SAP knowledge graphs and governance anchor enterprise AI agents for reliable automation

Enterprise AI agents that only rely on general knowledge fail to understand internal terminology and business processes. At VB Transform 2026, SAP senior solution advisor Max McPhee laid out a structured approach: ground agents in the company's own context using knowledge graphs and vector-embedded data, then wrap them with governance, ML validation, and strict identity controls. This is the difference between a chatbot that asks "What does that acronym mean?" and an agent that acts like a coworker.

What happened

McPhee described how the same principles used to onboard new employees apply to AI agents, but tuned for information retrieval. The key is combining knowledge graphs with vector-embedded data as a format agents can easily search and consume. This grounding prevents misinterpretation of enterprise shorthand, a problem acute in SAP environments where acronyms and tribal knowledge are the norm.

SAP is modernizing its 50-year-old governance framework to handle the flexibility of autonomous agents. Machine learning is making a comeback as a guardrail: customers run agents within processes and layer anomaly detection and ML-based validation on top, similar to SAP's existing intelligent approval recommendations.

Identity and permissions are a critical execution layer. Under SAP's model, both the human user and SAP's generative AI assistant Joule must be individually provisioned to access a system. Even if a user has rights to S/4, they cannot access it through Joule unless Joule itself has been granted that access. This closes the risk of agents routing around access controls.

Because SAP makes up only about 10% of many customers' landscapes, the company has invested in mapping the rest. LeanIX, described as "Google Maps for your architecture", process-mining company Signavio, and Berlin-based automation platform n8n are being embedded into Joule Studio, SAP's low-code agent builder. These tools let agents understand how enterprise systems interconnect, not just SAP's own modules.

McPhee also warned that on-premises legacy systems must be modernized before scaling autonomous agents. He compared running agents on outdated infrastructure to "driving a Ferrari around a dirt track."

Why AI builders should care

For developers building enterprise AI agents, the message is that general-purpose RAG is not enough. Without SAP Knowledge Graph and enterprise context grounding, agents will misinterpret internal language and make unsafe decisions. The governance layer -- including ML-based validation and dual identity checks -- provides a pattern for safe autonomous execution in regulated environments.

The integration of landscape-mapping tools like LeanIX and Signavio shows that agents need visibility into non-SAP systems to act across the full enterprise. Builders targeting SAP ecosystems must plan for connecting to ERP, CRM, and custom legacy systems, not just one vendor's stack.

Practical implications

• Adopt knowledge-graph-based grounding and vector embeddings as the primary data format for agents. Treat enterprise context like an onboarding document for every agent.
• Implement dual identity checks: both the user and the agent must be explicitly provisioned for each system access. Do not let one bypass the other.
• Invest in landscape mapping (LeanIX, Signavio, n8n, or equivalent) to give agents a unified view of system dependencies. Without that map, agents cannot reason about downstream impacts.
• Layer ML-based anomaly detection and validation as guardrails inside agent workflows, not just as a post-hoc audit.

Caveats

The evidence and quotes come from a VentureBeat article sponsored by SAP. Details may evolve as SAP releases more public documentation. The specific integrations and capabilities of LeanIX, Signavio, and n8n inside Joule Studio may have limitations not yet disclosed. The governance model described is SAP's approach; other vendors may use different strategies.