LangChain Raised $300 Million at a $3 Billion Valuation. The 7,000-Organization LangSmith Platform Is What Every Enterprise AI Agent Deployment Needs Next.
Microsoft put knowledge graphs into public preview in Foundry IQ on November 17, 2026, alongside Decision-1.1 reaching general availability. Here is what the architecture shift means for enterprise AI agents and when the complexity is actually worth it.
On November 17, 2026, Microsoft updated Azure AI Foundry with two announcements that reflect distinct phases of enterprise AI maturity. The Foundry IQ knowledge graph capability entered public preview, adding structured entity and topic extraction to the knowledge base layer that enterprise AI agents use for grounding. And Microsoft-Decision-1.1 reached general availability — a small, fast model for bounded decisions that is more than twice as fast as its predecessor at the same cost.
Taken together, the announcements describe a coherent architectural direction: Microsoft wants enterprise AI agents to query a richer, more structured representation of business knowledge, and to make faster, more reliable routing decisions at the agent workflow level. The question for enterprise technology leaders is not whether to pay attention to these updates, but whether the specific architecture they represent solves the context grounding problems their agents actually have.
The Foundry IQ Knowledge Graph: What Changed on November 17
Standard Foundry IQ knowledge bases already give enterprise AI agents access to curated institutional knowledge — policies, product documentation, operational procedures — through a retrieval pipeline that searches semantically similar content and returns relevant chunks to the agent for synthesis. The November 17 update adds a layer that standard retrieval cannot provide: structured knowledge graphs with entities and topics that enable relationship-aware queries across the enterprise document set.
This distinction matters more than it initially appears. Consider the difference between two enterprise questions.
The first question is "What is our refund policy for subscription customers?" Semantic search handles this well. The policy document exists, it contains the word "refund" and the phrase "subscription customers," and a similarity search will find and return it.
The second question is "What refund obligations do we have under contracts with enterprise customers in EMEA that were signed before our policy changed in Q2 2025?" Standard semantic search cannot answer this. It does not know what the relationships are between contracts, customers, regional classifications, and policy versions. It can find documents that contain these terms, but it cannot navigate the relationships between them to synthesize an answer.
That is the gap knowledge graphs address. The entities and topics capability extracts the facts and relationships from every document during indexing, stores them as a graph structure, and enables agents to query that structure directly — in natural language or in GQL — with source citations attached to each answer.
The Benchmark: What Microsoft's Accuracy Data Actually Shows
Microsoft has published benchmark data comparing knowledge graph queries to standard RAG, and the numbers are dramatic enough to warrant careful treatment. The company reports that entity graph search answered 93.5% of queries correctly versus 26.1% for standard RAG, while consuming 64.3% fewer tokens.
These figures come from Microsoft's internal evaluation. They have not been independently replicated. Standard caveats apply: benchmarks designed by a vendor to validate its own product will tend to emphasize query types where the new approach excels. The 93.5% vs 26.1% comparison likely reflects queries specifically designed to test relationship-aware retrieval — exactly the class of queries where knowledge graphs outperform semantic search by design.
| Retrieval Method | Accuracy (Microsoft benchmark) | Tokens vs. RAG |
|---|---|---|
| Standard RAG | 26.1% | Baseline |
| Foundry IQ Entity Graph Search | 93.5% | 36.4% of RAG baseline |
| Standard Foundry IQ knowledge base | +36% over direct search* | — |
*Microsoft's separate evaluation on the Frames benchmark reported a 36% gain in answer quality over directly searching all sources, using query planning and reflective search. These are distinct evaluations with different methodologies.
The more useful framing is directional: for queries that require navigating relationships between entities across multiple documents — contracts, compliance documents, organizational hierarchies, product catalogs — knowledge graphs address a structural limitation of chunk-based retrieval. For simple queries against a homogenous document set, the complexity cost of knowledge graph indexing likely does not justify the accuracy gain. The enterprise technology decision is whether the query distribution of a specific agent's workload tilts toward the class of queries where graphs win.
