Box announces Context Retrieval and Architect for AI agents

Box says much of the knowledge that runs a business sits in unstructured files such as contracts, policies, proposals, meeting notes, and customer records.

“The real opportunity with AI is not just helping individuals work faster, but rethinking how entire business processes get done,” Box CEO Aaron Levie said in Box’s Architect announcement.

According to Box, Box Context Retrieval, announced November 5, is an engine that turns the files in Box into a structured knowledge layer agents can reason over. Box says it combines precise keyword matching with semantic search powered by NVIDIA’s Nemotron Retriever embedding models, and that it indexes material that has been hard for AI to read, including scanned documents, images, audio, and video.

Instead of handing an agent a whole document, Box says it returns the chunks most relevant to the question. Levie told TechCrunch the system returns “chunks from the documents that are most relevant to the user’s query.” Box says every request is permission-aware, so an agent only sees content the requester is allowed to open. In its announcement, Box gave examples such as financial analysts asking questions about loan applications and insurance auditors reviewing claim packages across scanned forms, images, and recorded statements.

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. Those are Box’s projections, not independent test results. NVIDIA’s Kari Briski told TechCrunch she expects queries to be answered four times faster, with 20% fewer errors and 90% fewer tokens used. Box says the engine will be generally available to Enterprise Advanced customers early next year, and that NVIDIA’s open-weight Nemotron models, beginning with Nemotron 3 Ultra, are available in Box AI to Enterprise Advanced customers now.

Box also introduced Box Architect. According to Box, a business owner describes the outcome they want — for example, onboard a client with complete documentation, review a contract against company requirements, or route an insurance claim to the right decision-maker — and Architect uses frontier models plus the content and procedures already in Box to design the workflow, assemble the pieces, and recommend improvements over time.

For a client onboarding workflow, Box says that could mean an intake form, document analysis, routing rules, approvals, and a dashboard, built by agents Architect gathers for the job. Box says Architect redesigns the workflow from the ground up for what frontier models make possible, rather than reproducing the process a team runs today. Box says teams can review, test, and refine a workflow before it goes live, and that Architect can suggest performance updates afterward through conversational prompts. Box says customers will be able to choose models from providers including Anthropic, OpenAI, Google, NVIDIA, and IBM. AppStack Insider reports that Box plans an Architect beta for Enterprise Advanced customers in the fourth quarter; Box’s own Architect post does not state that timing.

Box also announced security updates. Box says it is accelerating unusual-download detection, strengthening defenses against account takeover and session hijacking, and expanding anomaly detection for suspicious sharing and mass deletion. Box says a new adaptive response can ask for stronger verification, limit downloads while allowing approved viewing, or end a session when the risk is severe.

Box says Agent Security adds classification-based policies that apply before external agents reach sensitive files, MCP guardrails that control which tools agents can use and which actions they can take, alerts when activity crosses set thresholds, and session governance with audit trails showing which agent acted and what it did. Box says Box Mount, coming soon, will let agents work directly with Box folders inside isolated sandboxes, including with NVIDIA’s OpenShell secure runtime. “AI agents need more than intelligence. They need the knowledge and context of the business to do useful work,” NVIDIA CEO Jensen Huang said in Box’s announcement. Box also introduced Box for Industries, starting with banking, public sector, and media and entertainment.

Box says Nemotron 3 Ultra is in Box AI now for Enterprise Advanced customers, and that Context Retrieval is planned for general availability to those customers early next year. BoxWorks runs November 5–6 in San Francisco; Box’s BoxWorks site lists Aaron Levie and NVIDIA’s Jensen Huang among featured speakers. Box’s projected figures — fourfold faster question-answering, a 20% Recall@5 gain, and 90% fewer tokens — are its own projections, not results from an independent test.