed environment.
Generative AI Hub
What it is: A governed gateway to external models — Azure OpenAI, Google Gemini, Mistral, and others.
Why it matters: Lets you use the best model for each task while keeping access control and audit logs in one place.
Joule Studio 2.0
What it is: A low-code agent builder inside BTP. Build, test, and deploy AI agents against your SAP processes.
Why it matters: Non-developers can configure agents. Developers can extend them with custom tools via MCP.
SAP Agent Gateway
What it is: The runtime that routes requests between AI agents and SAP backend systems. Uses A2A protocol.
Why it matters: Replaces custom RFC/OData wiring. One gateway handles agent-to-SAP and agent-to-agent calls.
HANA Cloud Vector Engine
What it is: Embedded vector search inside HANA Cloud. Stores and retrieves semantic embeddings alongside structured data.
Why it matters: Enables RAG (retrieval-augmented generation) directly against your existing HANA database — no separate vector DB.
SAP Domain Models
What it is: Pre-trained models built on SAP business logic and code. GA expected Q3 2026.
Why it matters: An AI that already understands what a cost centre, goods receipt, or dunning run actually means.
MCP and A2A: the two protocols that hold it together
Two open standards underpin the SAP Business AI Platform's agent architecture:
MCP (Model Context Protocol) — used to give AI agents structured, semantically enriched access to SAP business data and capabilities. SAP uses MCP internally for Joule agent connectivity and has shipped an MCP server for SAP Commerce Cloud (Storefront MCP Server, Q2 2026 GA). Any external AI agent that supports MCP can connect to your SAP system through an MCP connector without bespoke integration code.
A2A (Agent2Agent)— SAP's preferred protocol for communication between AI agents. Where MCP handles agent-to-data connections, A2A handles agent-to-agent orchestration: one AI agent handing off to another, multi-step workflows, human-in-the-loop approvals. The Agent Gateway (SAP's runtime for A2A) is currently unidirectional; bidirectional GA is expected Q4 2026.
Understanding both protocols is now a baseline requirement for any SAP AI integration engagement. They are not competing standards — MCP and A2A serve different layers of the same architecture.
What this means for your integration project
If you were planning an SAP AI integration project before Sapphire, your technical approach may need revisiting. Specifically:
- →Do not build custom RFC/OData wrappers for AI. The Agent Gateway is designed to replace this pattern. Build to the gateway, not directly to the backend.
- →Design for MCP from the start. Even if you are using Joule Studio today, MCP compatibility ensures your integration works with any future AI model or agent framework.
- →HANA Cloud Vector Engine removes the need for a separate vector database. If you are on HANA Cloud, you can run RAG directly against your existing data — no additional infrastructure.
- →SAP Domain Models (Q3 2026 GA) will materially change what agents can do out of the box. Projects starting now should account for these in the architecture — they will eliminate significant prompt engineering effort for standard SAP processes.
The window for first-mover advantage
The SAP Business AI Platform is real, but large portions of it are not yet GA. Agent Gateway bidirectional support lands Q4 2026. SAP Domain Models land Q3 2026. This means the organisations that start architecture and pilot work now will have production-ready implementations before their competitors have finished procurement approval.
The question is not whether to adopt this stack. It is whether to be the team that shapes how your organisation uses it, or the team that inherits what someone else decided.
Planning an SAP AI integration project?
We help enterprise teams architect and build SAP AI integrations using MCP, the SAP Business AI Platform, and Joule Studio. Tell us where you are and we will tell you what makes sense.
Get in touchSources: SAP Architecture Center (updated May 2026), SAP Sapphire 2026 keynote announcements, SAP News.