Agentic BI on Microsoft Fabric: Architecture and Implementation
Traditional BI is a passive model: a user asks a question, an analyst or a report answers it. Agentic BI flips that logic — the agent queries data, reasons over context, and delivers actionable insights where decisions are made, with no human in the loop.
This article walks through the Agentic BI architecture on the Microsoft stack, its key components, and the business value it unlocks.
Architecture overview
The architecture is organized into three vertical layers: Data, Intelligence, and Output.
flowchart TB
subgraph OUT["🖥️ OUTPUT — Microsoft Teams"]
direction LR
C1["💬 Copilot chat\nNL Q&A"]
C2["🔔 Smart alerts\nProactive nudges"]
C3["📋 Insight cards\nRoot cause + next"]
C4["⚡ Agentic actions\nOne-click execute"]
end
subgraph INT["🧠 INTELLIGENCE"]
direction LR
subgraph AF["Azure AI Foundry — pro-code"]
LLM["LLM Reasoning\nGPT-4o / Azure OAI"]
AS["Agent Service\nTool use + memory"]
end
subgraph CS["Copilot Studio — low-code"]
OR["Orchestration\nNL flows & triggers"]
CN["Connectors\nPower Platform · APIs"]
end
AF <-->|"⇄"| CS
end
subgraph DAT["📊 DATA — Microsoft Fabric"]
direction LR
subgraph SRC["Data Sources"]
LW["Lakehouse /\nWarehouse"]
PB["Power BI\nSemantic Model"]
RT["Real-Time\nIntelligence"]
MR["Mirrored DB /\nSQL Analytics"]
end
DA["🤖 Fabric Data Agent\nNL → T-SQL · KQL · DAX"]
subgraph CTX["Context & Intelligence"]
IQ["Fabric IQ\nContextual AI"]
GR["Graph\nRelationships"]
ON["Ontology\nSemantic schema"]
end
end
SRC --> DA
CTX --> DA
DA --> AF
DA --> CS
AF --> OUT
CS --> OUT
style OUT fill:#1a2a4a,stroke:#4a9eff,color:#fff
style INT fill:#2a1a4a,stroke:#9a4aff,color:#fff
style DAT fill:#1a3a2a,stroke:#4aff9a,color:#fff
style AF fill:#1e1040,stroke:#7a4aff,color:#fff
style CS fill:#2a1040,stroke:#b04aff,color:#fff
style SRC fill:#0f2a1a,stroke:#4aff9a,color:#fff
style CTX fill:#0f2a1a,stroke:#4aff9a,color:#fff
style DA fill:#0a3a1a,stroke:#4aff9a,color:#fff,font-weight:bold
DATA layer — Microsoft Fabric
Microsoft Fabric is the foundation of the architecture. It centralizes all data assets in a single governed platform.
| Component | Role |
|---|---|
| OneLake | Unified storage — one copy of the data across the entire platform |
| Lakehouse / Warehouse | Structured Delta tables, SQL endpoints — the agent’s primary query target |
| Power BI Semantic Model | Certified business metrics, pre-built KPIs and hierarchies |
| Real-Time Intelligence | Streaming data via Eventhouse / KQL DB for live anomaly detection |
| Copilot in Fabric | Native AI assistance built into the Fabric experience |
The Fabric Data Agent
At the core of the system sits the Fabric Data Agent — a natural language query engine that can:
- Interpret NL questions and translate them into T-SQL, KQL, or DAX
- Reason across multiple data sources in a single query
- Use an Ontology layer to resolve semantic ambiguities (e.g. “Region” contains “Store” owns “Sales”)
- Leverage Microsoft Graph to contextualise data with organisational relationships (users, teams, documents)
- Be enriched by Fabric IQ, the native AI layer that surfaces proactive insights and usage recommendations
INTELLIGENCE layer — Pro-code vs Low-code
This is where the agent’s reasoning is built. Two approaches coexist and complement each other.
Azure AI Foundry (pro-code)
The reference runtime for building custom agents using the Fabric SDK, Python, and Azure OpenAI — full control over reasoning and tool use.
- LLM reasoning: GPT-4o / Azure OpenAI
- Agent service: tool use management and context memory
When to use it: complex business logic, multi-step agents, external system integrations, fine-grained control over prompts and model behaviour.
Copilot Studio (low-code)
A low-code builder to expose the Fabric Data Agent as a conversational bot and orchestrate automated workflows via Power Automate.
- Orchestration: NL flows and automated triggers
- Connectors: Power Platform, third-party APIs
When to use it: fast exposure to business users, workflow automation, scenarios without complex reasoning logic.
The two approaches are not mutually exclusive — an agent built with AI Foundry can be surfaced through Copilot Studio.
OUTPUT layer — Microsoft Teams
The agent delivers its outputs directly in Teams — where users already work, with no new tool to learn.
| Capability | Description |
|---|---|
| Copilot chat | Natural language Q&A over Fabric data |
| Smart alerts | Proactive notifications triggered by thresholds or detected anomalies |
| Insight cards | Contextual cards with root cause analysis and next-best-action recommendation |
| Agentic actions | One-click action execution directly from the insight card |
| M365 publish | Direct deployment across the Microsoft 365 ecosystem |
Why Agentic BI?
1. Faster insights
From question to root cause in seconds — no waiting for an analyst or an ad hoc report.
2. Business user autonomy
Non-technical users get answers in natural language, directly in their daily work environment.
3. From insight to action
Recommendations surface where decisions are made. The action can be triggered in one click from the Teams notification.
4. Built on existing data
The architecture leverages the Lakehouse, Semantic Model, and Power BI reports already in place — no need to start from scratch.
What this actually changes
The shift is fundamental: from pull BI (the user goes looking for information) to push + act BI (the agent detects, notifies, and proposes the next step). Data no longer sits in a dashboard — it reaches the right person at the right moment, with context and next action already formulated.
The Microsoft stack — Fabric, AI Foundry, Copilot Studio, and Teams — provides the building blocks to close this loop end to end, with a level of governance and security that meets enterprise requirements.