Box partners with Nvidia to build context layer for AI agents

Solving the context bottleneck

Box is expanding its partnership with Nvidia to address a primary obstacle in enterprise AI: the lack of context. For autonomous agents to be useful, they need to understand the specific details of a company's operations, which are often buried in disorganized files. As Box CEO Aaron Levie noted, agents without this context are likely to make mistakes, which prevents companies from trusting them with real work.

Most corporate data is 'unstructured,' meaning it exists in formats like scanned documents, internal emails, policy manuals, and research archives rather than clean databases. Companies want to use this data to train or inform their AI tools, but they are often afraid of letting autonomous systems jump into these files without oversight. There is a significant need for a system that provides secure, permission-aware access to these archives.

The new partnership aims to create an intelligence layer that bridges this gap. By providing a way for agents to reason over messy content, Box hopes to move AI from simple chat interfaces to functional, autonomous workflows. This approach ensures that the AI stays grounded in the actual facts of the business while maintaining strict security boundaries.

How the retrieval engine works

The core of this update is a new tool called Box Content Retrieval. This engine is designed to transform messy enterprise content into a structured knowledge base that AI agents can easily navigate. It utilizes Nvidia’s specialized Nemotron Retriever embedding models to create a searchable index of various file types, including images, audio, and video files.

The system uses multimodal and hybrid search techniques to improve how the AI finds information. According to Box, this technology can improve search accuracy by 20% while reducing token consumption by up to 90%. This efficiency allows AI agents to find the necessary context much faster, potentially improving response times to worker questions by around four times.

A critical component of this technology is its integration with native file permissions. The system ensures that an AI agent can only access content that the specific user it is assisting is authorized to see. This prevents the agent from accidentally leaking secrets or viewing sensitive files that fall outside of its designated scope.

Open-weight models and costs

In addition to the retrieval engine, Box is bringing Nvidia’s frontier open-weight models to its platform. These include the new Nemotron 3 Ultra large language model, which will be available directly within Box AI. Because these models are open-weight, they can be accessed at more affordable prices than the proprietary models offered by companies like OpenAI or Anthropic.

Nvidia's open-weight models are specifically trained on domain-specific knowledge for highly specialized tasks. For example, financial analysts could use Nemotron Ultra 3 to query loan applications while staying within strict data boundaries. Legal teams might use it to identify contract precedents by understanding conceptual meanings rather than just searching for specific keywords.

By offering these models, Box aims to drive broader adoption across the industry. The company views open-weight models as a way to complement closed models while providing more flexibility for specialized workflows. This allows businesses to choose the right balance between high-end performance and cost-effective, specialized intelligence.

Orchestration and security

To manage these agents, Box is also introducing Box Architect, a tool launching in beta later this year. This platform allows workers to build and deploy agentic processes using natural language. It can connect to native Box tools to create complex workflows, such as a bank using it to automate client onboarding by verifying identity documents and generating signed agreements.

Because autonomous agents can perform multiple steps independently, Box is also introducing new threat detection tools. These capabilities are designed to identify the actions of 'rogue' agents in real-time. If an agent begins to delete massive numbers of files, exfiltrate data in bulk, or take over user accounts, the system will immediately flag the activity for review.

Administrators can further control these agents using classification-based policies and MCP guardrails. These guardrails dictate which third-party tools an agent is allowed to use. This layered security approach is intended to give organizations the confidence to let agents handle more complex, multi-step tasks without losing control over the underlying data.