Microsoft Fabric
Microsoft Fabric is the unified data platform where every workload reads from and writes to OneLake, so data never moves between engines. That is why its AI layer matters: the Data Agent, Eventhouse, and Ontology all operate directly on data where it already lives, rather than on a copy.
When to choose Fabric over Foundry: data already lives in Fabric or OneLake, users need NL query over Fabric data, or you need AI over real-time streaming data.
Architecture
OneLake
The storage foundation. Every Fabric workload - Lakehouse, Warehouse, Eventhouse, Power BI - reads from and writes to OneLake. You don't move data between engines; every engine uses the same copy.
| Capability | Description |
|---|---|
| Format | Delta Parquet, open format - no proprietary lock-in |
| Provisioning | Automatic with every Fabric tenant |
| Shortcuts | Reference data in Azure Data Lake, AWS S3, GCP Storage without copying |
| MCP server | AI agents access OneLake directly via MCP - browse workspaces, read files, query schemas |
| Foundry IQ connection | Connect Foundry IQ to OneLake, index files, expose as knowledge inside Foundry agents |
# AI agent accessing OneLake via MCP
agent = client.agents.create(
model="gpt-5",
name="data-agent",
tools=[
{"type": "mcp", "server_url": "https://my-fabric.fabric.microsoft.com/mcp"}
# Agent can browse workspaces, read Delta files, discover table schemas
]
)
OneLake for Microsoft Foundry | OneLake MCP
Eventhouse
High-performance time-series and streaming data store. The storage layer for Real-Time Intelligence.
| Capability | Description |
|---|---|
| KQL database | Kusto Query Language - built for time-series, logs, telemetry |
| Streaming ingestion | EventHub, Kafka, IoT Hub, custom event streams |
| Cross-table analytics | Joins, aggregations, anomaly detection across multiple streams |
| OneLake integration | Data mirrored to OneLake - available to all Fabric workloads without copying |
| MCP server | AI agents query Eventhouse live via KQL through the RTI MCP server |
| Skills | Reusable prebuilt KQL capabilities for AI tools |
Eventhouse is the right store when you need sub-second query latency over streaming or time-series data. It's not a replacement for the Lakehouse - it's for real-time data that arrives continuously.
Operations Agents
AI agents that monitor real-time data in Eventhouse and act autonomously. Replace dashboard-watching with continuous automated monitoring.
How it works:
- Define what to monitor (metric threshold, anomaly condition, pattern)
- Operations Agent queries Eventhouse continuously
- When condition is met - notify, trigger Power Automate, create ticket, call API
- No user trigger required
Example: IoT sensor anomaly in Eventhouse - Operations Agent detects it - triggers Power Automate - creates ServiceNow incident - notifies engineer in Teams.
Fabric Data Agent
Natural language query across all Fabric data - lakehouses, warehouses, Eventhouses, semantic models - through a single conversational interface.
Key differentiator: auto-selects the query language. Ask a question; the Data Agent decides whether to use SQL (warehouse), KQL (Eventhouse), or DAX (semantic model) based on what the data is.
User: "What were sales last quarter, and what's the current pipeline status?"
Data Agent:
- Routes "sales last quarter" → SQL query against Data Warehouse
- Routes "current pipeline status" → KQL query against Eventhouse
- Combines results → grounded answer
| Capability | Description |
|---|---|
| Multi-source | Query lakehouse, warehouse, Eventhouse, and semantic models in one conversation |
| Context-aware language selection | Auto-selects SQL, KQL, or DAX based on data type |
| Historical + real-time | Combines batch analytics with live streaming data in one response |
| Grounded answers | Responses backed by actual query results |
| Fabric IQ integration | Data Agent wired into Foundry and Copilot Studio via Fabric IQ |
Within multi-agent architectures, the Data Agent serves as the conversational analytics component - other agents delegate data questions to it.
Ontology
A shared, governed business model that gives consistent meaning to entities, metrics, and KPIs across all agents and tools in Fabric.
