CIO Walkthrough¶

Audience: a CIO and their leadership team. Goal: tell the story of Fabric as the operating system for a multi-agent retail business — one platform, an orchestrator that picks the right brain per question, and a fleet of agents that turn insight into governed, approvable action. Run time: ~15 minutes for the core path, ~25 with the deep cuts.

In this walkthrough we'll go end-to-end: we'll level-set on what's deployed and why the ontology matters, see the business in a Power BI report, ask it questions in plain English and watch the orchestrator choose between the semantic model, the relationship graph, and the live event stream, then act on it with two operations agents that draft real, costed, approvable decisions — and finally close the loop by watching those decisions stream back into Fabric as real-time telemetry. No ETL pipelines on stage, no black boxes — every answer shows its work.

Before we start — what's deployed, and why it's useful¶

Before the demo, let's level-set on what this is. Everything lives in one Fabric workspace — one security boundary, one billing meter, one copy of the data in OneLake. Four capabilities sit on that single foundation:

The point: these are not four products stitched together. The same Gold tables feed the report, the agents, and the models. The same Eventhouse that streams live sales also captures what the AI agents decide.

Why the ontology matters — the part that's hard to copy. A semantic model is great at aggregates ("net sales by region"). It is not built to pull together everything we know about one specific entity — a single customer's profile, their ML segment, and their churn prediction — into one connected answer. The typed business graph is. Name one customer by loyalty card and the ontology fuses their business attributes and their ML predictions in a single hop. Broad aggregations and deep scans, meanwhile, are routed to the semantic model, which handles them far more reliably — and the orchestrator picks the right brain automatically.

For the full architecture, see Agentic Application.

The interface — what's on screen¶

Two views do the heavy lifting:

• Dashboard — the Power BI report (≈70%) side-by-side with the Ask-your-data chat (≈30%). The CIO sees the visual and the conversation in one frame. There is no "ask the data agent vs. ask the ontology" toggle — the user just asks, and a background orchestrator routes each question.
• Agent Operations — a live approval-funnel dashboard that streams from the Eventhouse agent_actions table and auto-refreshes every few seconds.

And under every answer, a collapsible "how I reached this" trace shows the source, the router + the exact model that decided, the call made, the graph path or DAX, and a result preview. That's the CIO trust requirement, built in.

The walkthrough — see it, ask it, act on it¶

The flow has three acts: See the business → Ask the business → Act on the business.

Act 1 — See it (Power BI, ~3 min)¶

Open the Dashboard page, let the report load, then expand 2–3 visuals:

1. Executive sales & margin — net sales, gross margin %, channel mix. Sets the scale.
2. Inventory / stockout-risk visual — surfaces that some stores are running thin. The hook for Act 3, Agent 1.
3. Customer / churn visual — segments and at-risk value. The hook for Act 3, Agent 2.

Narrate: "This is the curated semantic model — one version of the truth. But a dashboard only answers the questions someone built a chart for. Watch what happens when I just… ask."

Act 2 — Ask it (chat, ~5 min)¶

Ask these in order. After each answer, expand the "how I reached this" trace so the room sees which brain answered and why.

Narrate over the traces: "Notice the app is choosing a different brain each time — the semantic model for the metric, the graph for the relationship, the live stream for right-now — and it's telling me which one and why. I never had to pick."

Act 3 — Act on it (operations agents, ~6 min)¶

This is the moment that lands. Ask a question that implies a decision, and the orchestrator hands off to a specialized operations agent that reads the real data, applies business rules, and drafts an approvable action.

Agent 1 — Inventory & Replenishment. Ask: "What products are at risk of stockout — and what should we do about it?"

The audience sees the answer quantify the exposure (~30 SKUs across 15 stores, ~$62K/day of sales exposed, 8 with no open reorder), the trace show the agent read the ML stockout-risk table → cross-check fact_reorders → join dim_products for price → apply a "14-day cover" rule, and a recommendation card with an evidence table and 8 drafted reorders, each with Approve / Dismiss. Narrate, then click Approve on one.

Agent 2 — Customer Retention. Ask: "Customer churn is rising — what can we do to retain them?"

The audience sees 8,286 customers predicted to churn = ~$534M of lifetime value at risk, the trace show the agent read the ML churn table → join customer_segments for LTV → select the top cohort → pull cross-sell anchors from the market-basket model, and a drafted win-back campaign with Approve / Dismiss.

Narrate: "Two different agents, same pattern: perceive the data, reason with business rules, recommend a specific action, and require a human approval. This is a multi-agent operations layer, not a chatbot."

Validate it — close the loop¶

Now let's prove it's a system, not a demo toy. Every agent decision and every human approval is written back into the Fabric Eventhouse as an event in the agent_actions table — the same real-time store that streams live sales.

Switch to the Agent Operations tab after Act 3. The CIO watches, in real time: how many actions the agents proposed, how many operators approved, the dollar value of approved replenishments and addressed churn LTV — each new draft and approval appearing within ~2 seconds. Behind it, it's plain KQL anyone can put on a Fabric Real-Time Dashboard:

agent_actions
| summarize actions = count(), ltv_at_risk = sum(ltv_at_risk) by agent, action_status
| order by agent asc

The takeaway: the AI isn't a black box bolted on the side. Its recommendations and your team's approvals become first-class, governed, queryable telemetry inside Fabric — auditable, measurable, and ready to drive the next dashboard or the next automation.

Recap — the narrative arc¶

In this walkthrough we showed a living digital twin of a national retailer on one Fabric workspace, with a team of AI agents on top of it that don't just answer questions — they recommend and execute the next best action, with a human in the loop, and write every decision back into the platform as real-time telemetry. To tell it in five beats:

1. "One platform." Sales history, the relationship graph, the live event stream, and the ML models are all in one Fabric workspace, one copy of the data.
2. "Ask, don't build." A dashboard answers the chart someone built. Here you ask in plain English and an orchestrator picks the right brain — and tells you which one.
3. "From answers to actions." Specialized operations agents turn an insight into a specific, costed, approvable action — with a human in the loop.
4. "Closed loop, governed." Every recommendation and approval is logged back into the Eventhouse and visible on a real-time dashboard — measurable and auditable, not a science project.
5. "This scales." The same pattern adds a Pricing agent, a Logistics agent, a Marketing agent — each reading the same Fabric foundation. Fabric is the operating system for a multi-agent retail business.

Appendix — deep-cut questions (if time allows)¶

These deliberately span the semantic model + ontology + ML in one ask:

• Revenue-at-risk: "Quantify the total lifetime value tied to high-churn-risk customers — dollars on the table if we do nothing." → the Retention agent's $534M headline, with a drafted campaign.
• Stockout → sales impact: "For products at high stockout risk, estimate the at-risk daily sales." → the Inventory agent's ~$62K/day, with drafted reorders.
• Best next action per store: "For each underperforming store, surface its top stockout risk, top churn-risk customers, and top cross-sell — one prioritized action list." → the roadmap slide: where the agent fleet is going.

Appendix — pre-flight checklist¶

• Fabric capacity Active (F64 auto-pauses overnight — resume before the demo).
• Local app running at http://127.0.0.1:8080; az login as the workspace admin.
• Streaming notebook running so "last 15 minutes" returns fresh data.
• agent_actions table and the proposals store cleared for a clean approval funnel (so the live numbers start at zero in front of the audience).
• One dry run of the Inventory and Retention questions (the first DAX reads take a few seconds).