Academy module

Databricks Mosaic AI on BDC — ML on SAP Data Products

As of 2026-11-08

For consultants scoping or reviewing a Databricks-side build (agent or ML model) against SAP data shared zero-copy via Business Data Cloud Connect. Covers the OpenSharing setup sequence and security model (GA October 2025), why semantic metadata syncing into Unity Catalog matters specifically for ML, Mosaic AI Agent Framework's narrow value proposition (Delta tables, registered ML models, vector search in one reasoning loop) versus a simpler direct chain, building tools via Unity Catalog functions and AI Search with MCP as a bridge to SAP-exposed tools, Unity Catalog/Unity Gateway governance as the Databricks-side equivalent of a Datasphere DAC review, model signatures tied to data-product version pinning, and the bidirectional write-back path that returns Databricks-computed results to SAP as governed assets. Closes with the Mosaic-AI-or-SAP-AI-Core decision and three exercises.

What you will learn

Create and verify an SAP BDC connection in Databricks (Data Ingestion → SAP Business Data Cloud) and record owner, connection count against the five-per-metastore limit and mounted catalogs
Check in Catalog that comments, primary keys, same-share foreign keys and `sap.PersonalData.*` tags arrived for a mounted share, and use Refresh Table when they have not
Write an ABAC policy sentence that masks `sap.PersonalData.isPotentiallyPersonal` columns while accounting for the default value true
Register a Unity Catalog SQL function as an agent tool and grant `EXECUTE` to a dedicated service principal
Share a keyed score table back to SAP BDC with the SAP BDC SDK, re-initialising `BdcConnectClient` before each call
Build and run a 40 to 60 item golden set that tests model, tool and agent answer separately and obtain the three sign-offs

Module overview

Who this is for. You have completed data products for AI agents and a client's data-science team wants to build machine learning and agents on Databricks against SAP data. It is not a Joule agent but a Databricks-side one, run by people who may never open the BDC catalog. By the end you can mount an SAP data product in Databricks, prove that its semantic metadata arrived, build a Unity Catalog function as an agent tool, score a business object with a registered model, and return the score to SAP consumers as a governed asset. The running case is a late-delivery risk score for purchase orders at a mid-size manufacturer.

Prerequisites

SAP Business Data Cloud Data Products for AI Agents or equivalent understanding of data products and governance metadata
Working familiarity with Databricks fundamentals: Unity Catalog, Delta Lake, workspace concepts
Basic understanding of zero-copy sharing / Delta Sharing (the concept “Delta Sharing Protocol”)

Outcomes

Trace a real zero-copy share and verify, not assume, that semantic metadata synced into Unity Catalog.
Decide correctly between Mosaic AI Agent Framework and a simpler direct chain for a given agent's requirements.
Configure Unity Catalog governance to mirror SAP-side governance tags rather than leaving default grants in place.
Design a write-back path that returns a Databricks-computed result to SAP as a governed, consumable asset.

Full module with the Consultant plan — €19.90/month. The full module adds: the decision framework · the end-to-end scenario walkthrough · the KPI scorecard · the anti-patterns · the code blocks · the knowledge check · the diagrams.

When is the Databricks Mosaic AI Agent Framework worth it on SAP data?

The module treats the Databricks Mosaic AI Agent Framework as a narrow fit: it earns its place when one reasoning loop has to combine Delta tables, registered ML models and vector search over SAP data products shared zero-copy through Business Data Cloud Connect. When the task does not need all three, a simpler direct chain is easier to govern. Tools are built from Unity Catalog functions and Vector Search, with MCP as a bridge to SAP-exposed tools.

Open in the app →

Edited by Cédric Mary · How this page is made