Home > Docs > Features > AI Copilot Configuration

🤖 AI Copilot Configuration & Setup¶

Last Updated: 2026-03-12 | Version: 1.0.0

📑 Table of Contents¶

• 🎯 Overview
• ⚙️ Enabling Copilot
• 📊 Copilot per Workload
• 📐 Configuration Best Practices
• 🔐 Security and Compliance
• 🏢 Enterprise Rollout
• ⚠️ Limitations and Known Issues
• 📚 References

🎯 Overview¶

Copilot in Microsoft Fabric provides AI-powered assistance across every major workload in the platform. It leverages Azure OpenAI models to help users generate code, build reports, create pipelines, write queries, and explain data -- all within the context of their specific workspace and data model.

Copilot Across Fabric Workloads¶

flowchart TB
    subgraph Copilot["🤖 Fabric Copilot (Azure OpenAI)"]
        ENGINE["AI Engine"]
    end

    subgraph Workloads["📦 Fabric Workloads"]
        PBI["📊 Power BI<br/>Report building, DAX, narratives"]
        DF["🔄 Data Factory<br/>Pipeline creation, expressions"]
        NB["📓 Notebooks<br/>Code gen, explain, debug"]
        SQL["🗃️ SQL<br/>Query gen, optimization"]
        KQL["🔍 KQL<br/>Query assistance"]
        DW["🏢 Warehouse<br/>T-SQL generation"]
    end

    subgraph Context["📋 Context Sources"]
        SM["Semantic Models"]
        SCH["Table Schemas"]
        META["Column Descriptions"]
        REL["Relationships"]
        CODE["Existing Code"]
    end

    Copilot <--> Workloads
    Context --> Copilot

    style Copilot fill:#6C3483,stroke:#4A235A,color:#fff
    style Workloads fill:#2E86C1,stroke:#1A5276,color:#fff
    style Context fill:#27AE60,stroke:#1E8449,color:#fff

Key Principles¶

⚙️ Enabling Copilot¶

Copilot must be enabled at multiple levels before users can access it. Follow this hierarchy from top to bottom.

Enablement Hierarchy¶

flowchart TB
    T["🏛️ Tenant Level<br/>Admin Portal"]
    C["⚡ Capacity Level<br/>Capacity Settings"]
    W["📁 Workspace Level<br/>Workspace Settings"]
    U["👤 User Level<br/>Security Groups"]

    T --> C --> W --> U

    T -->|"Must be ON"| C
    C -->|"Must be ON"| W
    W -->|"Must be ON"| U

    style T fill:#C0392B,stroke:#922B21,color:#fff
    style C fill:#E67E22,stroke:#CA6F1E,color:#fff
    style W fill:#2E86C1,stroke:#1A5276,color:#fff
    style U fill:#27AE60,stroke:#1E8449,color:#fff

Level 1: Tenant-Level Settings (Admin Portal)¶

Navigate to the Fabric Admin Portal to configure tenant-wide Copilot settings:

Admin Portal → Tenant Settings → Copilot and Azure OpenAI Service

Settings to configure:
┌────────────────────────────────────────────────────────────────────┐
│ Users can use Copilot and other features powered by Azure OpenAI  │
│ ├── Status: [Enabled / Disabled]                                  │
│ ├── Apply to: [Entire organization / Specific security groups]    │
│ └── Exclude: [Specific security groups to exclude]                │
├────────────────────────────────────────────────────────────────────┤
│ Data sent to Azure OpenAI can be processed outside your           │
│ capacity's geographic region, US, and EU data boundary             │
│ ├── Status: [Enabled / Disabled]                                  │
│ └── NOTE: Disable for FedRAMP / FISMA workloads                   │
├────────────────────────────────────────────────────────────────────┤
│ Users can use Fabric IQ for natural language queries               │
│ ├── Status: [Enabled / Disabled]                                  │
│ └── Apply to: [Entire organization / Specific security groups]    │
└────────────────────────────────────────────────────────────────────┘

⚠️ Warning: For federal agency workloads (USDA, SBA, NOAA, EPA, DOI, DOT/FAA), ensure the data residency setting is configured to keep data within the US boundary. Cross-region processing may violate FedRAMP and FISMA requirements.

