How to Connect Microsoft Fabric to SAP ERP Data: The Practical Guide

The analytics team is told SAP data is available. Three weeks later they learn what that means: a BASIS consultant will run an extract when asked, drop a CSV on a shared drive, and go back to his queue. That is not an integration. It is a favour, and it does not survive a monthly reporting cycle, let alone a demand forecast.

It stays that way because every proper route into SAP has a licensing, performance or contractual constraint bolted to it — and the person who owns that constraint (the SAP account team, the BASIS lead, the CFO who signed the SAP contract) is not in the room when the architecture gets drawn.

As of August 2026, six practical routes — each with a constraint to check first.

Two things to note. The ADF SAP CDC connector was historically the delta route via the ODP framework, but Microsoft warns SAP Note 3255746 blocks ODP RFC API calls from third-party clients including it. And no SAP connector in Fabric supports SAP as a destination — every one is source-only.

Microsoft's connector overview lists SAP BW Application Server and SAP BW Message Server for Dataflow Gen2; SAP BW Open Hub and SAP Table (each in Application Server and Message Server variants) for pipelines and Copy job; and SAP HANA database across all three experiences. All are source-only.

The split by experience matters: choose Dataflow Gen2 because that is where Power Query lives and you get SAP BW and SAP HANA — but no SAP Table. Choose pipelines or Copy job for SAP Table and Open Hub. Getting this wrong is a common early misstep that sends teams down a route their chosen experience cannot support.

Full extracts stop scaling when the extraction window exceeds the business's tolerance for stale data or the SAP system's tolerance for load, whichever comes first — in practice somewhere between 20 and 80 million rows on a single large document table. The SAP Table connector has no delta of its own; it has partitioning (PartitionOnInt, PartitionOnCalendarYear/Month/Date, PartitionOnTime) that splits one enormous read into parallel smaller ones, but you are still re-reading.

Copy job with the SAP ABAP Add-On is the more honest answer at volume — watermark-based incremental copy against long timestamps, short timestamps and DATS date columns, with support for transparent, pool and cluster tables, views and CDS views. True change data capture (deletes detected, no reliance on a timestamp column) is what the ODP framework gave you and what SAP Note 3255746 has restricted, so Microsoft now points customers to Datasphere mirroring or third-party tools for that.

Every built-in SAP connector routes through the on-premises data gateway. It needs version 3000.214.2 or later for Fabric pipelines, outbound firewall access to the Fabric front-end endpoints, and the SAP drivers installed on the gateway machine itself.

The implications bite in month three, not week one: the gateway becomes production infrastructure the data team owns and the infrastructure team patches — a Windows machine carrying the SAP .NET Connector or HANA ODBC drivers, through which every SAP row bound for OneLake passes. Cluster it (up to 10 members with automatic failover). For the ABAP Add-On route the firewall conversation widens to outbound HTTPS to OneLake from both the gateway and the SAP application server, plus inbound RFC from the gateway to SAP.

Four things must be agreed with the SAP BASIS and functional teams before a single extraction job is scheduled, because extraction runs on the production SAP system's work processes and competes with transactional users. Bring specifics, not a request for "access to SAP".

• The window — name the hours extraction may run, and the hard exclusions around month-end close
• The parallelism ceiling — partitioned reads open multiple RFC connections; agree the maximum
• The authorisations — for the SAP Table route, RFC destination authorisation and execute rights on S_SDSAUTH
• The transports — the ABAP Add-On route requires importing Microsoft's transports, a change to production ABAP moving through your dev-quality-production path and change advisory board

Confirm your extraction entitlement in writing with your SAP account team before you build — this is where confident consulting advice does the most damage. What is documented publicly: Microsoft Learn carries the ODP-RFC blocking warning (SAP Note 3255746) on its SAP CDC pages, documents Premium Outbound Integration as a prerequisite for the mirroring path, and documents which connectors exist. What is not publicly verifiable: the full text and revision history of the SAP Note, whether your contract permits a given extraction interface, and whether exposing lakehouse copies of SAP data to non-SAP-licensed users creates an indirect or digital access liability.

So put three questions to your SAP account team in writing and keep the reply: which extraction interfaces does our agreement permit; does routing SAP data to Power BI consumers change our digital-access position; and does installing the Microsoft ABAP Add-On affect support of our SAP system.

SAP stores business objects across multiple client-dependent tables with compound keys, conversion exits and language-dependent text tables. A material master is not MARA — it is MARA plus MAKT for descriptions, MARC for plant data and MARD for storage-location stock. This is where SAP-to-Fabric projects overrun: the extraction is built in a fortnight, the modelling is the project.

Four traps. The client field — nearly every client-dependent table carries MANDT as its first key, and you filter to your production client once, not in every report. Compound keys — a material document line is keyed on MBLNR + MJAHR + ZEILE, not a surrogate. Conversion exits — the ALPHA exit stores material 1234 as 000000000000001234 and prints it as 1234. Data types that are not what they look like — DATS fields are 8-character strings and an empty date is 00000000, not null. What survives this is a layered model: raw tables land untouched in bronze, client filtering and conversion exits applied once in silver, business entities assembled in gold. Skip it and you apply the ALPHA conversion in fourteen Power BI measures and fix it forever.

A connector does not fix master data — if the same vendor exists four times in LFA1 with different tax numbers, every route replicates that faithfully into OneLake. The gateway is a single point of failure until you cluster it, and a cost centre once you do. Preview features move — Copy job with the SAP ABAP Add-On was announced in preview in June 2026. Third-party tools earn their cost only above a threshold — they solve genuine CDC but add a subscription and a second vendor relationship.

The licensing position can change under you — SAP Note 3255746 has been revised repeatedly. And extraction does not deliver a report: the distance from a landed SAP table to a decision someone acts on is the whole modelling and adoption problem.

Five things, answerable within a week, none requiring a purchase:

• List the SAP tables you actually need for your first analytical area — usually 12 to 25 for inventory or order-to-cash, not "SAP data"
• Record each one's row count and monthly change rate — that sheet decides full extract versus incremental, and whether a third-party tool is justified
• Confirm your SAP release — S/4HANA any version, or ECC 6.0 EhP 8, opens the ABAP Add-On Copy job
• Book 45 minutes with BASIS and agree the extraction window, the parallelism ceiling and the authorisation objects, bringing the table list
• Send the three licensing questions to your SAP account team in writing and file the reply with the architecture document

A sensible first slice: one analytical area, landed through one route, reconciled to the decimal against the existing SAP report before anyone builds a second pipeline. We do the unglamorous part — the client filtering, the conversion exits, the reconciliation — because that is what makes the dashboards true.