Description

Fabric data agent has been available for a while, yet many implementations fall short of expectations. Without the right design, the agent may struggle to align with business context, produce inconsistent outputs, or require heavy manual intervention.

In this session, learn how to design smarter, more reliable agents by fine-tuning Agent Instructions, using the Data Agent SDK and training data sources.

Key Takeaways

• A virtual analyst that allows users to
• Lakehouse Warehouse Semantic
• Eventhouse Mirrored DB
• The conversational Data Agent allows users to interact
• Eventhouse KQL databases.
• Specify which data sources to route for
• Provide clarity on how to use

My Notes

Action Items

• [ ]

Resources & Links

Slides

Mastering Fabric Data Agents:
From Setup to Success
Wednesday March 18th
Personas
Modern Technology
Customer
Financial Advisors
Co-pilot / Chatbot
Data Management
Platform
Platforms
Portfolio Managers
…
Predictions
Fabric
Digital Experience
CRM & ERP
Document
Management
…
Marc Lelijveld
Technical Evangelist | Solution Architect
Macaw Netherlands
linkedin.com/in/MarcLelijveld
Data-Marc.com
Exploring solution areas
What are the alternatives to copilot?
What can you conclude from
the orders placed versus the
orders shipped?
Q&A
Microsoft Foundry
Copilot Studio
External tools
Copilot
Data Agents
External AI ecosystem
Q&A Visual
Natural Language Capabilities
Verified Answers

