Power BI Consulting in Riyadh for Manufacturing Companies: What the Work Involves

The meeting that starts most of these engagements looks the same in Riyadh as anywhere. A group operations director has three numbers for last month's output — one from the plant manager's Excel, one from a Power BI report built by the ERP team, one from finance — and they do not agree.

Power BI is not missing. It is usually everywhere. That is the actual problem: the reports are not wrong at random, they are wrong systematically, because each one encodes a private definition of yield, downtime, on-time delivery and cost per tonne.

It is mostly remediation, not new build: inventorying an existing report estate, consolidating many semantic models into a small number of governed star schemas, agreeing metric definitions with operations and finance, implementing row-level security, and making refresh reliable.

In delivery terms a first engagement runs Discover, Prototype, Deploy, and the first two settle the argument. Discover (1–2 weeks): run the Power BI admin API or metadata scanner across the tenant and produce a real inventory — workspaces, models, reports, refresh history, failure rates, gateway dependencies, orphaned owners. Prototype (2–3 weeks): pick one contested metric — OEE, OTIF, scrap rate, cost per tonne — and build one governed model that produces it correctly for every plant. Deploy (2–4 weeks): move the agreed model into a governed workspace, apply RLS, set deployment pipelines, retire the reports it replaces, and hand over documented DAX. A consultant who starts with visuals produces a prettier version of the disagreement.

Five concrete workstreams:

The most common technical cause of a distrusted estate is that each author connected Power BI Desktop straight to an ERP extract and built a private model. Agreed metric definitions are a facilitation job disguised as a technical one. RLS matters because Saudi industrial groups are frequently several legal entities under a holding company, sometimes with JV partners entitled to see one plant and not others. And refresh is where credibility is actually lost — a report showing stale data looks identical to one showing current data.

Most mid-market Riyadh manufacturing estates do not need a lakehouse on day one. Where the problem is inconsistent definitions across ERP data that fits in an import model, a governed Power BI semantic model on a dataflow or warehouse layer solves it — I say this against my own commercial interest.

The triggers that genuinely justify Microsoft Fabric and a OneLake lakehouse are specific: fact tables past roughly 100–200 million rows where import refresh windows break; OT data arriving via OPC-UA or MQTT at a frequency Power BI cannot sensibly ingest; multiple heavy source systems needing engineering before reporting. Migrating an inconsistent estate to Fabric produces the same disagreements on a larger capacity bill.

There is no live Azure region inside Saudi Arabia as of August 2026. Microsoft has confirmed the Saudi Arabia East region, with three availability zones, will be available for customer workloads from Q4 2026. This is the fact most often stated incorrectly in Saudi procurement, in both directions.

Two consequences: a Riyadh manufacturer wanting regional hosting today with full Fabric workloads (not Power BI alone) is choosing UAE North; and you should check your own tenant's home region (Help → About → "Your data is stored in") before anyone writes a residency clause into a contract.

Saudi Arabia's Personal Data Protection Law, issued under Royal Decree M/19 and enforced by SDAIA, has a narrow but real impact on a Power BI programme. Most manufacturing reporting data — production counts, downtime reasons, batch yields, machine states — is not personal data. The exposure sits in the datasets people forget: labour hours by named operator, safety incident records, driver telematics, visitor and gate-access logs.

Where personal data moves outside the Kingdom, the Regulation on Personal Data Transfer sets the conditions. In delivery terms that is three tasks: classify which semantic-model tables contain personal data, establish a lawful transfer basis, and document a risk assessment. One scoping point that saves money: the NCA's Cloud Cybersecurity Controls, requiring in-Kingdom hosting, apply to government organisations and Critical National Infrastructure operators — not to every private manufacturer, so confirm whether they actually apply to you before over-engineering residency.

Power BI supports Arabic in the browser-based service, but Microsoft documents that Power BI Desktop is not available in Arabic or Hebrew because Desktop does not support right-to-left languages, and that inside reports the layout of visuals does not flip for a right-to-left language.

For a bilingual Riyadh estate that means: metadata translations localise table, column and measure names in the model (tractable, done once); report label translations cover titles and captions, but hard-coded text in the report layout cannot be localised and page-tab names cannot be translated; data translation — Arabic product or customer names in the data itself — is the hardest and needs additional columns modelled into the source; and layout must be authored deliberately, with right-aligned cards, Arabic-capable fonts and locale number/date formats. Budget this as design work — it is routinely underestimated because the service being available in Arabic gets mistaken for reports being bilingual.

A Microsoft partner badge does not distinguish a team that has run a production Power BI estate from one that builds demos. The criteria that separate them: practitioner depth over badge count (ask who writes the DAX); the team that pitches is the team that delivers (names in the contract); willingness to fix scope (T&M-only on a remediation transfers discovery risk to you); production evidence not demos (ask their worst refresh failure and how it was found); industry data experience (shift calendars, batch genealogy and downtime hierarchies are not generic modelling); and honest limitations during the sale.

A consolidated model does not fix bad master data — if the same customer exists three times with three codes, the model faithfully reports three customers. Governed models do not stop shadow reporting — people rebuild in Excel when the governed report is slower to answer than a pivot table. RLS is not a substitute for workspace governance — it does not apply to workspace Admin, Member or Contributor roles, so an over-permissive workspace silently defeats it.

A remote delivery model has limits on the plant floor — semantic modelling, DAX and refresh work well remotely; reconciling MES and ERP disagreements needs someone on site. Neither Power BI nor Fabric makes you PDPL-compliant — they are tools inside a compliance position that requires classification, a lawful transfer basis and a documented risk assessment. And a regional Azure region does not automatically move your tenant: when Saudi Arabia East arrives, your existing home region does not change by itself.

Answer these five questions this week, before you brief anyone:

• How many semantic models in your tenant contain a measure named some variant of OEE, yield or on-time delivery — and do they agree?
• What percentage of published reports have been opened in the last ninety days?
• What is your tenant's home region, and does any contract you have signed make a residency commitment inconsistent with it?
• Which semantic-model tables contain personal data of employees or contractors, and where does that data physically sit?
• Which single contested metric, if everyone agreed on it, would shorten your monthly operations review the most?

That last answer is your first-slice scope: one metric, one governed model, every plant, six weeks. Disclosure — MyData Insights is a practitioner-led practice in the last row above; apply the same criteria to it that you apply to anyone else.