SAP's Case for Why Enterprise AI Agents Need More Than a Chatbot

Key points

• SAP senior solution advisor Max McPhee told VB Transform 2026 that AI agents feel like coworkers only when they are fed company-specific context, not just general knowledge.
• SAP uses knowledge graphs and vector-embedded data, a format that makes stored information easy for software agents to search and retrieve, to give agents that internal context.
• SAP's Joule AI assistant must be independently granted access to each company system, meaning an agent cannot be used to sneak around existing security permissions.
• SAP has acquired mapping tool LeanIX and process-analysis company Signavio to help agents understand the non-SAP software that makes up most enterprise IT environments.
• McPhee warned that companies running outdated on-premises systems may hit speed limits as they expand agent use.

Most company chatbots know a lot about the world. They know almost nothing about your company.

That gap, according to Max McPhee, a senior solution advisor at SAP, is exactly why so many enterprise AI pilots stall out before they do anything genuinely useful. Speaking at VB Transform 2026 in a session reported by VentureBeat, McPhee argued that the move from a chatbot to an AI agent, software that can carry out multi-step tasks on its own without a human approving every click, depends almost entirely on feeding that agent the right internal knowledge first.

"Where we're starting to see more emergent behavior of it feeling like a coworker rather than an assistant," McPhee said, "is where we're able to provide context on the actual enterprise."

What does "company context" actually mean?

It means the agent knows your acronyms, your approval workflows, and your internal shorthand before it starts work. McPhee compared it to onboarding a new employee, but adapted for software.

SAP does this using knowledge graphs, which are structured maps of how pieces of company information connect to one another, combined with vector-embedded data, a technique that stores text in a format a software agent can search quickly and accurately. The practical result: an agent that understands what your internal project codes mean, rather than stopping to ask.

"Being able to provide that tribal knowledge in the format that's easy for it to consume," McPhee said, "helps to provide a really nice result with your agents versus a chatbot that might say, 'Well, what does that acronym mean?'"

Who controls what the agent can actually touch?

Governance is the part most vendors gloss over. SAP does not.

Under SAP's model, both the human user and Joule, the generative AI assistant built into SAP's cloud products, must separately hold permission to access any given system. A user with access to SAP's S/4HANA finance platform cannot use Joule to reach it unless Joule has also been explicitly granted that access. The agent cannot become a back door.

McPhee also pointed to a quiet comeback for an older discipline: machine learning models, the statistical pattern-recognition systems that predate today's chatbot era, are being layered on top of agents as anomaly detectors. If an agent starts behaving oddly, the ML layer flags it.

"It's becoming a bit of a revival of machine learning," he said.

What happens when SAP is only part of the picture?

For most large companies, SAP handles a fraction of total software. McPhee acknowledged customers who tell SAP bluntly: "You're only 10% of my landscape."

SAP's answer has been to buy its way to a fuller map. Its acquisitions of LeanIX, which McPhee likened to "Google Maps for your architecture," and process-mining company Signavio are meant to chart the non-SAP majority. SAP has also invested in Berlin-based automation firm n8n and is embedding it into Joule Studio, its low-code tool for building agents.

One last caution from McPhee: none of this works well on old infrastructure. "You're going to probably run into throughput issues," he said. "You've got to upgrade the track first if you want to drive a Ferrari."