Without Ontology: one agent defines "churn" as 90-day no-purchase, another defines it as subscription cancellation. Both pull different numbers. Reports and agents disagree.
With Ontology: "churn" is defined once, governed centrally, consumed by the Data Agent, Fabric IQ, Copilot Studio agents, and Power BI identically.
| Capability | Description |
|---|---|
| Shared business model | Define entities, relationships, and metrics once |
| Agent context layer | Fabric agents and RTI components consume Ontology for consistent reasoning |
| Automated actions | Ontology-defined conditions trigger business processes via alerts and workflows |
| Governance | Permissions, versioning, audit trail built in |
| Fabric IQ | Ontology is the business context layer behind Fabric IQ |
Fabric IQ
The intelligence layer that connects Fabric data to the broader Microsoft AI ecosystem.
| Component | What it connects |
|---|---|
| Fabric IQ in Foundry Toolboxes | Foundry agents access Fabric data through one MCP endpoint |
| Fabric IQ as Foundry IQ source | Fabric semantic models as knowledge bases in Foundry |
| Fabric IQ in Copilot Studio | Business user agents grounded in Fabric data |
| Ontology-grounded | Reasoning uses the shared business model |
Fabric IQ is what makes Fabric data visible to agents that live outside Fabric. Without it, every integration is custom.
Fabric AI Services Summary
| Service | Status | Use for |
|---|---|---|
| OneLake | GA | Unified storage, all Fabric workloads |
| OneLake MCP server | GA | Direct agent access to OneLake data |
| Eventhouse | GA | Time-series, streaming, KQL analytics |
| Real-Time Intelligence | GA | Full streaming pipeline - ingest to act |
| Operations Agents | GA | Autonomous monitoring and response |
| Fabric Data Agent | Preview | NL query over all Fabric data |
| Ontology | Preview | Shared business context layer |
| Fabric IQ | Preview | Connect Fabric data to Foundry and Copilot Studio |
| Power BI Copilot | GA | NL report creation, Q&A, DAX generation |
| Notebook AI assist | GA | Copilot in Spark/Python notebooks |
When Fabric vs Foundry
| Scenario | Use |
|---|---|
| AI query over existing Fabric data | Fabric Data Agent |
| Real-time streaming + AI | Eventhouse + Operations Agents |
| Consistent business context across agents | Ontology |
| Connect Fabric data to Foundry agents | Fabric IQ - Foundry IQ integration |
| Custom RAG over documents and files | Foundry IQ directly |
| Pro-code agent development | Microsoft Foundry + MAF |
| Business user agent building | Copilot Studio + Fabric IQ as knowledge source |
Microsoft AI Platform compares all four ways to build on Microsoft side by side, if the choice is wider than this pair.
Getting Started
| Resource | Link |
|---|---|
| Fabric portal | fabric.microsoft.com |
| Fabric Data Agent | docs |
| Ontology | docs |
| Eventhouse | docs |
| AI agents for RTI | docs |
| Operations Agents | docs |
| OneLake MCP | docs |
| Fabric pricing | azure.microsoft.com/pricing/details/microsoft-fabric |
Key Takeaways
- OneLake is the only copy of the data. Every Fabric workload reads from it directly - no ETL between engines. This makes Fabric significantly cheaper to operate than multi-platform stacks at scale.
- Eventhouse is not a Lakehouse replacement. Use it for real-time and time-series data that arrives continuously. Use the Lakehouse for historical batch data.
- Fabric Data Agent handles the query language decision - SQL, KQL, or DAX - based on context. This is the key differentiator over building your own text-to-SQL pipeline.
- Ontology solves the metric consistency problem that kills multi-agent accuracy. Define "churn" once; every agent and report uses the same definition.
- Fabric IQ is the integration point between Fabric and the rest of Microsoft AI. Without it, every Foundry or Copilot Studio agent that needs Fabric data requires a custom connector.
- Operations Agents replace dashboard-watching. The ROI case is straightforward: continuous monitoring at near-zero marginal cost vs. human monitoring shifts.
Next: AI Search - Fabric holds the data; AI Search is how agents retrieve from it at query time.