AI News
Box Unveils Box Architect and NVIDIA-Powered Context Retrieval for Enterprise AI Agents
Box used its BoxWorks 2026 conference in San Francisco on November 5 to lay out its next step in enterprise AI: a set of agent tools built with NVIDIA, led by a new workflow builder, Box Architect, and a retrieval engine, Box Context Retrieval, that turns a company's stored content into a knowledge layer AI agents can work from. The announcements came in three Business Wire releases published the same morning, and the keynote paired Box chief executive Aaron Levie with NVIDIA chief executive Jensen Huang.
The pitch is that AI agents are only as useful as the context they can reach. A model on its own can write and reason, but it cannot answer a question about a contract, a claim or a research file unless it can find the right document, respect who is allowed to see it, and quote it accurately. Box, which already stores that content for its enterprise customers, is now selling the layer that connects the two.
What Box Context Retrieval does
Box Context Retrieval is a retrieval engine built on NVIDIA's Nemotron Retriever embedding models, specifically the Nemotron 3.5 Retriever family: Embed, Rerank and Parse. It turns enterprise content into what Box calls a structured, permission-aware knowledge layer for agents, handling scanned documents, images, audio and video, not just clean text files. Search is multimodal and hybrid, combining keyword matching with semantic search, so an agent can find a clause in a scanned contract or a moment in a recorded meeting as well as a line in a document.
It is not available yet. Box says Context Retrieval will reach general availability for Enterprise Advanced customers in early 2027. Alongside it, NVIDIA's open-weight Nemotron models, beginning with Nemotron 3 Ultra, are now available inside Box AI for Enterprise Advanced customers.
Box Architect builds workflows from plain language
Box Architect is the second headline product. It builds, manages and optimises agentic workflows from plain-language descriptions, so a team can describe a business process and get a working agent workflow rather than wiring one together step by step. Workflows can be designed, tested and deployed with human review at key decisions, and they connect to Box Extract, Automate, Apps and Forms, so an agent can pull data from a document, trigger an approval and file the result without leaving Box.
Each step of a workflow can use a different model, with a choice across Anthropic, OpenAI, Google, NVIDIA and IBM, so a team is not locked to one provider for every task. Architect is due in beta for Enterprise Advanced customers in the fourth quarter of 2026. Levie put the ambition plainly: “The real opportunity with AI is not just helping individuals work faster, but rethinking how entire business processes get done.”
Security controls for agents are live now
The agent security controls announced alongside the two products are generally available now for Shield and Enterprise Advanced customers. They include classification-based access policies, which limit what an agent can touch based on how content is classified; MCP Guardrails, which limit which tools agents can use, including whether they can create, move or share files externally; and threshold-based activity alerts that flag unusual agent behaviour.
Deeper threat protection follows early next year, also for Enterprise Advanced and Shield customers: faster detection of anomalous downloads, session binding and device-bound credentials to fight session hijacking and account takeover, and severity-based automatic responses to bulk exfiltration and mass deletion. Box also announced Box Mount with NVIDIA OpenShell, combining content governance with a secure runtime for agent execution.
Box for Industries, and the customer numbers
A separate release the same day introduced Box for Industries, available now for Banking, Public Sector, and Media & Entertainment, with Legal, Insurance, Life Sciences, Education and Retail to follow. The industry packages carry FedRAMP High, GovRAMP High and DoD IL4 compliance. Box's customer example is Norinchukin Bank, which Box says cut contract-review time by about 88% using its tools — that figure is Box's own customer claim.
The benchmark figures in the announcement need the same treatment. Box says customers “will have the potential to improve search accuracy by 20 percent (Recall@5), reduce token usage by 90 percent, and accelerate question-answering speeds fourfold.” That is Box's wording from its own press release, and the company did not disclose the baseline those improvements are measured against anywhere in the announcement — its chief technology officer, Ben Kus, did not specify the comparison at the investor briefing either. Treat them as Box's claims, not independent results.
What Box did not say
No pricing, revenue or adoption figures were given for any of the new products, and the three most important pieces are not shipping yet: Context Retrieval is early 2027, Architect is a fourth-quarter beta, and the deeper threat protection is early next year. What is live today is the agent security controls and Box for Industries. Huang's line from the keynote sums up the bet Box and NVIDIA are making together: “AI agents need more than intelligence. They need the knowledge and context of the business to do useful work.” Levie's version is the same idea from the other side: “Enterprise AI reaches its full potential when models and agents can securely access the unique context of a business.”