The gap between a chatbot that answers questions and an agent that does real work has always been about context. Max McPhee, senior solution advisor at SAP, made that point bluntly at VB Transform 2026: the difference is grounding agents in a company's own reality rather than general knowledge. That means moving beyond the familiar pattern of retrieving information and toward systems that understand the messy, acronym-laden, customized landscape most enterprises actually run on. It's a useful corrective to the hype cycle. For anyone who has watched a Unlock LLM Training: A Practical Guide to Distributed Algorithms explain how models learn, the leap from raw capability to dependable execution is precisely where most initiatives stall.
What stands out in McPhee's argument is the insistence on knowledge graphs and vector-embedded data as the foundation. He frames it like onboarding a new employee: you don't hand them a dictionary and hope for the best. You give them the tribal knowledge, the internal shorthand, the unwritten rules that make an organization function. For agents, that means structuring data in a format they can actually retrieve and reason over. This is not a theoretical nicety. It's the difference between an agent that processes an invoice and one that knows which approval chain actually matters when a vendor's name is misspelled. The parallel to how humans navigate Exploring Paragraph Structure: How LLMs Navigate Token Space is worth sitting with: just as meaning emerges from how tokens relate to one another, enterprise utility emerges from how data points connect.
The governance piece is where SAP's history gives it a genuine edge, and McPhee leans into that without overstating it. He describes a "revival of machine learning" as a validation layer on top of agent behavior, using anomaly detection to catch mistakes before they become costly. That's a smart, sober framing. More interesting is his point about identity and permissions: both the human and the assistant must hold the access rights. That's not bureaucracy for its own sake. It closes the loophole where an agent becomes a backdoor around access controls. In practice, that means the agent's capabilities are bounded by the same rules that govern the people using it. For enterprises worried about autonomous systems running wild, this is the kind of concrete guardrail that deserves attention.
The honest tension in McPhee's position is that most customers tell SAP it's only 10% of their landscape. He acknowledges this head-on, pointing to acquisitions like LeanIX and Signavio as the map and the process layer for everything else. That's pragmatic. But his warning about on-premises systems is the detail worth watching: you can't run a Ferrari on a dirt track. If you're building agentic workflows, the underlying infrastructure has to keep pace. The takeaway for readers is simple: don't start with the agent. Start with whether your data is structured, governed, and connected enough to support one. Because the agent will only ever be as good as the context you give it, and no amount of model capability will fix a landscape that wasn't built for it.