BoxWorks 2026: Box Gives AI Agents a Memory for Your Files

Box used its BoxWorks 2026 conference in San Francisco to make its biggest bet yet on AI agents, unveiling Box Context Retrieval, a new engine powered by NVIDIA Nemotron models that lets agents find the right information inside company files, and Box Architect, a tool that designs and builds whole workflows from a plain-language description. Box says the retrieval engine could improve search accuracy by 20 percent, cut token usage by 90 percent and make answers four times faster, while NVIDIA's open-weight Nemotron models, starting with Nemotron 3 Ultra, are available inside Box AI now. Both headline products will sit in the Enterprise Advanced plan, and Box gave no financial updates at the briefing.

What Box Context Retrieval is

Box Context Retrieval is a retrieval engine that turns a company's messy, unstructured content into a structured knowledge layer that AI agents can reason over. That sentence needs unpacking, because it is the core of the announcement. Most of the information a business runs on does not live in neat databases. It lives in contracts, policy manuals, customer emails, research archives, scanned forms, images, audio recordings and video. An AI agent asked to answer a question about that material has two bad options today: read whole files, which is slow and expensive, or guess from too little context, which produces wrong answers. Retrieval is the step that finds the right passages first, so the agent only reads what matters.

Box's engine does that with NVIDIA's Nemotron Retriever embedding models, which create an index of content that the agent can search. It combines keyword search, which matches exact words, with semantic search, which understands meaning and intent, so it can find a relevant clause even when the query uses different words from the document. It uses multimodal indexing to make scanned documents, images, audio and video searchable, and it retrieves relevant sections within documents rather than forcing the agent to process an entire file. Crucially, it respects existing file permissions, so an agent can only see content its user is allowed to see.

Illustration: large tech-conference keynote stage with big screen and audience

The numbers Box claims, and what they do not prove

Box's headline figures are striking. In the company's own recall benchmarks, the tool improved accuracy by more than 20 percent, reduced token consumption by approximately 90 percent and enabled agents to deliver answers four times faster. Box's own blog states the figures as potential gains: up to 20 percent better search accuracy measured as Recall at 5, 90 percent lower token use and fourfold faster question answering. Those are vendor figures, and Box did not specify the comparison used for them during the briefing. Until independent testing exists, they should be read as what Box expects, not what customers have measured.

Availability is staggered. Box Context Retrieval will be generally available to Enterprise Advanced customers early next year, while the NVIDIA Nemotron models, beginning with Nemotron 3 Ultra, are available now inside Box AI to Enterprise Advanced customers. Levie said the advanced agentic and security work logically belongs in that tier, so customers weighing an upgrade should note the gap: the models are here today, but the retrieval engine is still months away.

Box Architect: describe the outcome, get a workflow

Box Architect is the second headline product, and it tackles a different problem. Building an AI workflow today means configuring extraction, automation, apps and approvals step by step, which needs expertise most business teams do not have. Architect flips that around. A team describes the outcome it wants, such as onboarding a client with complete documentation, or reviewing a contract against company requirements, and Architect draws on the organisation's content and procedures already in Box to propose the stages, decisions and handoffs needed, then builds the supporting forms, metadata templates, AI agents, automations and Box Apps, and links them together.

Architect is not tied to one model provider. Customers will be able to choose models from Anthropic, OpenAI, Google, NVIDIA and IBM for each step of a workflow, using lower-cost models for simpler steps and more advanced models for complex reasoning. That flexibility is a deliberate contrast with single-vendor agent platforms, and it mirrors how enterprises actually buy AI: different models for different jobs, with cost matched to the work. Architect enters beta for Enterprise Advanced customers in the fourth quarter of 2026, and Box has not disclosed pricing.

Illustration: office worker sorting stacks of documents beside a laptop showing search results

Security, governance and the rest of the announcements

Box paired the new capabilities with security and governance controls, and those may be the announcements chief information officers act on first, because they are available now. Agent security controls are generally available, with classification-based access policies, guardrails on actions such as external sharing, activity-threshold alerts, and agent session governance and audit trails. New threat protection, including detection of bulk downloads, session hijacking, account takeover and mass deletion, arrives early next year, with adaptive responses that can demand stronger verification, restrict downloads or end a session.

Announcement	What it does	Availability
Box Context Retrieval	Permission-aware retrieval engine on NVIDIA Nemotron Retriever	Generally available early next year, Enterprise Advanced
NVIDIA Nemotron models	Open-weight models, starting with Nemotron 3 Ultra, inside Box AI	Available now, Enterprise Advanced
Box Architect	Designs, builds and optimises agentic workflows from a plain-language outcome	Beta in Q4 2026, Enterprise Advanced
Agent security controls	Classification policies, guardrails, session governance, audit trails	Generally available now
Threat protection	Detection of bulk exfiltration, hijacking, takeover and mass deletion	Early next year
Box for Industries	Editions for banking, public sector, media and entertainment	Announced at BoxWorks 2026

Illustration: business team collaborating around a table with laptops and a workflow diagram on a whiteboard

Why this matters beyond Box customers

Box's announcements point to a shift in enterprise AI that goes beyond one vendor. The first wave of enterprise AI was about chatbots that answered questions. The current wave is about agents that do work, and agents need two things chatbots did not: reliable access to the company's own content, and permission to act on it safely. Box is betting that the company which already stores and governs that content has a natural advantage in supplying both.

That bet connects to themes INORYVA has covered across the agent landscape. Claude Duet Anthropic showed Anthropic pushing agents across tools, Claude can now dream to remember explored how agents keep context over time, and the OpenAI Decisions API showed agents being given structured ways to make choices. Box is solving the other half of the context problem: not what the agent remembers, but what it is allowed to read right now. The same pattern of agents entering real business workflows is visible in Claude Commerce, which extends Claude into shopping and transactions.

Key Takeaways

Box announced Box Context Retrieval at BoxWorks 2026 on 5 November, a permission-aware retrieval engine powered by NVIDIA Nemotron Retriever that gives AI agents relevant passages from company files instead of whole documents.
Box claims potential gains of 20 percent better search accuracy, 90 percent lower token use and four times faster answers, but these are vendor benchmarks with unspecified baselines and should be treated as expectations, not measured results.
Box Architect, entering beta in Q4 2026, designs, builds and optimises agentic workflows from a plain-language outcome, using Box Extract, Automate, Apps, Forms, Doc Gen and Sign, with model choice per step.
NVIDIA's open-weight Nemotron models, starting with Nemotron 3 Ultra, are available now in Box AI for Enterprise Advanced customers, while Context Retrieval arrives generally early next year.
Agent security controls are generally available now, Box for Industries launched for banking, the public sector and media and entertainment, and Box gave no financial updates or Architect pricing at the briefing.

Conclusion

BoxWorks 2026 was Box's clearest statement yet that its future is not storing files but supplying the trusted context AI agents need to work with them. Context Retrieval tackles the problem every enterprise agent deployment hits first, which is finding the right information without breaking permissions or budgets, and Architect tackles the second, which is turning a business goal into a working process without a team of specialists. The NVIDIA partnership gives both efforts credible engines and a secure runtime. The open questions are whether the claimed 20 percent, 90 percent and fourfold gains survive contact with real customer data, and whether enterprises are ready to let agents act on their most sensitive content. The answers will come from the customers who test them first.

Search This Blog

INORYVA