Enterprise Data for AI Agents: From Legacy to Semantic

Bridging Deterministic and Probabilistic Data Worlds

Fabiane Nardon's presentation provides a compelling overview of the challenges and solutions for architecting enterprise data layers to support AI agents. A key insight is the fundamental shift required from data optimized for human-driven applications to data that can effectively serve token-sensitive, latency-aware AI reasoning loops. The proposed hybrid approach, leveraging both transactional systems for real-time writes and business logic, and data platforms for historical analysis, semantic search, and enrichment, offers a pragmatic path forward. The emphasis on data mesh principles, particularly the concept of data products with defined ownership, interfaces, and SLAs, is crucial for managing the complexity and ensuring governance in an agent-driven data ecosystem. This aligns well with modern data architecture trends and offers a structured way to integrate AI capabilities without sacrificing control.

However, a significant concern lies in the practical implementation of semantic models and ontologies. While the presentation correctly identifies the power of Semantic Web technologies (RDF, OWL) and their synergy with LLMs trained on vast datasets, the effort involved in creating and maintaining these ontologies at scale remains a substantial hurdle. The presenter acknowledges this by mentioning the difficulty of ontology creation, even with LLMs assisting. For many organizations, especially those with complex, bespoke enterprise systems, developing comprehensive and accurate ontologies that capture the nuanced meaning of their data will be a long and resource-intensive process. Furthermore, the 'token-hungry' nature of LLMs, while addressed by optimized data retrieval, still poses a cost challenge. The presentation touches upon cost as a factor in balancing deterministic and non-deterministic models, but a deeper dive into cost optimization strategies beyond reducing token overhead, such as efficient vector embeddings or smaller, specialized models, could have been beneficial. The discussion on precision also highlights the inherent trade-offs between deterministic transactional systems and probabilistic AI, a point well-made but one that necessitates careful management of user expectations and error handling.

Key Points

• Enterprise data, traditionally optimized for applications and analysts, is ill-suited for the demands of AI agents (token-hungry, latency-sensitive, unpredictable queries).
• A hybrid data access strategy is recommended: transactional systems for writes, fresh data, and embedded business logic; data platforms for stale data, historical processing, semantic search, and data enrichment.
• Data mesh architecture, with its concept of data products (governed, discoverable, well-documented data units with owners and SLAs), provides a strong foundation for organizing and serving data to AI agents.
• MCP tools should be designed as part of data products to inherit governance and ensure business-specific data retrieval.
• Semantic Web technologies (RDF, OWL) are crucial for resolving semantic ambiguities inherent in enterprise data, enabling AI agents to understand context more accurately by leveraging ontologies.
• LLMs, trained on vast datasets, can effectively interpret Semantic Web technologies, making them a viable solution for bridging the semantic gap in enterprise data.

---

📖 Source: Presentation: Architecting the Data Layer for AI Agents: From Transactional Systems to MCP and Semantic Models