Power BI in Saudi Arabia: From Shelfware to AI-Driven Reporting with MCP Agents

Walk into ten Saudi mid-market companies that signed Power BI Pro seats during the post-2022 Vision 2030 BI buildout. In nine of them, the same pattern: a 30-tab workspace built by a consultant, three dashboards that the CEO opened once, and a $20-per-user-per-month line item that no one has the budget signature to cancel.

This is Power BI shelfware, and it is a uniquely Saudi problem. Office 365 enterprise penetration in KSA is among the highest in MENA. ZATCA Phase 2, the Saudi Open Data initiative, and bank-grade reporting mandates pushed every CFO toward a Microsoft BI footprint. The licenses got bought. The activation work — the part where the dashboards actually drive decisions — mostly did not happen.

The good news: the gap between a dormant Power BI tenant and an AI-driven reporting layer that the business uses every day is not a fresh six-figure consulting engagement. It is an activation pattern built on Model Context Protocol (MCP) servers, agent-pushed narratives, and Arabic-native generation that sits on top of the semantic model you already paid for.

Why Power BI Shelfware Is the Default Outcome in KSA

Three forces converge to leave dashboards unused:

1. The dashboard-as-deliverable trap. A System Integrator shows up, asks for KPIs, builds 22 visuals across 4 tabs, hands over a PBIX file, and invoices. Adoption is not in scope. Six months later the data refresh has broken twice, the executive sponsor changed roles, and the workspace is read-only by default.

2. Power BI rewards browsing, not deciding. A dashboard answers "what is the number?" — it does not answer "what should I do?". For a regional director comparing 14 branches in Riyadh, Dammam and Jeddah, the cognitive cost of opening Power BI, picking filters, and synthesising a narrative is higher than asking a finance analyst to send a WhatsApp summary. So the analyst becomes the dashboard.

3. Arabic narrative generation is weak in Copilot. Power BI Copilot, Microsoft's native AI summarisation layer, produces English output of acceptable quality and Arabic output that ranges from "stiff" to "unusable for a board report." For a KSA business where the audit committee reads Arabic and the auditors file in Arabic, the AI layer that ships with the product does not meet the language bar.

The result is a tenant full of correct numbers that nobody consults. The cost of inaction is the licensing bill plus the salary of the analyst who keeps doing manually what the platform was supposed to automate.

The Activation Pattern: AI Agents on Top of Power BI

The activation play is not "rebuild your dashboards." It is add three layers:

• A semantic-model bridge — an MCP (Model Context Protocol) server that exposes your Power BI datasets to a language model the way the dataset designer intended (measures, dimensions, RLS).
• A narrative agent — an AI agent that can be asked "summarise April for the Eastern Province retail division in Arabic, flag anomalies, and give me three things to ask the regional GM" — and produces a one-paragraph answer grounded in the dataset.
• A push runtime — scheduled workflows that send Arabic monthly narratives to Teams, email or WhatsApp instead of waiting for a human to open a dashboard.

None of these replace Power BI. The dashboards stay where they are. The activation layer makes them useful to people who currently ignore them.

Layer 1 — MCP Server for Power BI

MCP is the open protocol Anthropic published in late 2024 to let language models call tools and read structured data via a uniform interface. For Power BI, an MCP server typically exposes:

• list_datasets — workspaces and datasets the agent can query
• list_measures — measures defined in the semantic model
• query_dataset — execute a DAX or natural-language query and return a typed result
• get_dataset_metadata — table schema, relationships, RLS roles

Authentication uses a Microsoft Entra service principal with read-only Power BI Service permissions, scoped per workspace. Row-Level Security (RLS) defined in the semantic model is respected automatically — the agent sees the same slice a user with the same security role would see.

For a Saudi bank or government entity where PDPL compliance is on the line, this is the safest place to plug AI in: the agent never sees raw transactional data outside the semantic model's boundaries, and audit logs in Power BI capture every query the agent executes.

Layer 2 — The Narrative Agent

The agent is where the value shows up. A well-built narrative agent for KSA looks like this:

• Reads the dataset via the MCP server
• Has a persona system prompt that defines tone (formal Arabic for board reports, conversational for ops standups)
• Knows the business context (fiscal calendar, branch geography, regulatory deadlines)
• Has a fact-checking step — every numeric claim in the narrative is regenerated from the dataset before being included
• Outputs Arabic-native text (not translated from English) when configured for Arabic mode

For a Saudi retailer, the agent's monthly output might be a 6-paragraph Arabic narrative covering same-store sales by region, top SKU movers, returns flagged for ZATCA review, and three plain-language anomalies ("Riyadh North branch closed for renovation 8 days — adjust YoY comparison").

This is what Copilot does not do well: the synthesis and the language.

Layer 3 — Push, Don't Pull

The third layer is the runtime: the activation only works if the narrative shows up where executives already are. In KSA, that is Teams (for enterprise) and WhatsApp (for ops and field teams).

Typical scheduling:

• 06:30 daily — Arabic operations brief to regional WhatsApp groups
• Monday 08:00 — weekly performance summary to executive Teams channel
• Month-end + 2 days — board-grade Arabic narrative emailed to CFO with PDF attachment

The dashboards stay available for the rare moment someone wants to drill. But 90 percent of the consumption shifts to the push layer, where it matches how Saudi executives already work.

