Microsoft Promotes 'Grounding' for Enterprise AI
Microsoft is positioning its "grounding" technology as the "invisible infrastructure" for enterprise AI. The company emphasizes that persistent context, workflow orchestration, and real-world integration are key differentiators for building trust and reliability in agentic automation systems.
- The "invisible infrastructure" for grounding is materialized in services like Foundry IQ, a knowledge layer within Azure AI Foundry powered by Azure AI Search that provides "agentic retrieval" for AI agents, abstracting away the need to build custom RAG pipelines. - Microsoft is advancing beyond standard RAG with techniques like GraphRAG, a method from Microsoft Research that constructs a knowledge graph from text to enable more complex reasoning and answer global, dataset-wide questions. The GraphRAG library is also available on GitHub. - For agentic systems that interact with user interfaces, Microsoft has developed Phi-Ground, a model family focused on GUI grounding that translates natural language instructions into precise screen coordinates for actions like clicks and inputs. This technology has been integrated into features like Vision Highlighting in Windows Copilot. - Microsoft's concept of "agentic retrieval" evolves traditional RAG by using an LLM to decompose complex questions into multiple sub-queries that can be executed in parallel across different knowledge sources before synthesizing a grounded response. - The Azure AI Foundry Agent Service supports native integration with third-party systems like Elasticsearch, allowing agents to be grounded in external, proprietary data through a dedicated Foundry Connection and customizable search templates. - Microsoft's research in this area extends to multimodal grounding, with models like KOSMOS-2 designed to connect text to specific objects and their locations within images, thereby reducing object-related hallucinations. - From a security perspective, enterprise grounding within Microsoft 365 Copilot leverages the Microsoft Graph to ensure that AI responses only surface data and content that the individual user already has explicit permissions to view. - New features in Bing Webmaster Tools provide content owners with metrics on how their data is being used for grounding AI-generated answers, including citation tracking and visibility into the specific "grounding queries" used by AI systems to retrieve their content.
