Microsoft Fabric in a Live Manufacturing Estate

It is 06:00 on a Monday morning at a mid-market manufacturing site. The shift supervisor opens the operations review. The data is current. The overnight ADF pipeline ran at 03:15. The OEE figure for Line 3 is sitting at 61% — 14 points below the 75% target. A downtime event logged at 02:40 is flagged. By 07:15, the plant manager has asked a question in plain English and has an answer. By 07:30, a corrective action is open and assigned.

That is what Microsoft Fabric looks like in practice. Not a platform diagram. An operating cadence.

The data movement layer runs overnight. Azure Data Factory pulls production batch records from SAP S/4HANA — confirmed production orders, goods movements, quality inspection results. For businesses running SAP ByDesign, the same ADF pipelines pull sales orders, financial postings, and inventory levels. Where a MES sits on the plant floor — capturing machine state, cycle time, and reject counts — a separate ADF pipeline or an OPC-UA / MQTT stream lands that data in OneLake alongside the ERP records.

Everything lands in OneLake in Delta Parquet format. Delta Parquet is an open table format — it is not a proprietary silo. If another system needs to read those records, it can, without moving the data. The quality inspection results from SAP S/4HANA and the reject count from the MES sit in the same lake, with the same governance boundary managed by Microsoft Purview.

By 03:45, the data is there. By 04:00, the Power BI semantic layer has access to it via Direct Lake — no Import refresh, no scheduled pipeline to wait for. The OEE, OTIF, scrap rate, and downtime root-cause classifications are available to the model as soon as ADF finishes writing.

The morning operations review dashboard opens in Power BI. It is running in Direct Lake mode — which means it is reading the data that landed at 03:45, not the data from the previous evening's Import refresh. For a site that runs a night shift, the difference is meaningful. A downtime event that occurred at 02:40 is visible at 06:00. It would not have been visible until the following evening under an Import-mode dataset on a 22:00 refresh schedule.

The dashboard shows four metrics on the first page.

**OEE by line** — Overall Equipment Effectiveness, calculated from the MES availability, performance, and quality data joined to the SAP production order. Line 3 is at 61%.

**OTIF by customer** — On Time In Full, calculated from the SAP S/4HANA delivery records against the promised date on the sales order. Two customers are showing a developing OTIF miss for the current week.

**Scrap rate by product family** — calculated from the SAP quality inspection results. One product family is running at 4.2% scrap against a 2% target.

**Downtime root cause** — a ranked list of downtime categories from the MES, sorted by total minutes lost in the past 24 hours.

No manual consolidation. No Excel file built at 05:30. No phone calls to the previous shift to find out what happened.

The plant manager types a question into Copilot in Power BI: "Show me the top three root causes for Line 3 downtime in the last 48 hours."

Copilot reads from the semantic model. The model has a well-defined Downtime measure, a RootCause dimension sourced from the MES classification codes, and a correctly configured fiscal calendar. The answer comes back in seconds — a ranked bar chart: conveyor jam (43 minutes), changeover overrun (38 minutes), unplanned maintenance (27 minutes).

That is a useful answer because the model is well built. If the RootCause dimension had ambiguous codes — "other," "mechanical," "TBD" — the Copilot answer would be accurate but operationally useless. The intelligence is in the data model, not in the AI layer.

Power Automate has already acted on the OEE drift before the plant manager asked the Copilot question.

A Power Automate flow monitors the OEE measure in the Fabric dataset. When OEE on any line drops below 65% for two consecutive hours, the flow triggers. It sends an alert to the plant manager and the maintenance supervisor on Teams. It creates a corrective-action record in Power Apps — the field operations application the site uses to log and track corrective actions. It assigns the record to the maintenance lead responsible for Line 3.

By 07:30, the corrective-action record exists, is assigned, and has a due date. The plant manager did not need to create it manually. The OEE drift was captured at 02:40, the alert fired at 03:00, and the corrective-action record was waiting in Power Apps when the maintenance supervisor started their shift at 06:00.

That is the Act layer of the Unify · Predict · Act model working in a live manufacturing context.

Fabric replaces the analytics layer — the data warehouse, the Import-mode datasets, the manual morning reports built in Excel. It does not replace the people running the operation.

The shift supervisor who interprets the OEE figure, understands which line is underperforming for structural reasons versus one-off events, and decides whether to escalate or manage it within the shift — that judgment is not automated. Copilot provides faster access to the data. It does not provide operational experience.

The MES or SCADA system on the plant floor continues to do what it does — managing machine state, capturing cycle time, controlling process parameters. Fabric ingests from it. It does not replace it.

The master-data stewards who maintain the production routing in SAP S/4HANA — the BOM structures, the work centre definitions, the quality inspection plans — continue to be necessary. If the routing in SAP is wrong, the OEE calculation built on top of it is wrong. Fabric will calculate and display the wrong figure efficiently and consistently. That is not an improvement.

The 06:00 frame applies equally to an FMCG distribution centre or a 3PL warehouse.

For an FMCG site, replace OEE with fill rate and case pick accuracy. The SAP ByDesign delivery notes and the WMS pick records land in OneLake overnight via ADF. The morning dashboard shows fill rate by SKU, pick accuracy by zone, and vehicle departure compliance against the dispatch window. A Power Automate alert fires when a shipment misses its dispatch window — and creates a notification in the customer's delivery exception log automatically.

For a 3PL operator, replace scrap rate with mis-pick rate and dock utilisation. The corrective-action workflow in Power Apps captures the mis-pick investigation. The OTIF dashboard shows performance by customer account. The Operations Director's weekly review uses the same data as the client-facing performance report — because there is one dataset, not two reconciled spreadsheets.

A mid-market FMCG manufacturer in the GCC running SAP S/4HANA and a third-party MES engaged MyData Insights to build this operational layer on Microsoft Fabric. The presenting problem was a 4–6 hour reporting lag — the morning operations review was built manually by the planning team from overnight exports, and was consistently incomplete by the time the 07:00 site meeting started.

Within six weeks, OneLake held the canonical operational data from SAP S/4HANA and the MES, ADF pipelines ran at 03:00, Power BI Direct Lake delivered OEE, scrap rate, and OTIF figures current to within 90 minutes of the run, and Power Automate routed the OEE-drift alert to the maintenance supervisor automatically. The 4–6 hour lag was gone. The manual morning assembly was gone. The planning team reclaimed roughly 90 minutes per shift-handover.

If the MES or SCADA system on the plant floor does not have an accessible data output — no API, no file export, no OPC-UA feed — the data engineering work required to connect it to OneLake is a separate project. That work needs to be scoped at the OT/IT integration layer before Fabric is relevant.

If the site is still running a fully on-premises ERP with no cloud connectivity, the ADF integration approach requires a self-hosted integration runtime and network configuration that adds complexity. It is solvable — but it is not a Fabric problem, it is an infrastructure problem that sits upstream of Fabric.

Fabric replaces the analytics layer. It does not resolve infrastructure that was never built to be connected.

Discover: map the SAP S/4HANA or SAP ByD source tables, the MES or SCADA output format, and the three metrics the Operations Director reviews every morning. Prototype: build the ADF pipelines to OneLake, construct the Direct Lake Power BI model for those three metrics — OEE, OTIF, scrap rate — and demonstrate the 06:00 live operations review with no manual assembly required. By week six, the shift supervisor has a dashboard that is current at shift start and the operations team has a Power Automate alert for OEE drift.