If you haven’t already, check out Arun Ulag’s hero blog “FabCon and SQLCon 2026: Unifying databases and Fabric on a single, complete platform” for a complete look at all of our FabCon and SQLCon announcements across both Fabric and our database offerings.
As enterprises’ use of large language models (LLMs) evolves from generating text to driving decisions, the path to the answer has come to matter as much as the answer itself. Enterprises are moving beyond AI that merely generates responses toward systems that reason, justify, and allow for inspection. This shift drives the need for AI that can ground explanations in each enterprise’s complex information environment, which requires the system to perform explicit reasoning over organizational, transactional, and behavioral relationships.
That’s why we’re introducing a graph-powered AI reasoning harness in preview through Fabric data agent and graph integration, which translates natural language intent into Graph Query Language (NL2GQL) and performs graph-based retrieval augmented generation (often called “graph RAG”) via deterministic traversals.
What is neurosymbolic AI? Why does it matter for enterprises?
Neurosymbolic AI combines neural models (such as LLMs) with symbolic representations (entities, relationships, and rules) to interpret natural language and perform step-by-step reasoning over structured knowledge – so the path to an answer can be validated, not just the output.
Text-based neural AI models are strong at summarization and generation, but they often fall short when enterprises need relationship-aware reasoning (e.g., tracing connections across linked records) and explanations that can be inspected and trusted.
Symbolic AI meets these needs by reasoning over entities, relationships, and rules – well-suited to operations management, risk and fraud analysis, compliance and investigations, and other situations where understanding what connects to what (and why) is as important as the answer. Without symbolic grounding, enterprise AI agents struggle to reliably explain, justify, or govern their decisions.
Symbolic AI reasoning via graphs
Graphs are a practical symbolic representation because they make relationships first-class. Organizations can model enterprise data and processes as nodes and relationships, which enables fast, explicit traversal of multi-hop connections without flattening them into complex joins.
In Microsoft Fabric, you can harness the power of graph-based reasoning by adding graph models or querysets as knowledge sources in data agent. Under the hood, an NL2GQL service interprets your intent, validates it against the graph model, generates a GQL statement, and executes deterministic traversals. The data agent then responds with answers grounded in the nodes, edges, and paths returned by execution – so users can inspect the “why,” not just read generated text.
While setting up a data agent, you can optionally add other supported data sources, agent instructions, data source instructions, and sample queries. These additional inputs can expand the agent’s knowledge base or further define its behaviors.
An example of graph-powered AI reasoning – cross-selling recommendations
Adventure Works is a multinational outdoor sportswear and equipment retailer. The Revenue Operations team wants to scale cross-selling opportunities leveraging “you might also like” product recommendations based on customers’ similar purchase patterns in terms of product categories and subcategories.
With graph-powered agents, Adventure Works can quickly surface recommendations by mapping connections across salespeople, customers, products, and categories – without relying on brittle similarity pipelines.
Adventure Works’ revenue operations
The company records its revenue operations using the following tables and columns inside those tables:
| Orders
• SalesOrderID • OrderDate • DueDate • ShipDate • EmployeeID • CustomerID • TotalDue • ProductID • OrderQty • UnitPrice • UnitPriceDiscount • LineTotal | Customers
• CustomerID • FirstName • LastName | Product Subcategories
• SubCategoryID • CategoryID • SubCategoryName |
| Salespeople
• EmployeeID • FirstName • LastName • JobTitle • OrganizationLevel • Country | Products
• ProductID • ProductNumber • ProductName • ModelName • StandardCost • ListPrice • SubCategoryID | |
| Product Categories
• CategoryID • CategoryName |
The challenge with SQL
While Adventure Works could solely rely on SQL for this work, this quickly pushes the team into
- Deriving a customer’s “shopping context” set for each order line (PurchaseCountry, CategoryID, SubCategoryID, ProductID), which is complicated by the country column being in the Salespeople table.
- Comparing customers to customers (often a self-join) to compute similarity based on overlap in countries they buy in and categories/subcategories they buy from.
- Aggregating products purchased by similar customers, excluding what the target customer already bought.
To do this for every customer, Adventure Works would be in “pairwise comparison” territory (customer-customer similarity), which tends to get expensive and complex without precomputation, materialized views, and extra derived tables (customer×country, customer×subcategory, co-purchase matrices).
Graph-powered agent to the rescue
Adventure Works has adopted Fabric Graph, a Microsoft-managed, horizontally scalable, native graph data system. Thus, the company didn’t need to allocate significant engineering resources to building custom similarity pipelines or maintaining derived tables.
With Fabric Graph’s no-code user interfaces, a business analyst was able to create a graph model from Adventure Works’ operational data in OneLake.
Figure: Graph model for Adventure Works' sales data. The "country" node type has been created from the corresponding column in the "salespeople" table.
Then the analyst connected the graph to Fabric Data Agent and asked the following question:
Which products should be recommended together because they are frequently bought by similar customers who make purchases in the same countries and buy from the same categories or subcategories?
A few seconds later, a recommendation for Adventure Works’s cross-selling playbook appears:
Figure: Cross-selling recommendations based on the sales data.
To provide these recommendations, the NL2GQL service:
- Found customers who placed multiple orders.
- Traversed the graph model from orders to products, subcategories, and categories.
- Constrained similarity by shared country and category context.
- Counted product pairs frequently purchased together.
- Ranked the strongest co-purchase signals per country.
This workflow is reflected in the inspectable GQL query statement:
Figure: GQL query trace allows for greater explainability.
This query shows that the agent’s answer is grounded in explicit graph traversal, not probabilistic text generation. While parts of this analysis can be performed via SQL, evolving the question quickly leads to brittle, unreadable queries. Graph-powered agents keep the reasoning model stable as intent evolves.
Conclusion
As AI systems move from answering questions to facilitating actions, enterprises need reasoning they can inspect, govern, and trust. Graph-powered AI reasoning provides a durable foundation for agents that must operate over the complexities of real-world relationships.
Learn more
- How to create a graph model in Fabric
- How to create a data agent in Fabric