Box launches Context Retrieval, a new enterprise search engine built with Nvidia Nemotron models to supply AI agents with permission-aware information from company files. The product was announced at BoxWorks on November 5 alongside Box Architect, a system that creates multi-step business workflows from plain-language descriptions.

 

The launch targets a persistent enterprise AI problem: models can reason over a prompt, but useful agents also need timely access to contracts, policies, images, recordings and other internal material. Box is pairing retrieval, model choice and workflow controls while keeping content under its existing permissions and governance rules.

 

Box's BoxWorks announcement introduces four product layers:

  • Nemotron-powered context retrieval
  • Multimodal and hybrid enterprise search
  • Plain-language workflow design
  • Agent access and threat controls

 

Related Research

 

Box Launches Context Retrieval With Nvidia Nemotron

Box Context Retrieval uses Nvidia Nemotron Retriever embedding models to transform enterprise files into representations that software can search by meaning as well as keywords. The engine is intended to return the most relevant portions of a document, reducing the amount of text an agent must send to a larger language model.

 

In its product announcement, Box said customers could potentially improve Recall@5 search accuracy by 20%, reduce token use by 90% and answer questions four times faster. Those are vendor projections for the new system, not independent performance results across every customer workload.

 

The architecture combines semantic embeddings with conventional keyword search. That hybrid approach matters when a user knows an exact contract clause, account number or product code, while semantic matching can find related concepts expressed with different wording. Permission checks restrict retrieval to content the requesting user is already allowed to access.

 

Box also plans multimodal indexing for scanned documents, images, audio and video. The company says new or changed files will become available to agents in real time, an important requirement for systems expected to work from current policies, customer records or project materials rather than a stale search index.

 

How Nemotron Grounds Enterprise AI Agents

Retrieval is often the boundary between a capable model and a useful business agent. Without a reliable search layer, an agent may consume excessive context, miss a relevant file or produce an answer from incomplete material. Embeddings narrow the search before the reasoning model performs the more expensive final task.

 

Box is also adding Nvidia's open-weight Nemotron models to Box AI, beginning with Nemotron 3 Ultra. Customers on the Enterprise Advanced plan can choose those models for specialized or high-volume work alongside models from other providers, extending Box's existing multi-model strategy.

 

The partnership adds Nvidia OpenShell as a secure runtime for agent execution. Box Mount supplies governed access to content, while OpenShell applies policies and supervision around the code and tools an agent uses. The aim is to preserve file permissions when an agent moves from answering a question to taking an action.

 

Reuters confirmed the November 5 launch and Box's performance targets. TechCrunch separately reported that the engine uses Nemotron's embedding system to return relevant chunks instead of making an agent read every document, which explains the proposed token and latency savings.

 

Box Architect Builds Workflows From Plain Language

Box Architect expands the announcement beyond search. Users will describe a process or goal in ordinary language, and the system will propose stages, decisions and handoffs using Box content and procedures. Teams can then test the workflow with real files, insert human approvals and choose a model for each step.

 

The product can assemble components including Box Extract, Automate, Apps, Forms, document templates and electronic signatures. Box describes examples such as contract review, bank onboarding, government permitting, insurance claims and employee onboarding, each combining document collection, extraction, review and routing.

 

That approach shifts agent building toward process design rather than individual prompts. It could lower the technical barrier for business teams, but the quality of generated workflows will still depend on clear procedures, accurate source documents and careful testing of exceptional cases before deployment.

 

Box says administrators will be able to publish standardized workflows and monitor results, while Architect recommends changes over time. Human review remains configurable at key decisions, a necessary control for legal, financial, government and other processes where an automated mistake can carry regulatory or contractual consequences.

 

Availability and Agent Security Controls

Nemotron models, starting with Nemotron 3 Ultra, are available now in Box AI for Enterprise Advanced customers. Box Context Retrieval is scheduled for general availability on that plan early next year. Box Architect is expected to begin a beta rollout to Enterprise Advanced customers in the fourth quarter.

 

The company is also developing threat detection for bulk downloads, session hijacking, account takeover, suspicious sharing and mass deletion. Planned responses include restricting downloads, requiring stronger authentication or terminating a session. Those features are expected for Enterprise Advanced and Shield customers early next year.

 

Existing agent security controls are already generally available on the same plans. They include classification-based access policies, MCP Guardrails that limit which content tools agents can invoke and threshold alerts when automated activity exceeds a customer-defined limit.

 

The staged availability is a material limitation: the retrieval engine, Architect beta and new threat detection are not all broadly deployed today. Buyers will need production measurements covering recall, latency, token cost and permission enforcement before Box's stated gains can be judged across different file collections and workflow types.

 

Feature photograph: Box headquarters in Redwood City, California, by Jeanjung212, sourced from Wikimedia Commons under CC BY-SA 4.0. The image was cropped to 16:9, resized and converted to WebP; the derivative is shared under the same license.