NVIDIA and SAP Add Trust to Specialized AI Agents

At SAP Sapphire, with NVIDIA founder and CEO Jensen Huang joining via video, SAP and NVIDIA expanded their collaboration to run specialized AI agents more safely within enterprise systems.

SAP is embedding NVIDIA OpenShell into the SAP Business AI Platform, and SAP engineers are also joining the joint development of the open-source codebase. OpenShell provides isolated execution environments, filesystem- and network-level policy enforcement, and infrastructure-level isolation to reduce damage when an agent fails.

The engineering focus is on runtime hardening, policy modeling, enterprise identity integration, and audit/governance hooks. NVIDIA has also long run its finance, supply chain, and logistics operations on SAP as a customer, so both companies view the control requirements needed in real enterprise environments through the same lens.

Enterprise agents are not simple chatbots—they are executing entities that touch the system of record and cross application boundaries. That's why the two layers' roles are separated.

• OpenShell asks whether this task can be executed safely.
• The Joule Studio runtime determines whether this task should be performed at all.

In NVIDIA's proposed AI five-layer cake, SAP is highlighted as the core of the application layer. In SAP environments handling finance, procurement, supply chain, and manufacturing workflows, roles, permissions, data boundaries, where inference is performed, and audit trails must all be managed together.

Additionally, NVIDIA NemoClaw is now offered directly within Joule Studio, letting SAP customers move custom agents into production faster without building security scaffolding from scratch. Together, the two companies aim to first establish a trust layer that enterprises can entrust their data to, then scale autonomous AI on top of it.