Cost Comparison: Power BI Copilot vs Custom MCP Activation

The honest tradeoff:

The math turns favorable on custom MCP for any organisation where the executive Arabic narrative is a recurring deliverable. If the only consumer of insights is one English-speaking founder, Copilot is fine. If twelve regional managers need a monthly Arabic narrative their CFO can sign off on, the custom layer pays back in months.

Where This Pattern Fits in KSA

Three sectors where the activation pattern delivers fastest in Saudi Arabia:

Retail and quick-service — Branch performance varies sharply across regions. Owners want a regional brief, in Arabic, before the morning meeting. ZATCA Phase 2 e-invoicing dataset is naturally Power BI-shaped. Anomaly detection (refund spikes, missing buyer IDs, duplicate invoices) is high value.

Banking and finance — Risk dashboards exist but are consumed by analysts, not by branch managers. An Arabic narrative agent that walks a branch manager through their portfolio health weekly is a behaviour-change tool, not a reporting tool.

Government and semi-government — Vision 2030 KPIs are tracked centrally but rarely cascaded into department-level narratives. PDPL compliance means the data stays inside the Microsoft tenant — MCP architecture suits this because the agent runs against the semantic model, not raw PII.

What Noqta Does — And What We Do Not

To be explicit, because this matters in a market full of "AI consultants" who promise everything:

We do:

• Build MCP servers against existing Power BI semantic models
• Train narrative agents on customer-specific personas and Arabic tone of voice
• Set up push runtimes (Teams, WhatsApp, email) with scheduling and audit logs
• Integrate non-PBI data sources (Odoo, custom ERPs, ZATCA e-invoice APIs) into the same agent

We do not:

• Build your Power BI dashboards from scratch — your existing SI or in-house team owns that
• Re-architect your data warehouse
• Sell Microsoft licensing
• Replace your finance team

This positioning matters because it is the only way the engagement stays small, fast, and high-ROI. We sit on top of what you already paid for.

When Not to Do This

A few honest disqualifiers:

• You do not have Power BI in production yet. Solve that first; activation comes after.
• Your Power BI workspace has data quality problems. No AI layer fixes garbage in. We audit the semantic model first — if measures are broken, we flag it and pause.
• You need real-time second-by-second data. Power BI is not a streaming platform. The activation pattern inherits that constraint.
• You are a sub-50 employee company. The fixed setup cost does not amortise. Use Copilot.

Getting Started: A Three-Week First Project

A typical first activation engagement for a Saudi mid-market client:

• Week 1 — Audit existing Power BI tenant. Map workspaces, datasets, RLS roles. Pick the dataset where an Arabic narrative would deliver the most decision-making value. Define one persona.
• Week 2 — Build the MCP server, wire it to the chosen dataset, run the agent in a sandbox. CFO reviews three test narratives in Arabic.
• Week 3 — Push runtime setup. Schedule first monthly narrative. Onboard the executive recipients. Hand over runbook.

After three weeks you have one workflow live, one persona productionised, and a measurement of how much time the activation saves the analyst who was doing the manual synthesis. Subsequent personas (sales, operations, board) ship in days, not weeks, because the MCP server is already in place.

Frequently Asked Questions

Do we need to migrate off Power BI? No. The activation pattern explicitly preserves your existing dashboards, datasets, and licensing. We add three layers on top.

Is this just Power BI Copilot under a different name? No. Copilot is Microsoft's built-in summarisation feature limited to what Microsoft ships. A custom MCP agent has custom personas, enforced fact-checking, Arabic-native generation, and push-to-Teams/WhatsApp scheduling — none of which Copilot does today.

Will the AI hallucinate numbers? The narrative agent is built with a regeneration step: every numeric claim in the output is re-queried from the dataset before the narrative is finalised. Hallucination of numbers is structurally prevented. Hallucination of judgment (e.g. "this is good performance") is constrained by the persona prompt and reviewed during onboarding.

How does this work with PDPL and ZATCA compliance? The MCP server uses a Microsoft Entra service principal scoped to read-only Power BI Service access. Row-Level Security is respected. Audit logs in Power BI capture every query. The agent runs inside your Microsoft tenant boundary if you self-host the runtime, which is the standard PDPL-compliant configuration.

Can we trial this without committing to a full project? Yes — we offer a free audit of your existing Power BI tenant. We map the dormant assets, identify the highest-ROI activation, and give you a fixed-price scope for a three-week first project.

Does this work with Arabic dashboards or only English? The MCP layer is language-agnostic — it reads the semantic model. The narrative agent is configured per-language. We typically run two narrative modes for Saudi clients: formal Arabic for board and audit output, and conversational English or Arabic for operational briefings.

What about Tableau, Qlik, or Looker? The same activation pattern applies — MCP server, narrative agent, push runtime. The connector layer differs. For Saudi clients the conversation is almost always about Power BI because that is what is deployed; for other platforms the engagement structure is identical.

If your Power BI tenant looks more like shelfware than a decision-making layer, the fix is not to buy more licenses or commission a new dashboard. It is to activate what you already own with an AI layer that produces Arabic narratives, pushes them where executives already are, and respects the security boundaries Microsoft already gave you.

Book a free Power BI activation audit →