Challenge
A new plant, and six systems that each knew part of the story
The plant had been running for a few months when the leadership team started asking questions that no single system could answer. How far are we from nameplate? At this week's rate, how many days of work are sitting in the yard? Which supplier's material is getting worse, and is it costing us yield?
Every answer existed somewhere. Business Central held receipts, purchasing and production orders. The scale house weighed every truck that came through the gate. The line historian logged shredder load, a separate log recorded downtime, and the lab's LIMS held an assay for every lot. Market prices lived in a spreadsheet. A weekly view meant exporting from four of those and matching supplier names by hand, and by the time it was finished the week had moved on.
Finance had a smaller version of the same problem. Receipts and supplier bills were captured in Dext, then typed into Business Central a second time.
Strategy
Put working reports in front of leaders before asking for access
A data platform proposal usually asks for weeks of source access before anyone sees a chart, so the first meeting is spent on what people might want. This one went the other way round. The whole platform was built first, on sample data shaped like the plant's own systems, down to the tables, the keys and the inconsistent supplier names that real scale tickets carry.
That moves the first conversation somewhere more useful. A plant manager looking at a supplier scorecard can say "rank it by grade, not by delivery" in about ten seconds, which is a decision a requirements workshop would have taken an afternoon to reach.
It also limits what changes later. Connecting the real systems replaces the ingestion layer and leaves every table above it alone. Business Central stays the system of record throughout. Databricks joins it with the plant floor, the lab and the market, and takes over none of its jobs.
Solution
One governed lakehouse, nine leadership reports and a Genie space
Seven sources land in bronze through Lakeflow as they arrive, with lineage on every row. Silver types and de-duplicates them, reconciles units and supplier names, and enforces quality rules. Gold holds the tables behind each report with one definition per metric, so "processed" means the same thing on the ramp-up chart as it does in the mass balance. Unity Catalog governs all three layers, which come to 12 bronze tables, 15 silver and 10 gold.
The reports are AI/BI dashboards on the gold layer:
- Ramp-up against nameplate, week by week, on a 14-day rolling average
- Feedstock received against feedstock processed, and the days of work waiting in the yard
- Yard inventory by age and feedstock type
- Line health, and unplanned downtime by cause
- A mass balance that follows every tonne from supplier to output stream
- Recovery over the ramp
- A supplier scorecard covering delivery against contract, rejected loads, grade trend and yield
- Receiving exceptions, which match each Business Central receipt to its scale ticket and flag loads weighed but never received
- Market exposure, which values each production order at that day's price and its own assay, so price and grade effects show up separately
Genie sits on the same gold tables. A leader can type "what caused the most unplanned downtime on Line 1 in August?" and get the answer with a chart and the SQL behind it, so anyone who doubts a number can check it against the query that produced it.
The Finance workstream is smaller, and deliberately boring. Dext has no public API, so Zapier's export-when-ready trigger hands each item to Power Automate. The flow maps supplier, GL account and dimensions from a SharePoint list that Finance owns, then creates the purchase invoice in Business Central with the receipt attached. The Dext item ID is stored as the external document number, which means nothing can post twice. Business Central's own approval workflow decides what gets posted, and anything that breaks a rule goes to a Finance review queue in Teams with the reason. Each item also lands in bronze, so spend sits next to plant cost in the same lakehouse.
Results
Leadership saw its own reports before a single live system was connected
The pipeline, all 37 tables, the dashboards and the Genie space were built and tested end to end. The values in them are samples chosen to behave like a plant in its first months. They aren't measurements of anyone's operation, which is why none of them are quoted here.
What the demonstration settled was scope. Both workstreams go live inside ten weeks, starting with read access to Business Central and exports from the scale house, historian and lab in week one. The Dext flow is built in a Business Central sandbox and runs in parallel with Finance's manual process before it replaces it.
What it sets up
Once the plant's data is in one governed place, the next pieces are extensions rather than new projects. Models on line signals can warn days before a bearing fails. Feedstock can be priced by predicted grade, supplier by supplier, and yard lots sequenced by grade and value against line capacity. Lot genealogy from supplier to customer shipment falls out of the lineage that already exists, and a second site joins the same tables, reports and definitions on its first day.