Level 2: Capacity-Level Settings¶

Each Fabric capacity can independently enable or disable Copilot:

Capacity Settings → Copilot
  ├── Enable Copilot for this capacity: [On / Off]
  └── Note: Requires F2 or higher SKU (Copilot not available on F1)

Level 3: Workspace-Level Settings¶

Each workspace can control Copilot availability:

Workspace Settings → General → Copilot and Azure OpenAI
  ├── Allow Copilot and other AI features: [On / Off]
  └── Default: Inherits from capacity setting

Level 4: User-Level Access¶

Control which users can access Copilot through Entra ID security groups:

Recommended Security Group Structure:
├── sg-fabric-copilot-all          (All Copilot users)
│   ├── sg-fabric-copilot-pbi      (Power BI Copilot users)
│   ├── sg-fabric-copilot-de       (Data Engineering Copilot users)
│   ├── sg-fabric-copilot-sql      (SQL/KQL Copilot users)
│   └── sg-fabric-copilot-admin    (Full Copilot access)
├── sg-fabric-copilot-excluded     (Users excluded from Copilot)

Verification Checklist¶

📊 Copilot per Workload¶

📊 Power BI Copilot¶

Power BI Copilot is the most user-facing Copilot experience, helping report builders and consumers interact with data through natural language.

Capabilities¶

DAX Generation Examples¶

Casino Domain:

User: "Create a DAX measure for casino hold percentage"

Copilot generates:

Hold Percentage =
DIVIDE(
    SUMX(
        'SlotPerformance',
        'SlotPerformance'[CoinIn] - 'SlotPerformance'[CoinOut]
    ),
    SUM('SlotPerformance'[CoinIn]),
    0
) * 100

User: "Now create one for month-over-month change in hold percentage"

Copilot generates:

Hold Pct MoM Change =
VAR CurrentMonth = [Hold Percentage]
VAR PreviousMonth =
    CALCULATE(
        [Hold Percentage],
        DATEADD('Date'[Date], -1, MONTH)
    )
RETURN
    DIVIDE(CurrentMonth - PreviousMonth, PreviousMonth, 0) * 100

Federal Domain:

User: "Create a measure showing total USDA crop production with year-over-year comparison"

Copilot generates:

Crop Production YoY =
VAR CurrentYear = SUM('CropProduction'[ProductionValue])
VAR PreviousYear =
    CALCULATE(
        SUM('CropProduction'[ProductionValue]),
        SAMEPERIODLASTYEAR('Date'[Date])
    )
RETURN
    IF(
        NOT ISBLANK(PreviousYear),
        DIVIDE(CurrentYear - PreviousYear, PreviousYear, 0) * 100,
        BLANK()
    )

Narrative Visual Configuration¶

Add Narrative Visual to Report:
  1. Insert → Narrative Smart Visual
  2. Configure:
     ├── Summary type: [Auto / Custom prompt]
     ├── Custom prompt: "Summarize key slot performance trends,
     │    highlighting any machines below 5% hold percentage
     │    and comparing to last week"
     ├── Tone: [Professional / Casual / Executive]
     ├── Length: [Short / Medium / Detailed]
     └── Refresh: [With report / On demand]

🔄 Data Factory Copilot¶

Data Factory Copilot helps data engineers build and debug pipelines and dataflows.

Capabilities¶

Example: Pipeline Creation¶

User: "Create a pipeline to ingest USDA crop production data daily"

Copilot scaffolds:
  Pipeline: pl-usda-crop-production-daily
  ├── Web Activity: Call USDA NASS QuickStats API
  │   ├── URL: https://quickstats.nass.usda.gov/api/api_GET/
  │   ├── Method: GET
  │   └── Parameters: source_desc=SURVEY&sector_desc=CROPS&...
  ├── Copy Activity: API Response → Bronze Lakehouse
  │   ├── Source: Web Activity output
  │   ├── Sink: lh_bronze.bronze_usda_crop_production
  │   └── Format: Delta
  ├── Notebook Activity: Run silver transformation
  │   └── Notebook: nb-silver-usda-crop-production
  └── On Failure: Send email notification

📓 Notebook Copilot¶

Notebook Copilot assists data engineers and scientists with code generation, explanation, and debugging within Fabric notebooks.

