Box and NVIDIA have expanded their partnership around enterprise AI with the introduction of Box Context Retrieval, an engine designed to help AI agents find relevant information in company content. Announced on November 5, the system uses NVIDIA Nemotron models to retrieve information from files stored in Box. The companies say it aims to make enterprise content more accessible to AI workflows while preserving the permissions that govern who can see particular documents. The collaboration also brings NVIDIA’s open-weight Nemotron models into Box AI and adds security controls for agent execution.
Enterprise AI systems often struggle not because a model cannot generate fluent text, but because it cannot reliably access the right internal information. Business knowledge is spread across contracts, policy documents, customer records, scanned forms, images, audio and video. If an agent retrieves the wrong file or ignores access restrictions, it may produce an incorrect answer or expose information to a user who should not see it. Box and NVIDIA are positioning Context Retrieval as a content layer that addresses these retrieval and governance requirements.
Box says its engine uses NVIDIA Nemotron Retriever embedding models to turn enterprise data into a structured knowledge layer. The company says it can retrieve permission-aware information across millions of files, including formats that have historically been difficult to search with text-only systems. The announcement describes multimodal indexing for scanned documents, images, audio and video, alongside hybrid search that combines keyword matching with semantic understanding. That approach aims to help agents find relevant material even when a user does not know the exact wording used in a document.
Permission-aware retrieval is central to enterprise use
Access control is a basic requirement for corporate search. An employee may be allowed to view a project folder but not legal files, payroll records or another customer’s documents. If an AI agent uses a shared index without preserving those boundaries, it can become a new route for data leakage. Box says Context Retrieval will return information based only on the content the user is permitted to access. That claim is important, but customers will still need to test how permissions are enforced across nested folders, shared links, inherited access and changing employee roles.
Box says customers could improve search accuracy by 20% on Recall@5, reduce token usage by 90% and accelerate question-answering speeds fourfold. These are company-provided performance claims and should be evaluated against the benchmarks, datasets and workloads used in testing. Real enterprise environments vary widely in file quality, access patterns, languages and content volume. The results customers experience will depend on how their information is organised and whether the retrieval system can interpret their particular documents and workflows.
The partnership also includes NVIDIA OpenShell, which Box says will help extend security and governance to agent execution. Retrieval is only one part of an agent workflow. After finding information, an agent may call tools, update records or trigger business processes. Those actions need policy enforcement, supervision and a clear record of what the agent accessed and changed. Box and NVIDIA are working to connect Box’s content security with a secure runtime for agents, although the announcement does not mean every future agent workflow is automatically safe.
Availability and use cases
Box described potential use cases across financial services, legal, healthcare, insurance and education. A financial analyst could query loan applications within authorised records; a lawyer could search contract clauses using different terminology; a clinical researcher could find updated protocols; and an insurance auditor could examine claim files that include scanned forms, images and recordings. These examples show why multimodal retrieval may be useful, but each industry has additional rules governing data access, retention and auditability.
The announcement reflects a larger shift in enterprise AI from general model capability to reliable access to business context. A capable model still needs accurate, current and authorised information before it can complete work safely. Box and NVIDIA are trying to make that content layer a managed part of agent infrastructure. The value will depend on retrieval quality, permission enforcement, integration with business systems and transparent evidence for generated answers. Customers should test the system against their own documents before relying on it for consequential workflows.