AI Dev 25 X NYC | Christoph Meyer, Lars Heling: Improving AI Agent Discovery with a Knowledge Graph

Why It Matters

By leveraging a knowledge graph to supply semantic context and process sequencing, SAP enables AI agents to reliably automate complex, cross‑system business tasks, unlocking faster, error‑free digital transformation for enterprises.

Summary

In this AI Dev 25 session, SAP Business AI leaders Christoph Meyer and Lars Heling explain how a knowledge graph can dramatically improve the discovery and execution of AI agents within SAP’s enterprise ecosystem. They introduce Joule, SAP’s AI‑driven business co‑pilot, which orchestrates processes across ERP, supply chain, HR and procurement by invoking a vast catalog of APIs. Central to Joule’s effectiveness is the SAP Knowledge Graph, which enriches raw API metadata with semantic context and business‑process relationships, turning a chaotic landscape of 3,500 services and 110,000 endpoints into a navigable, actionable map.

The presenters outline two core challenges for AI agents: (1) API discovery, where relevant services are scattered across multiple data sources and described with domain‑specific jargon, and (2) business‑process sequencing, where APIs must be called in a precise order that varies by customer configuration. To address these, they combine a vector‑based retrieval layer that embeds user utterances and API descriptions with a graph‑based augmentation that pulls in upstream and downstream process edges. This hybrid approach surfaces not only the directly requested APIs (e.g., a purchase‑order service) but also prerequisite steps such as purchase‑requisition creation, ensuring the agent has a complete toolset to fulfill the request.

Lars walks through a live demo using a toy knowledge graph of 100 APIs. He shows how a simple natural‑language request—"create a purchase order for five pencils"—is transformed into a set of candidate APIs via embedding similarity, then enriched by querying the graph for related process edges. The result is an enhanced discovery list that includes both the purchase‑order API and its required predecessor, illustrating how the graph’s ontology (entities, relationships, transitive properties) enables inference and context‑aware retrieval. Christoph’s demo further highlights the flexibility of constructing custom embedding texts from graph queries, allowing rapid adaptation to new domains without re‑engineering the underlying model.

The session underscores the strategic value of embedding a knowledge graph between data lakes and AI agents: it provides a unified semantic layer that reduces integration friction, accelerates automation, and mitigates errors caused by missing or mis‑sequenced API calls. For enterprises adopting AI‑augmented workflows, this architecture promises faster time‑to‑value, lower maintenance overhead, and a scalable path to extend AI capabilities across heterogeneous SAP landscapes.

Original Description

Christoph Meyer, Principal AI Scientist and Lars Heling Senior Knowledge Engineer at SAP SE, discussed at AI Dev 25 x NYC how to improve AI agent discovery and execution using a knowledge graph.

LLMs make agents fluent; knowledge graphs make them effective. This session showed how a knowledge graph gives AI agents the semantic and process context to discover and safely invoke the right tools and APIs across complex enterprise systems.

They covered key techniques—semantic retrieval, process-aware API connectivity, and how they align with the Model Context Protocol (MCP) and run a demo where an agent applied these methods.

-------------

Join us at AI Dev 26 x San Francisco! Tickets: https://ai-dev.deeplearning.ai