Capabilities¶

Example: PySpark Code Generation¶

User: "Read the bronze USDA crop production table, filter for corn in the
       top 5 producing states, calculate year-over-year production change,
       and write to silver layer"

Copilot generates:

from pyspark.sql import functions as F
from pyspark.sql.window import Window

# Read bronze table
df_bronze = spark.read.format("delta").load(
    "Tables/bronze_usda_crop_production"
)

# Top 5 corn-producing states (by most recent year)
top_states = (
    df_bronze
    .filter(F.col("commodity_desc") == "CORN")
    .filter(F.col("statisticcat_desc") == "PRODUCTION")
    .groupBy("state_name")
    .agg(F.sum("value").alias("total_production"))
    .orderBy(F.desc("total_production"))
    .limit(5)
    .select("state_name")
    .rdd.flatMap(lambda x: x)
    .collect()
)

# Filter and calculate YoY change
window_spec = Window.partitionBy("state_name").orderBy("year")

df_silver = (
    df_bronze
    .filter(F.col("commodity_desc") == "CORN")
    .filter(F.col("statisticcat_desc") == "PRODUCTION")
    .filter(F.col("state_name").isin(top_states))
    .withColumn("prev_year_production",
                F.lag("value").over(window_spec))
    .withColumn("yoy_change_pct",
                F.when(F.col("prev_year_production").isNotNull(),
                       ((F.col("value") - F.col("prev_year_production"))
                        / F.col("prev_year_production") * 100))
                .otherwise(None))
    .withColumn("processed_timestamp", F.current_timestamp())
)

# Write to silver layer
df_silver.write.format("delta").mode("overwrite").save(
    "Tables/silver_usda_corn_production_yoy"
)

print(f"Silver table written: {df_silver.count()} rows for states: {top_states}")

🗃️ SQL Copilot¶

SQL Copilot assists with T-SQL query generation and optimization in Fabric Warehouse and SQL analytics endpoints.

Capabilities¶

Example: Cross-Domain Query¶

User: "Write a query that joins casino player data with SBA loan data
       to find players who are also small business owners in the same state"

Copilot generates:

-- Cross-domain analysis: Casino players who are SBA loan recipients
-- Note: This query requires appropriate cross-workspace access permissions
SELECT
    p.player_id,
    p.player_name,
    p.state AS player_state,
    p.loyalty_tier,
    p.lifetime_value,
    s.loan_number,
    s.business_name,
    s.loan_amount,
    s.approval_date,
    s.jobs_reported
FROM gold_player_value p
INNER JOIN silver_sba_7a_loans s
    ON p.state = s.borrower_state
    AND p.player_name = s.borrower_name  -- Approximate match
WHERE p.loyalty_tier IN ('Gold', 'Platinum')
ORDER BY p.lifetime_value DESC;

⚠️ Warning: Cross-domain joins like the example above require careful data governance review. Ensure that joining PII across domains complies with your organization's data use policies and applicable regulations.

🔍 KQL Copilot¶

KQL Copilot assists with writing and optimizing Kusto Query Language queries in Eventhouse.

Capabilities¶

Example: Anomaly Detection Query¶

User: "Write a KQL query to detect anomalous spikes in slot machine
       error rates over the past 7 days, grouped by floor location"

Copilot generates:

// Detect anomalous error rate spikes by floor location
let lookback = 7d;
let bin_size = 1h;
let sensitivity = 1.5;  // Lower = more sensitive
SlotTelemetryRaw
| where Timestamp > ago(lookback)
| summarize
    TotalEvents = count(),
    ErrorCount = countif(EventType == "error")
    by FloorLocation, bin(Timestamp, bin_size)
| extend ErrorRate = round(todouble(ErrorCount) / TotalEvents * 100, 2)
| order by FloorLocation asc, Timestamp asc
| make-series ErrorRate = avg(ErrorRate)
    on Timestamp
    step bin_size
    by FloorLocation
| extend (anomalies, score, baseline) =
    series_decompose_anomalies(ErrorRate, sensitivity)
| mv-expand
    Timestamp to typeof(datetime),
    ErrorRate to typeof(double),
    anomalies to typeof(int),
    score to typeof(double),
    baseline to typeof(double)
| where anomalies == 1  // Positive anomalies (spikes)
| project FloorLocation, Timestamp, ErrorRate, baseline, score
| order by score desc

📐 Configuration Best Practices¶

Semantic Model Preparation¶

The single most impactful thing you can do to improve Copilot accuracy is to prepare your semantic model with rich metadata.

