Safely manage your Zendesk from the AI assistant you already use, via the Deltastring MCP. Beacon configuration platform
← Back to news
ai

Why SAP says enterprise AI agents need knowledge graphs and governance

SAP's position on enterprise AI agents reveals a critical gap between current deployments and what organisations actually need to move beyond conversational interfaces into genuine process automation. Max McPhee's argument centres on two foundational requirements: knowledge graphs that ground agent decisions in accurate, contextualised business data, and governance frameworks that prevent the proliferation of uncontrolled AI systems. This distinction matters because most CX teams currently operate within platforms designed for assisted intelligence—Zendesk's answer bot or Salesforce's Einstein—rather than true autonomous agents that can execute multi-step workflows without human intervention. The implication is stark: teams investing in agent capabilities without first establishing knowledge infrastructure and governance controls are building on sand. For organisations already running Agentforce or similar agent-enabled platforms, this raises an uncomfortable question about whether their current data architecture can actually support the autonomous execution these tools promise, or whether they're inadvertently creating shadow agents that operate with incomplete or contradictory information.

The broader ecosystem context amplifies this concern. Shadow AI agents are multiplying across enterprises, often deployed by individual teams without central visibility, whilst governance frameworks haven't kept pace with deployment velocity. For CX leaders, this creates a dual challenge: your support teams may already be experimenting with AI agents to handle routine inquiries, but without knowledge graphs connecting customer data, product information, and business rules, those agents operate in isolation. They can't learn from previous interactions, can't access real-time inventory or account status, and can't escalate intelligently when context demands human judgment. SAP's argument essentially positions knowledge graphs and governance not as nice-to-have infrastructure but as prerequisites for agent reliability—without them, you're deploying systems that will either hallucinate answers or require constant human override, defeating the automation premise entirely.

The practical consequence for CX operations is that agent maturity requires investment upstream, not just in the agent layer itself. Teams need to audit their data architecture, establish clear ownership of knowledge assets, and implement governance policies that define which agents can access which systems and under what conditions. This is substantially different from the incremental approach many organisations have taken with chatbots, where poor answers simply meant a ticket got routed to a human. Autonomous agents making decisions about customer accounts, refunds, or service levels without proper knowledge grounding and oversight create compliance and reputational risk. The question becomes whether your organisation's current data governance and knowledge management practices are mature enough to support agent autonomy, or whether you're better served consolidating agent capabilities within a single platform where governance is built in rather than bolted on.