Box announced an expanded partnership with NVIDIA at its BoxWorks 2026 conference in San Francisco, introducing a new AI engine designed to serve as the backbone for autonomous enterprise agents. The system, branded as Box Context Retrieval, integrates NVIDIA Nemotron Retriever embedding models to help organizations turn vast amounts of unstructured content into a high-speed knowledge base for AI applications.
The integration focuses on solving the “context bottleneck,” where AI agents struggle to find and process relevant information across enterprise silos quickly and cost-effectively. According to company specifications, the new retrieval system provides a 20% improvement in search accuracy—specifically measured by the Recall@5 benchmark—and accelerates question-answering speeds by four times compared to previous iterations.
Efficiency is a core component of the new architecture. Box reported that the system achieves a 90% reduction in token usage. This reduction is significant for enterprise scale, as it allows AI agents to process large document sets with substantially lower compute costs. By optimizing how data is retrieved and fed into large language models (LLMs), the system enables “Agentic RAG” (Retrieval-Augmented Generation), where agents can autonomously navigate complex file structures to find precise answers without manual prompting.
To address the security implications of autonomous agents, Box is also introducing new threat detection tools and a framework dubbed “NVIDIA OpenShell.” These tools are designed to monitor agent activity in real-time to prevent “rogue agents” from accidentally exfiltrating sensitive data or deleting critical files. The security layer is intended to provide administrators with visibility into how AI agents interact with corporate data, ensuring that automated workflows adhere to existing permission structures.
The rollout of these features is split between immediate updates and future releases. While NVIDIA Nemotron models are being integrated into the existing Box AI interface immediately, the full Box Context Retrieval engine is scheduled for general availability in early 2027. Access to these new AI and security tools will be targeted specifically at customers on the Enterprise Advanced and Shield subscription tiers.
