Box is expanding its partnership with NVIDIA to address two persistent challenges in enterprise AI: giving agents reliable access to company knowledge and controlling what they can do with that information. Announced at BoxWorks 2026, the collaboration introduces Box Context Retrieval, powered by NVIDIA Nemotron retrieval models, adds Nemotron open-weight models to Box AI and integrates NVIDIA OpenShell into the companies’ work on secure agent execution. The strategy positions enterprise content management as a governed knowledge layer for AI agents, rather than simply a repository of files.
Article: Enterprise AI Needs More Than a Powerful Model
The announcement brings together three layers of the enterprise AI stack: retrieval, model choice and runtime governance. Each addresses a different obstacle to deploying agents in business environments.
The first is the problem of finding reliable information. Enterprise content is often distributed across contracts, policy documents, presentations, scanned forms, customer records and other unstructured files. Traditional keyword search can miss relevant material when a user or agent does not know the exact terminology used in a document. A general-purpose language model, meanwhile, can produce a plausible response without necessarily locating the right internal evidence.
Box Context Retrieval is designed to close that gap. It uses NVIDIA Nemotron Retriever embedding models to create a structured representation of enterprise content that AI systems can search. The engine combines semantic retrieval, which looks for meaning and context, with keyword matching. Box says this will help agents find relevant information across millions of files, including scanned documents, images, audio and video.
The system is also designed to respect existing content permissions. An agent should retrieve only material that the user is authorized to access, rather than treating the entire corporate repository as a single pool of information. That distinction is particularly important in organizations where financial records, employee information, customer contracts and legal documents have different access restrictions.
Box says Context Retrieval could improve search accuracy by 20% on Recall@5, cut token usage by 90% and make question-answering four times faster. These are company-provided performance expectations, not independently verified benchmarks disclosed with the announcement. Recall@5 measures whether relevant results appear among the top five retrieved items; it should not be interpreted as a 20% reduction in hallucinations or errors across all AI tasks.
Nemotron adds another model option
The second part of the partnership expands the model choices available inside Box AI. NVIDIA’s Nemotron family of open-weight models, beginning with Nemotron 3 Ultra, is now available to Enterprise Advanced customers.
The move fits Box’s multi-model strategy. Rather than tying every workflow to one model provider, the company wants customers to select models based on the task, performance requirements, cost and level of control. Open-weight models can give organizations more flexibility to inspect, adapt or deploy model technology within supported environments, although open weights alone do not mean that every component of a model’s training data, development process or serving infrastructure is open.
NVIDIA describes Nemotron 3 Ultra as a 550-billion-parameter mixture-of-experts model, with up to 55 billion parameters active for a given token. It is designed for complex reasoning and long-running agent tasks. Bringing this model into Box AI gives customers another option for specialized workloads, although the announcement does not provide a direct comparison of its performance against the other models available in Box.
The distinction matters for enterprise buyers. A model may be strong at reasoning, but its value in a business workflow also depends on the quality of the information it receives, the permissions attached to that information and the ability to complete tasks without exceeding its authority. Retrieval and model selection therefore need to work together.
OpenShell addresses the risks of agent execution
The third component focuses on what happens after an AI agent receives information and decides what to do with it.
As agents move from answering questions to taking actions across enterprise systems, organizations face a different category of risk. An agent may need to retrieve a contract, prepare a summary, update a workflow or communicate with another application. If its access is too broad, a mistaken instruction or compromised agent could have consequences beyond an inaccurate answer.
Box and NVIDIA are combining Box’s content security and governance capabilities, including Box Mount, with NVIDIA OpenShell. NVIDIA describes OpenShell as a secure runtime that isolates agent workloads, restricts file and network access through policies, and supervises requests outside the agent’s own process. Its architecture includes sandboxing, policy enforcement and audit logs for allow-and-deny decisions.