Naming Conventions That Help Copilot¶

Description Guidelines¶

Write descriptions as if explaining to a new team member who has domain knowledge but no knowledge of your data model:

# Table Description Template
table: gold_slot_performance
description: >
  Daily aggregated slot machine performance metrics. One row per machine per gaming day.
  Contains financial metrics (coin-in, coin-out, hold percentage), utilization metrics
  (hours played, session count), and maintenance flags. Source: Silver layer slot telemetry.
  Refresh: Daily at 4:00 AM ET. Grain: Machine + Date.

# Column Description Template
column: hold_pct
description: >
  Casino hold percentage - the portion of wagered money retained by the casino.
  Formula: (CoinIn - CoinOut) / CoinIn * 100.
  Normal range: 2-15%. Values outside this range may indicate machine issues.
  Business context: Higher hold = more revenue per dollar wagered.

Measure Descriptions¶

-- Include descriptions as DAX comments AND in the model metadata
-- Both help Copilot understand intent

// Measure: Total Revenue
// Description: Total net gaming revenue across all slot machines.
//   Calculated as CoinIn minus CoinOut (casino win).
//   Use for: Financial dashboards, floor performance analysis.
//   Time intelligence: Supports YoY, MoM, WoW comparisons.
//   Related measures: Hold Percentage, Average Daily Revenue
Total Revenue =
    SUMX(
        'SlotPerformance',
        'SlotPerformance'[CoinIn] - 'SlotPerformance'[CoinOut]
    )

Column-Level Documentation Checklist¶

Workspace Organization for Copilot¶

Recommended Workspace Structure:
ws-gaming-domain/
├── Lakehouses/
│   ├── lh_bronze      (Raw data - minimal Copilot interaction)
│   ├── lh_silver      (Cleansed - Copilot-ready with descriptions)
│   └── lh_gold        (Business ready - primary Copilot target)
├── Semantic Models/
│   ├── sm_gaming_ops   (Fully documented - Power BI Copilot)
│   └── sm_compliance   (RLS enforced - restricted Copilot)
├── Reports/
│   ├── rpt_floor_ops   (Copilot-enabled for consumers)
│   └── rpt_compliance  (Copilot with RLS filters)
├── Notebooks/
│   ├── nb_silver_*     (Copilot for code gen)
│   └── nb_gold_*       (Copilot for code gen)
└── Eventhouses/
    └── evh_operations  (KQL Copilot for queries)

🔐 Security and Compliance¶

Data Boundary Architecture¶

flowchart TB
    subgraph Boundary["🔒 Data Boundary"]
        subgraph User["👤 User Context"]
            ROLE["User Roles & Groups"]
            RLS["Row-Level Security"]
            CLS["Column-Level Security"]
        end

        subgraph Copilot["🤖 Copilot Processing"]
            INPUT["User Prompt"]
            CTX["Model Metadata<br/>(schema, descriptions)"]
            GEN["Generated Query/Code"]
        end

        subgraph Data["💾 Data Access"]
            QUERY["Query Execution<br/>(filtered by RLS/CLS)"]
            RESULT["Filtered Results"]
        end
    end

    ROLE --> Copilot
    RLS --> QUERY
    CLS --> QUERY
    INPUT --> CTX --> GEN --> QUERY --> RESULT

    style Boundary fill:#f9f9f9,stroke:#333,color:#000
    style User fill:#C0392B,stroke:#922B21,color:#fff
    style Copilot fill:#6C3483,stroke:#4A235A,color:#fff
    style Data fill:#2E86C1,stroke:#1A5276,color:#fff

RLS/CLS with Copilot¶

Copilot queries are executed in the security context of the requesting user. This means:

💡 Tip: Test Copilot responses under each RLS role to verify that restricted data is not leaked through AI-generated narratives or summaries.

