Field Note from Sapphire 2026: The 'Almost Right' Doctrine

Christian Klein took the stage at SAP Sapphire on May 12, 2026 and said something quietly devastating to the way most enterprise AI vendors are currently selling.

"Accurate, not almost right."

That line will get less coverage than the partnership announcements that followed. It deserves more. Because the rest of the keynote, including a $5.2 billion investment in n8n, a strategic embed of Anthropic's Claude across SAP's flagship products, and a co-developed secure agent runtime with NVIDIA, only makes sense if you accept what Klein said first.

If "almost right" is the disqualifier, the AI-on-ERP debate is no longer about model quality. It is about controls.

What Klein Actually Said

The phrase "almost right" is doing heavy lifting. It signals two things at once.

First, it concedes that current frontier models, including the ones SAP just partnered on, will sometimes return outputs that look correct but are not. That is not a flaw to be debugged away. It is a property of how these systems work.

Second, it asserts that in the workflows SAP customers care about, closing the books, approving an expense, rerouting a supplier order, "almost right" is the same as wrong. Quarterly close cannot tolerate a 95 percent correct number. Treasury cannot tolerate an "almost approved" payment. HR leave inquiries cannot tolerate a confident hallucination about parental policy.

That is not a controversial claim. What is controversial, and what Klein was actually doing, is using that observation to redefine where the AI-on-ERP fight is won.

If you accept the doctrine, you stop competing on which model is smartest. You start competing on which platform can put a sufficiently tight controls layer between the model and the system of record.

Three Things SAP Shipped That Only Make Sense Under This Doctrine

The Sapphire announcements were dense. Strip away the marketing surface and three moves carry the strategic weight.

One. Claude Wired Through MCP, Not Bolted On Top

SAP did not announce a Claude chatbot inside Joule. It announced Claude integrated into S/4HANA, SuccessFactors, and Ariba via Model Context Protocol. That distinction matters.

A chatbot lives outside the approval graph. An MCP-integrated reasoning model lives inside it. When Claude executes an action inside SAP, it inherits the same identity, the same role permissions, the same audit trail, and the same approval workflow that a human would. The model is powerful. The cage around it is what makes it sellable to a CFO.

Two. NVIDIA OpenShell as the Runtime Cage

This is the announcement that got the least coverage and tells you the most about SAP's architecture. OpenShell is an open-source secure runtime for autonomous agents, co-developed by SAP and NVIDIA. It enforces filesystem and network policy at the execution layer. It sandboxes agent operations. It limits blast radius if agent logic fails.

The clean way SAP explained the division of responsibility is worth memorizing.

OpenShell answers: "Can this action safely execute?" Joule Studio runtime answers: "Should this action happen at all?"

That split is the entire control architecture in two sentences. Technical safety at the runtime layer. Business policy at the orchestration layer. Both wrapped around the model.

Three. The SAP Knowledge Graph

Underdiscussed. Possibly the most important piece. A structured map of business entities, units, processes, and the relationships between them. Without it, every "AI agent on SAP" is a chatbot guessing at the schema. With it, agents reason inside the real semantics of the customer's instance. Not synthetic embeddings of documentation. Actual entity relationships.

That is what grounding looks like when the vendor owns the system of record.

The Deeper Read

Holger Mueller at Constellation Research called the Autonomous Enterprise the first real ERP vision SAP has had this side of the millennium. He is right, but the framing is worth tightening.

What SAP is actually saying is this: when AI agents start touching real money inside real ERP systems, the winning vendor is not the one with the smartest model. The winning vendor is the one that can guarantee the model behaves inside the same controls that govern human users.

That is a posture only a system-of-record vendor can credibly take. Anthropic cannot promise it from the outside. Neither can OpenAI. Neither can the hyperscalers. They can promise model quality. They cannot promise approval-graph enforcement, role-based identity, audit trails, or sandboxed execution against your specific entity model. That is the structural advantage SAP just monetized.

It also explains why SAP could afford to be model-agnostic where it counts. Anthropic for primary reasoning. Mistral and Cohere for sovereign deployments. Google and Microsoft for agent-to-agent interop. SAP is not betting on a model. It is betting that the controls layer is where customers will actually pay.

What This Means for Buyers Right Now

A few practical implications for CFOs, CIOs, and ERP program leads.

If your AI strategy is anchored on which model you are evaluating, you are anchored on the wrong axis. The model is a commodity input. The controls integration is the differentiator. Ask vendors how their AI agents inherit your existing approval workflows, role permissions, and audit requirements. If the answer is "we will build it on top," that is a multi-year project, not a feature.

If you are on ECC and weighing the S/4HANA cloud move, note SAP's quiet but important policy choice. On-prem customers get AI scenarios conditional on a cloud commitment. Agents are now the migration carrot. The economics of waiting just changed.

If you are already on RISE, the inclusion of three Joule Assistants in the first year and the doubling of the RISE on Azure success program means the activation cost is being subsidized aggressively. Use that window.

What I Am Writing Next

This is a field note, not the full analysis. Three follow-ups will go up over the next two weeks.

A deeper architectural read of OpenShell and the Joule Studio runtime split, and what it means for any vendor trying to build secure agents.

A buyer's framework for evaluating "agentic ERP" claims, including the questions to ask that separate marketing from architecture.

A counter-take on where the SAP strategy is exposed, specifically the commoditization risk on the Joule Work conversational layer that every major vendor is now shipping.

If you walked away from today's keynote with the impression that SAP added Claude, you saw the surface. The actual announcement is that the largest ERP vendor in the world just made governance the entire pitch. That is not a small move. It is the move that decides who wins enterprise AI for the next five years.

If you are sketching your enterprise's response to today's news and want a second pair of eyes before it goes to your CIO or audit committee, reach out.

Shubhendu Tripathi is an AI and ERP strategy consultant based in Toronto, advising organizations on digital transformation, enterprise AI adoption, and technology leadership. Connect on LinkedIn or reach out at tripathis@qubittron.com.