Entities and Topics: How the Architecture Works
Foundry IQ knowledge graphs have two components: Knowledge Graph Entities and Knowledge Graph Topics.
Entities extract facts and the relationships between them from enterprise documents during the indexing process. The extraction uses an ontology — either one you provide if your organization has an existing knowledge model, or one automatically generated by an agent skill from your document set. The resulting graph stores facts as nodes and relationships as edges. An entity query asks the graph about connections between nodes, returning relationship paths with source citations. Direct GQL queries are also supported for applications that need programmatic access to the graph structure.
Topics organize documents into a thematic hierarchy — a way of grouping the enterprise document set by subject domain, business process, or regulatory area. The topics hierarchy gives agents category-level context that complements the entity-level fact retrieval. A document about a supplier contract might contain entities (the supplier, the payment terms, the deliverables), and also belong to a topic (vendor management, procurement policy, regional compliance). The combination lets agents understand both specific facts and the broader context in which those facts are relevant.
An agent skill provided by Microsoft automatically generates an initial topic ontology from a new knowledge source. This reduces the manual configuration burden that has historically made knowledge graph implementations expensive — building an ontology for a large enterprise document set could previously take months of knowledge engineering work. The automatic generation is a reasonable starting point; most enterprise deployments will iterate on the generated ontology based on specific query requirements, but the starting point is significantly better than a blank schema.
The Four Layers of Microsoft IQ
Foundry IQ is one part of a broader Microsoft IQ framework announced alongside these updates. Understanding where Foundry IQ sits within that framework matters because the full architecture describes Microsoft's approach to enterprise AI context — the problem of giving agents the right information to act on the right questions.
| Microsoft IQ Layer | What It Contains | Primary Use Case |
|---|---|---|
| Work IQ | Microsoft 365, Teams, SharePoint, enterprise collaboration | People, meetings, decisions, organizational knowledge |
| Fabric IQ | Microsoft Fabric analytics, data warehouses, real-time data | Structured data, metrics, analytics |
| Foundry IQ | Curated documents, policies, knowledge bases | Authoritative enterprise knowledge |
| Web IQ | Real-time external information | Current market data, external context |
The design logic is a layered context model: agents have access to organizational knowledge (Work IQ), structured data (Fabric IQ), authoritative documents (Foundry IQ), and external context (Web IQ). Each layer is independently useful; the enterprise value is in the combination. An agent handling a customer escalation might simultaneously access the customer's history from Fabric IQ, the relevant policy documents from Foundry IQ, the account team's recent communications from Work IQ, and current market conditions from Web IQ.
This architecture directly competes with the custom retrieval pipelines that most enterprises have built by combining separate vendors: a vector database for document retrieval, a data warehouse connector for structured data, a Microsoft Graph integration for collaboration data, and a web search API for external context. Microsoft's IQ framework consolidates those retrieval layers into a single managed product with shared authentication, shared observability, and 350-plus third-party connectors for enterprise data sources that are neither Microsoft products nor standard databases.
Signal's analysis of the IBM and Microsoft Imagine Enterprise platform described the broader strategic logic: Microsoft is using its position across the enterprise stack — productivity tools, cloud infrastructure, database products, and now AI agents — to build a distribution advantage in enterprise AI that is difficult for single-product AI companies to replicate. The IQ framework is the context grounding expression of that strategy.
Decision-1.1 GA: Speed for Agent Decision Nodes
Microsoft-Decision-1.1 reaching general availability on the same day as the Foundry IQ knowledge graph update is not coincidental. Decision-1.1 is designed for the routing and classification layer of agent workflows — the decision nodes where an agent must quickly determine what to do next without the latency overhead of a large language model call.
The key characteristics:
Speed: More than 2x faster than its predecessor, Microsoft-Decision-1, at the same cost. For latency-sensitive decision nodes in production agent workflows, this matters: a routing decision that took 500ms with Decision-1 takes under 250ms with Decision-1.1.