Audit Logging¶

All Copilot interactions are captured in the Fabric unified audit log:

// Query Copilot audit events
search in (FabricAuditLogs)
    "CopilotInteraction"
| where TimeGenerated > ago(7d)
| project
    TimeGenerated,
    UserId,
    WorkspaceName,
    WorkloadName,
    Activity,
    CopilotPrompt = tostring(parse_json(AdditionalInfo).prompt),
    CopilotResponse = tostring(parse_json(AdditionalInfo).response_type),
    DataAccessed = tostring(parse_json(AdditionalInfo).tables_accessed)
| order by TimeGenerated desc

Data Residency Configuration¶

For federal compliance workloads:

Admin Portal → Tenant Settings → Copilot and Azure OpenAI Service
  → "Data sent to Azure OpenAI can be processed outside your
     capacity's geographic region, US, and EU data boundary"
  → Set to: DISABLED for federal workloads

Opting Out of Data Collection¶

Organizations can control how Copilot data is used:

Admin Portal → Tenant Settings → Copilot
  ├── Allow Microsoft to use Copilot data to improve Fabric Copilot
  │   └── Set to: DISABLED for sensitive/federal workloads
  └── Log Copilot interactions for admin review
      └── Set to: ENABLED for compliance tracking

🏢 Enterprise Rollout¶

Phased Enablement Strategy¶

flowchart LR
    subgraph Phase1["📋 Phase 1: Pilot (Weeks 1-4)"]
        P1A["Enable for IT/Data team only"]
        P1B["Test with non-sensitive data"]
        P1C["Document findings"]
    end

    subgraph Phase2["📋 Phase 2: Expand (Weeks 5-8)"]
        P2A["Enable for power users"]
        P2B["Add domain workspaces"]
        P2C["RLS/CLS validation"]
    end

    subgraph Phase3["📋 Phase 3: General (Weeks 9-12)"]
        P3A["Enable for all users"]
        P3B["Full workspace rollout"]
        P3C["Monitoring & optimization"]
    end

    Phase1 --> Phase2 --> Phase3

    style Phase1 fill:#E67E22,stroke:#CA6F1E,color:#fff
    style Phase2 fill:#2E86C1,stroke:#1A5276,color:#fff
    style Phase3 fill:#27AE60,stroke:#1E8449,color:#fff

Phase 1: Pilot (Weeks 1-4)¶

Phase 2: Expand (Weeks 5-8)¶

Phase 3: General Availability (Weeks 9-12)¶

Training Materials¶

Prompt Engineering Tips for End Users¶

Example Prompt Library¶

# Casino Domain Prompts
casino_prompts:
  - category: "Floor Performance"
    prompts:
      - "Show slot revenue by floor section for the past 24 hours"
      - "Which machines have the lowest hold percentage this week?"
      - "Compare weekend vs weekday coin-in for high-limit slots"
      - "What is the trend in player sessions over the past month?"

  - category: "Compliance"
    prompts:
      - "How many CTR filings were generated today?"
      - "Show players with transactions approaching the $10,000 threshold"
      - "List all W-2G events for the current tax year by player"

# Federal Domain Prompts
federal_prompts:
  - category: "USDA"
    prompts:
      - "What was total corn production by state in 2025?"
      - "Show the year-over-year change in soybean yield"
      - "Which crops had the largest acreage increase this year?"

  - category: "EPA"
    prompts:
      - "Top 10 facilities by toxic release volume in 2024"
      - "Show the trend of PM2.5 concentrations in California"
      - "Which chemicals had the largest year-over-year increase in releases?"

  - category: "NOAA"
    prompts:
      - "List all Category 4+ hurricanes in the last 10 years"
      - "What was the average temperature in Phoenix each month last year?"
      - "Show severe weather alerts by state for this week"

Measuring Copilot ROI¶

⚠️ Limitations and Known Issues¶

General Limitations¶

Workload-Specific Limitations¶

Known Issues (as of March 2026)¶

📚 References¶

🔗 Related Documents¶

• Fabric IQ -- Natural language analytics powered by Copilot
• Real-Time Intelligence -- KQL Copilot in Eventhouse
• Data Mesh Enterprise Patterns -- Copilot governance across domains
• Security -- Security and compliance framework
• Architecture -- System architecture overview

📝 Document Metadata - Author: Documentation Team - Reviewers: Fabric Admin, Security, Data Engineering, BI Team - Classification: Internal - Next Review: 2026-06-12