Microsoft Fabric for Manufacturing: Honest Notes After Six Months in Production

We have now delivered Microsoft Fabric implementations for five manufacturing clients — three in the GCC (a packaging company in Dubai, a food manufacturer in Riyadh, and an EPC contractor in Abu Dhabi) and two in India (a mid-market auto-components manufacturer in Pune and an FMCG distributor in Hyderabad). The use cases span OEE reporting, predictive maintenance on compressor data, demand forecasting, and supply chain visibility. The ERP environments include SAP ByDesign, SAP Business One, and Dynamics 365.

This is not a theoretical review — it is notes from production deployments, based on what clients actually use, what broke in month two, and what we would do differently on a new engagement. Microsoft's own documentation is accurate but optimistic. The gap between the demo environment and a production manufacturing data environment is significant.

The OneLake architecture is genuinely good. One storage layer across lakehouses, warehouses, and Power BI datasets means you eliminate the data duplication that characterised most Azure data estate designs of the previous five years. The packaging client in Dubai ran four separate Azure Blob storage accounts feeding different downstream consumers. Moving to OneLake cut storage cost by 40% and eliminated three sets of nightly sync jobs that were the source of a persistent reporting discrepancy.

Power BI Direct Lake mode is the most meaningful improvement to the reporting layer in the past three years. Queries that took 8-12 seconds in Import mode take under two seconds in Direct Lake against a properly partitioned Delta table. For a plant manager looking at real-time OEE on a shop floor tablet, that difference is the difference between a tool that gets used and one that gets abandoned.

The Fabric Data Pipeline layer is mature and reliable. We have run ETL pipelines for six months without a failure that was Fabric's fault (there have been failures — all of them were source system issues). The monitoring and alerting through Azure Monitor integration is adequate for production use. The cost management through Fabric capacity units is more predictable than the per-query pricing model in Azure Synapse Analytics.

Fabric Mirroring is the most overmarketed feature in the current release. The pitch is near-real-time replication from SQL databases and Cosmos DB into OneLake without pipelines. The reality in production manufacturing environments is more complicated. Mirroring from Azure SQL works well. Mirroring from on-premise SQL Server (which is where SAP B1 and many Dynamics deployments run) requires Azure SQL Managed Instance as an intermediary or Azure Data Factory as a bridge — which is exactly the pipeline infrastructure that Mirroring was supposed to replace.

Fabric Activator — the real-time alerting and automation component — is genuinely useful but requires more configuration than the demos suggest. The trigger logic is straightforward. The challenge is defining alert conditions in manufacturing terms: you want an alert when OEE on Line 3 drops below 65% for more than 20 minutes, not when any field in a table changes. Building that alert logic requires a structured event stream, a defined KPI calculation, and a threshold management layer that Activator itself does not provide out of the box.

The Fabric AI features — anomaly detection, forecasting, and the Copilot integration — are early-stage for manufacturing use cases. The forecasting PREDICT() function works well on clean time series. In manufacturing, time series is rarely clean: it has shift pattern seasonality, planned downtime events that are not machine failures, material changes that affect consumption rates, and batch campaigns that distort demand signals. Getting the AI features to produce results that plant managers trust requires significant feature engineering that the marketing materials do not acknowledge.

SAP is not a native Fabric connector. There is no "connect to SAP" button in the Fabric interface. The integration requires either an Azure Data Factory SAP connector (for S/4HANA and ECC), direct JDBC or HANA database access (for ByDesign and on-premise systems), or the SAP Business One Service Layer API (for B1). In all cases, this is custom development work, not configuration.

The SAP OData extraction approach — connecting Fabric pipelines to SAP OData services — works reliably for moderate data volumes. For a manufacturer running 50,000 production orders per month, the extraction is fast enough. For a company running 500,000 movements per month, you need a delta extraction strategy: only pulling changed records since the last load. Building a proper delta load on SAP B1 or ByD OData, with watermark tracking and error handling, is a week's work. It is not difficult, but it is not plug-and-play.

Microsoft Fabric is the right platform for most mid-market manufacturing analytics in 2026. The OneLake architecture solves real problems. Direct Lake is a genuine step change in reporting performance. The platform is stable enough for production use, and the Microsoft investment in the roadmap means it is unlikely to be deprecated or significantly restructured in the next three to five years.

The implementation, however, is not as simple as Microsoft's documentation implies. The SAP integration requires custom development. Mirroring from on-premise systems requires additional infrastructure. The AI features require data science expertise to tune for manufacturing-specific patterns. Budget six to ten weeks for the foundation build, plan the SAP integration as a separate workstream, and do not rely on Mirroring as a substitute for a properly designed extraction architecture.

The clients who have had the most success with Fabric in manufacturing are the ones who treated the first three months as a data foundation project — getting the SAP integration right, building the governing semantic model — rather than going straight to the dashboard layer. The clients who went directly to dashboards found themselves rebuilding the foundation later, at higher cost and with lower team morale.