Reliability scoring: Decision-1.1 returns a confidence score alongside its decision. Applications can use this score to implement dynamic escalation: defer a decision to a human reviewer when confidence is below a threshold, or route to a larger model for synthesis when the routing decision itself is uncertain. This is a production reliability feature, not a research capability.
Benchmark position: Microsoft reports a score of 68 on the Artificial Analysis Decision Index v2.1. As with the knowledge graph benchmarks, this is a vendor-reported figure. The more useful framing is that Decision-1.1 is positioned as a cost-efficient specialized model for decision tasks, not as a frontier reasoning model — the right comparison is task-specific efficiency, not general benchmark leadership.
Visual decision support: Decision-1.1 also supports image analysis, extending its bounded decision capability to visual inputs. For workflows that require routing or classification based on images — document processing, quality inspection, visual content moderation — this expands the range of agent decision nodes it can serve.
Together with Foundry IQ knowledge graphs, Decision-1.1 addresses both ends of the agent information problem: richer context retrieval (knowledge graphs) and faster, more reliable task routing (Decision-1.1). These are complementary capabilities in a multi-step agent workflow where the agent first queries the knowledge graph to understand the relevant enterprise context, then routes to the appropriate next action based on what it found.
Why Standard RAG Is Not Enough for Enterprise Knowledge
The architectural case for knowledge graphs over RAG comes down to the nature of enterprise knowledge itself. Enterprise documents are not independent. A supplier contract references pricing schedules. A pricing schedule references product categories. Product categories have regulatory classifications. Regulatory classifications determine what disclosure requirements apply to contracts. This chain of references is the structure that RAG cannot navigate and knowledge graphs are designed to traverse.
Microsoft's execution containers for enterprise AI agents addressed the execution-level isolation problem: how to run agents with the security boundaries that enterprise IT governance requires. The knowledge graph update addresses the information-level problem: how to give agents access to the relationships within enterprise knowledge that make them genuinely useful for complex, context-dependent tasks.
The RAG limitation is particularly acute in regulated industries. Healthcare compliance agents need to understand the relationships between patients, treatments, billing codes, and regulatory requirements. Financial services agents need to navigate the relationships between products, customers, jurisdictions, and applicable regulations. Manufacturing quality agents need to trace relationships between specifications, suppliers, components, and certifications. In each case, the enterprise knowledge has a graph structure that chunk-based retrieval approximates poorly.
The enterprise AI context gap that makes agents fail before they can deliver value is fundamentally the same problem that the knowledge graph capability is designed to fix: agents deployed without access to the structured enterprise context they need to make relationship-aware decisions produce outputs that look plausible but fail on edge cases that require navigating the connections within the enterprise knowledge base.
The Enterprise Implementation Playbook: When to Use Knowledge Graphs
Not every enterprise agent deployment needs knowledge graphs. The complexity of indexing a knowledge graph — defining or generating an ontology, mapping entities during indexing, managing the graph structure as documents are updated — adds operational overhead relative to standard RAG that is only worth bearing when the query distribution genuinely requires relationship-aware retrieval.
The following decision framework helps enterprise AI teams determine when knowledge graphs justify the implementation cost:
1. Audit your agent's current failure modes. If your agent consistently fails on queries that require navigating relationships between entities — connecting information across multiple documents, tracking changes in policy or regulation over time, understanding the dependencies between contracts and their associated schedules — knowledge graphs are likely addressing your actual bottleneck. If your agent's failures are mostly in finding relevant documents at all, the problem is retrieval quality rather than retrieval structure, and graph indexing will not fix it.
2. Assess your domain's ontological structure. Knowledge graphs deliver the most value in domains where the enterprise knowledge has a natural graph structure: regulatory compliance, contract management, product catalog management, organizational hierarchy, and supply chain relationships. For domains with relatively flat document structures — FAQs, general policy documents, training materials — the graph overhead will add cost without commensurate accuracy gains.