In the Box collaboration, the intended result is to give agents access to the content they need while adding controls around the tools and actions they can use. This is a significant architectural distinction: content permissions govern what information an agent can retrieve, while runtime controls govern how an agent interacts with its execution environment and external services.
The companies’ participation in the Open Secure AI Alliance provides a broader context for the work. However, the announcement describes an expanded collaboration and a direction for safer agent workflows, not evidence that every possible agent action is now protected against misuse. Security will depend on how policies are configured, how integrations are implemented and how organizations monitor agent behavior.
Enterprise use cases extend beyond document search
The potential applications span industries where decisions depend on large volumes of sensitive or difficult-to-search information.
Financial analysts could query loan applications while respecting customer-level access restrictions. Legal teams could search for clauses and precedents even when contracts use different terminology. Clinical researchers could retrieve updated protocols and policies, while insurance auditors could examine claim packages containing scanned forms, photographs, recorded statements and video evidence.
These examples illustrate why multimodal retrieval matters. Enterprise knowledge is not limited to text documents, and an agent that can search only conventional files may overlook relevant evidence. Making different content formats searchable could broaden the range of workflows that organizations can automate.
The challenge will be ensuring that retrieval results are accurate, current and traceable. In regulated environments, an answer grounded in the wrong version of a policy or an incomplete claim record can be more problematic than no answer at all. Permission-aware retrieval is an important control, but it does not replace source validation, human review or broader compliance processes.
Availability and the enterprise adoption question
The rollout is divided between capabilities available now and those still planned. NVIDIA Nemotron models, beginning with Nemotron 3 Ultra, are available in Box AI for Enterprise Advanced customers. Box Context Retrieval is scheduled for general availability to the same customer tier in early 2027.
That timing means organizations evaluating the announcement should distinguish between the currently available model integration and the forthcoming retrieval engine. The wider vision of combining content retrieval, open-weight models and secure agent execution will depend on how these components work together in production.
The partnership also reflects a shift in enterprise AI competition. Cloud providers and model developers have largely focused on improving model intelligence, inference economics and agent frameworks. Content platforms such as Box are emphasizing a different requirement: connecting those models to proprietary business knowledge without losing control over access and governance.
For Box, the opportunity is to make its content platform a foundational layer for enterprise agents rather than a passive destination for storing and sharing files. For NVIDIA, the collaboration extends Nemotron and OpenShell into an established enterprise content environment.
The commercial test will be whether customers see measurable improvements in retrieval quality, response speed and operating costs while maintaining appropriate security controls. If the performance claims hold up under independent testing and the governance model proves practical at scale, the collaboration could strengthen the role of enterprise content platforms in agent-based automation. Until then, the announcement represents a significant product direction, with the most important Context Retrieval capability still awaiting general availability.
Market Landscape
Enterprise AI is moving from standalone assistants toward systems that retrieve proprietary information, coordinate multiple agents and execute tasks across business applications. This creates demand for three connected capabilities: retrieval-augmented generation (RAG), flexible model infrastructure and runtime security.
Box is using its content-management position to address the knowledge layer, while NVIDIA supplies embedding models, open-weight reasoning models and agent-runtime technology. The strategy competes less directly with individual chatbot products and more with the growing ecosystem of enterprise search, vector databases, RAG platforms and agent orchestration tools. The differentiator will be whether Box can combine existing content permissions with accurate retrieval and controlled execution without adding excessive complexity.
Top Insights
- New retrieval engine: Box Context Retrieval uses NVIDIA Nemotron Retriever models for semantic and keyword-based search across enterprise content.
- Permission-aware AI: The system is designed to retrieve information based on users’ existing access rights.
- Open-weight model choice: Nemotron 3 Ultra adds another model option to Box AI for enterprise workloads.
- Agent security: NVIDIA OpenShell adds sandboxing, runtime policy enforcement and auditability to the agent execution layer.
- Staggered availability: Nemotron models are available to Enterprise Advanced customers, while Context Retrieval is planned for early 2027.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