Synonyms
AI Instructions
Descriptions
A Linguistic Schema
Warehouse
Eventhouse
Semantic Model
Lakehouse
OneLake
Fabric Data Agents
Consumer
Fabric
Foundry
Copilot
Studio
Teams
MCP
Server
M365
Endpoint
A virtual analyst that allows users to
interact with and gain insights from
enterprise data in OneLake
Consume your data agents in
M365 Copilot and as remote MCP
Server in VS Code
New
Data Agent
Data agents now support
unstructured data through Azure AI
Search
New
OneLake
Data agents now support
Fabric IQ Workload through
Ontology
New
Lakehouse Warehouse Semantic
Model
Eventhouse Mirrored DB
SQL DB
AI Search
Fabric IQ
Consumer
Introducing Fabric Data Agents
Fabric
Copilot
Studio
The conversational Data Agent allows users to interact
with data naturally, enhancing accessibility and usability.
Seamlessly reason over multiple data sources, including
Lakehouses, warehouses, semantic models and
Eventhouse KQL databases.
Azure AI
Foundry
REST
Publish/ Share Data Agent
Creator
Improved chat canvas for creators with new debugging
capabilities, making it easier to understand and refine
responses.
Power BI
Copilot
Teams
Tools
Context and Configurations
OneLake
Your Data Agent can be consumed inside and outside of
Fabric. Stay tuned for upcoming integrations with
Copilot Studio, Teams, Azure AI Foundry and your own
custom applications.
Lakehouse
Warehouse
Semantic
Model
Eventhouse
KQL DB
Data Agent building blocks
• Easy configurable
• Exposable
• Multi-source
• Developer-friendly
Creator configurations
Instructions (schema)
Agent Instructions
Intended Use Cases:
Verified answers
Data Source Instructions
Intended Use Cases:
Example Queries
Intended Use Cases:
• Specify which data sources to route for
specific questions
• Provide clarity on how to use
each source
• Reflect advanced logic
• Guide how the agent interprets user intent
• Guide table selection for topics
• Examples should highlight join patterns or
expected values
• Ensure correct query logic
What to include:
• Map terminology to user intent
• Define steps for handling key questions
• Set response style and tone
What to include:
What to include:
• Descriptions of key tables
& columns
•
Representative set of questions a consumer
may ask
• Recommended table usage for certain topics
• Example column values and
filter patterns
Different types of supported sources
Lakehouse
Warehouse
Eventhouse
Semantic Model
And related sources (like mirrored sources)
Data Agent orchestrator
Lakehouse
Table
selection
Agent instructions
NL2SQL
Warehouse
Source
instructions
Data
Agent
Source selection
Example
queries
Eventhouse
NL2KQL
Passthrough
NL2DAX
Open AI query
rephrasing
Semantic Model
Prep data for AI
The Data Agent flow
Tell me more about
customer 3243
Account details for
customer 3243
Detailed
Customer Profile
Data Agent
Orchestrator
The data agent delegates
questions to specific tools or
data sources.
Data Sources are specialized
at taking a natural language
question and generating a
response – query or result.
Data Sources & Tools supply
their response back to the
orchestrator – allowing it to
process the information and
think about the next step or
return the summarized result.
Find all the survey comments
from customer id: “3243”
How many purchases did
user 3243 make?
Lakehouse
CustomerDetails
Eventhouse
KQL DB
SELECT *
FROM Profiles
WHERE CID=3243
GameBehaviors
| where Behavior == “play“
| summarize hrs =
sum(HoursPlayed)
| where CID=3243
AI Search Tool
“Loved the game! Found the
update times to be quite slow
though”
Semantic models: Prep data for AI
Three main pillars
• Simplify schema
• Verified answers
(links to visuals)
• AI Instructions
Semantic models: Prep data for AI – Simplify schema
Only keep what is relevant!
Hide report-specific measures
•
•
•
•
•
Conditional formatting
Measures outputting texts like visual titles
Disconnected tables likely cause issues
→ build relationships!
Keep in mind that Visual Calculations are not part of
the semantic model!
Semantic models: Prep data for AI – Verified answers
Only keep what is relevant!
Hide report-specific measures
•
•
•
•
•
Conditional formatting
Measures outputting texts like visual titles
Disconnected tables likely cause issues
→ build relationships!
Keep in mind that Visual Calculations are not part of
the semantic model!
Semantic models: Prep data for AI – Verified answers
Semantic models: Prep data for AI – AI instructions
Unstructured guidance (textual)
No guarantee the guidance is followed exactly –
see it as a hint!
Response formatting
•
•
•
•
•
Style
Tone
Business specific logic – for example:
•
•
•
Define what is a top performer – high or low?
Peak / off-season: which months are these?
Keep in mind!
The answers will only be as good
as your data!
Image from Urban Dictionary: https://x.com/urbandictionary/status/958044898212630528
Define success
Reliable answers
Secured
Well adopted
SDK
• Programmatic Management: Create, update, and delete Data Agent artifacts seamlessly.
• Data Source Integration: Easily connect to and integrate multiple data sources, for enhanced data
analysis and insight generation.
• OpenAI Assistants API Support: Use the OpenAI Assistants API for rapid prototyping and
experimentation.
• Workflow Automation: Automate routine tasks to reduce manual efforts and improve operational
efficiency.
• Resource Optimization: Optimize configuration and management of Data Agent resources to better
align with specific customer needs.
Diagnostics
“AI is wrong” to structured root cause analysis!
Where to start your investigation when the answer is not what you’re looking for?
Permissions
Notes for model
Prompt & grounding
Generated query analysis
• RLS / CLS
• Item level access
• governance policies
• Clear question
• Proper metadata
Source grounding
• Semantic model, lakehouse, warehouse, eventhouse
• Relationships and item naming
• Tone & voice
• Context (what is a good performer, high or low?)
• Hints (example queries)
• Diagnostics
Capacity & performance
• Storage mode (for semantic models)
• Model size
• Load on the capacity
Diagnostics
“AI is wrong” to structured root cause analysis!
Where to start your investigation when the answer is not what you’re looking for?
User prompts
Data source schemas
and configurations
File metadata and
contents
Environment and
artifact identifiers
(GUIDS)
Any proprietary
information you may
have entered
Diagnostics
Source
instructions
Schema
information
…and much more!
Permissions
Data access
Data source permission requirements
Semantic model
User
Agent
Source
Build permissions (for now! Read permissions only soon)
Agent access
Permission
Description
View
Config
Pass-through mechanics
•
User credentials flow to the data source
Default (None)
N/A
•
Permission evaluated per-user, per-query
•
RLS/OLS etc. are baked-in
View details
View the configuration
and settings, but make
no changes.
Edit and view
details
Edit the configuration
and settings.
Service principals (SPN) are NOT supported,
Access requires a user identity.
Edit Agent
Golden record testing
• Establishes trust by validating your Data Agent
against a known, authoritative dataset.
• Results are compared one-to-one
• Enables objective evaluation
Gains confidence in the data agent, produces
accurate, explainable and repeatable answers
Demo
Agent Architecture
One to rule them all?
Microsoft Foundry
Sales
Copilot Studio
Customer
Satisfaction
{…}
OneLake
M365 Copilot
Microsoft Foundry
Copilot Studio
Part of a larger ecosystem
Data
Agents
Data Agent API
(Open AI standard)
MCP
Fabric Data Agent External (Python) Client
Wrap-up
• Fabric Data Agents enable natural, governed access to data across OneLake, semantic models,
lakehouses, warehouses, and eventhouses.
• Strong foundations matter most, with well-prepared schemas, verified answers, and clear AI
instructions driving accuracy and trust.
• Security and diagnostics are essential, ensuring correct permissions, explainable behavior, and
structured root-cause analysis when results differ.
• Data Agents fit into a broader ecosystem, integrating with Fabric, Power BI, Copilot Studio,
Microsoft Foundry, and custom applications.
Key takeaway:
Reliable AI answers come from reliable data, thoughtful design, and continuous validation.
How was
the session?
Complete Session Surveys in
for your chance to WIN
PRIZES!