3. Start with the auto-generated ontology. Microsoft's agent skill that automatically generates a topic ontology from your document set is the right starting point for most deployments. Run the graph with the auto-generated ontology, evaluate accuracy on a representative sample of the query types your agent handles, and iterate on the ontology based on where the automatic generation falls short. Building a bespoke ontology before evaluating the auto-generated one wastes engineering time on problems that may not exist.
4. Instrument the comparison. Run both a standard Foundry IQ knowledge base and a knowledge graph configuration in parallel on a representative sample of your agent's real queries before committing to full graph deployment. Microsoft's internal benchmark showing 93.5% vs 26.1% is from an evaluation designed to show the gap; your query distribution may show a smaller gap. Know your own numbers before making the architecture decision.
5. Plan for ontology maintenance. Knowledge graphs add a maintenance overhead that RAG does not: as documents are updated, added, or removed, the entity extraction must run again, and significant changes to the enterprise knowledge structure may require ontology updates. For rapidly changing domains — pricing, promotional offers, news-sensitive content — the ontology maintenance cost may exceed the accuracy benefit.
What the November 17 Updates Signal About Microsoft's AI Strategy
The combination of knowledge graphs in public preview and Decision-1.1 going GA on the same day is a signal about where Microsoft believes the enterprise AI agent market is heading. Both updates address the production reliability of enterprise AI agents — not their capability ceiling (that is what frontier model updates address) but their operational maturity.
Signal's coverage of Microsoft's MAI model family described Microsoft's first-party model strategy: purpose-built models for specific enterprise tasks rather than a single frontier model for all tasks. Decision-1.1 is an expression of that strategy at the specialized end: a model that does one thing — make bounded decisions — extremely fast and reliably, with a confidence score that enables dynamic escalation.
Foundry IQ knowledge graphs are an expression of the same strategy applied to context infrastructure: rather than giving every agent access to a flat document store and asking it to reason about relationships, structure the knowledge base in a way that makes relationship-aware queries tractable. The architectural philosophy is the same as the model philosophy: specialized components for specialized problems, composable into agent workflows that handle enterprise complexity.
For enterprises that have already deployed AI agents and are running into the ceiling of what standard retrieval can accomplish, the November 17 updates offer a concrete upgrade path that does not require rebuilding the agent. For enterprises that are planning new agent deployments in relationship-rich domains — legal, compliance, procurement, healthcare — the knowledge graph capability gives them a production-ready foundation that was not available six months ago.
The Menlo Ventures 2026 State of Enterprise AI report found that enterprise AI spending hit $177 billion and incumbents held 91% of that budget. Microsoft's IQ framework is a direct expression of why incumbents hold that position: the enterprise decision to build context infrastructure on Azure AI Foundry is a decision that compounds over time, as the knowledge graph accumulates entity data and the topic ontology is refined. Each iteration of that refinement increases the cost of migration to a competing platform. That is the stickiness mechanism Microsoft is building with every Foundry IQ update.
Takeaway: Microsoft's Foundry IQ knowledge graphs enter public preview with vendor benchmark data showing a 93.5% vs 26.1% accuracy advantage over standard RAG for relationship-aware queries. The honest application of that data is that knowledge graphs solve a specific structural problem — navigating relationships across enterprise documents — that standard RAG cannot address, and that the accuracy gap will be much smaller for agents whose queries are primarily about finding relevant documents rather than navigating between them. The Decision-1.1 GA is a less headline-grabbing but practically important update: 2x faster routing decisions with confidence scoring change the operational economics of production agent workflows in ways that compound across millions of agent calls. Together, the updates represent Microsoft's consistent strategy of building specialized, composable enterprise infrastructure rather than competing at the frontier model level it cannot win on cost alone.
Frequently Asked Questions
What is Microsoft Foundry IQ?
