NVIDIA embeds agents with SAP

Enterprise AI is moving out of the chat window and into the software that actually runs companies. That is the big shift here. On May 12 at SAP Sapphire, SAP and NVIDIA said SAP is embedding NVIDIA OpenShell into the SAP Business AI Platform so AI agents can act inside finance, procurement, supply chain, and manufacturing workflows with tighter security and governance controls. ### What changed here? This is not just another “we partner on AI” announcement. SAP is making OpenShell the runtime security layer for SAP AI agents, including custom agents built in Joule Studio, and SAP engineers are co-developing the open-source project with NVIDIA rather than simply plugging it in as a vendor component. ### What is OpenShell actually doing? Basically, it is the fenced yard around the agent. (blogs.nvidia.com) OpenShell provides isolated execution environments, policy enforcement at the filesystem and network layers, and containment meant to limit damage if an agent goes off script or fails in some ugly way. That matters because an enterprise agent is not just drafting text — it may touch files, credentials, APIs, and systems of record. ### Why is SAP the interesting place to do this? Because SAP sits on top of the workflows where mistakes are expensive. Finance close, procurement approvals, supply planning, factory operations — these are not toy use cases. SAP’s pitch is that agents need to be grounded in business context, permissions, and process rules, not just wrapped around a general model and told to “figure it out.” SAP also says its new platform unifies Business Technology Platform, Business Data Cloud, and Business AI into one governed environment. (blogs.nvidia.com) ### Why does “governed execution” matter so much? The catch with agents is that the risk changes when they move from answering to acting. A chatbot can be wrong and annoy you. An agent can be wrong and post entries, move data across systems, trigger workflows, or hit the wrong endpoint. SAP’s framing is that older chatbot-era controls are not enough once agents operate continuously and cross application boundaries. So the hard problem is no longer just model quality — it is permissions, auditing, containment, and identity-aware execution. (news.sap.com) ### Is this part of a bigger SAP push? Yes — and that is what makes the announcement land. The same day, SAP launched its “Autonomous Enterprise” push, introduced a unified SAP Business AI Platform, and said its Autonomous Suite will deploy more than 50 domain-specific Joule assistants that orchestrate a subset of over 200 specialized agents across finance, supply chain, procurement, HR, and customer experience. One example: an Autonomous Close Assistant meant to compress financial close from weeks to days. (news.sap.com) ### Where does NVIDIA fit beyond the runtime? NVIDIA has been working with SAP across the stack already — model development, inference, and enterprise deployment. In March, SAP said NVIDIA NIM microservices were improving inference performance on SAP workloads, with up to a 20% gain versus another open-source serving engine. So this week’s news looks like the next layer up: not just faster models, but safer action. (news.sap.com) ### Why now? Because enterprises are done being impressed by demos alone. SAP had already been pushing AI agents into manufacturing and supply-chain workflows this spring, with the emphasis on execution inside real processes and measurable operational ROI. Once you cross that line, trust stops being a compliance footnote and becomes the product. ### Bottom line The real story is not that SAP and NVIDIA want more AI in the enterprise. (news.sap.com) Everyone does. The story is that the value is shifting toward the execution layer — the part that decides whether an agent can safely do work inside the systems companies actually depend on. If that layer holds up, agents stop being copilots and start becoming infrastructure. (news.sap.com 1) (news.sap.com 2)