Microsoft Foundry IQ is the enterprise knowledge layer within Azure AI Foundry that gives AI agents shared, structured access to an organization's institutional knowledge. It is part of the broader Microsoft IQ framework, which includes Work IQ (knowledge from Microsoft 365 and Teams), Fabric IQ (data and analytics from Microsoft Fabric), Foundry IQ (curated documents, policies, and enterprise knowledge bases), and Web IQ (real-time external information). Foundry IQ serves as the authoritative knowledge layer for AI agents built on Azure AI Foundry, letting multiple agents access the same enterprise context — policies, product documentation, operational procedures, and regulatory requirements — without each agent team maintaining its own retrieval pipeline. It supports more than 350 third-party connectors and is generally available, with the knowledge graph capability (entities and topics) now in public preview.
What is the difference between Foundry IQ knowledge graphs and RAG?
Standard RAG (Retrieval-Augmented Generation) retrieves chunks of text from a document store based on semantic similarity to a query, then passes those chunks to the language model to synthesize an answer. The limitation is that RAG operates at the chunk level — it finds documents that are similar to a query but cannot directly navigate relationships across multiple documents. Foundry IQ knowledge graphs add a structured relationship layer: entities represent facts extracted from documents, and the graph stores how those entities relate to each other. When an agent queries the knowledge graph, it follows relationship chains across documents rather than relying on semantic similarity alone. Microsoft's internal benchmark found that entity graph search answered 93.5% of queries correctly versus 26.1% for standard RAG, while consuming 64.3% fewer tokens — though these figures come from Microsoft's own evaluation and have not been independently replicated.
What are entities and topics in Foundry IQ knowledge graphs?
Knowledge Graph Entities and Knowledge Graph Topics are the two components of Foundry IQ's structured knowledge capability, both in public preview as of November 2026. Entities extract specific facts and the relationships between them from enterprise documents during the indexing process. For example, indexing a contract document would extract entities for the parties, the terms, the obligations, and the dates, and store the relationships between them in a queryable graph structure. Queries can be made in natural language or directly in GQL (Graph Query Language), and each answer includes source citations. Topics group documents into a thematic hierarchy — clustering all documents related to a product line, a business process, or a regulatory domain — and an agent skill automatically generates an initial topic ontology from the document set, reducing the manual configuration work. Together, entities enable relationship-aware queries and topics enable category-aware queries that standard semantic search cannot handle.
What is Microsoft Decision-1.1?
Microsoft Decision-1.1 is a small, specialized language model designed for making bounded decisions within software systems — routing tasks, selecting options, classifying inputs — rather than generating open-ended text. It reached general availability in Azure AI Foundry on November 17, 2026. Decision-1.1 is more than twice as fast as its predecessor Microsoft-Decision-1 at the same cost, making it suitable for latency-sensitive decision nodes in agent workflows. It can return a reliability score alongside its decision, allowing applications to defer a decision to a human or escalate to a larger model when uncertainty is high. According to Microsoft, Decision-1.1 scores 68 on the Artificial Analysis Decision Index v2.1. It was developed in part from Google Gemma 4 26B A4 and also supports image analysis, extending its decision-making capability to visual inputs.
How do you implement knowledge graphs in Azure AI Foundry?
Implementing knowledge graphs in Foundry IQ requires Azure AI Foundry access and follows four main steps: First, create or connect a knowledge base in Foundry IQ and add your document sources through the portal or API. Second, add a knowledge graph configuration — either providing your own ontology if you have an existing knowledge model, or using the agent skill that automatically generates an initial topic ontology from your document set. Third, configure the graph to map entities during the indexing process using your ontology, which extracts facts and relationships and stores them in the queryable graph structure. Fourth, connect the knowledge base to your agent using the Foundry IQ APIs, which support natural language queries and direct GQL queries with source citations. Microsoft notes that the knowledge graph approach adds indexing complexity relative to standard RAG, so it is most appropriate for domains where relationship-aware queries are a significant part of the agent's workload.