Box Context Retrieval and Architect: BoxWorks 2026 AI, Explained

Box has spent twenty years as the place where enterprises store their documents. At BoxWorks 2026, the company made its case for what those documents are for next: feeding AI agents that do real work. On November 5, 2026, Box announced two products at its San Francisco conference, Box Context Retrieval and Box Architect, built with NVIDIA, and both aimed at the same problem. An agent is only as useful as the information it can reach, and most enterprise information sits in files the agent cannot read well, cannot search precisely, or is not allowed to touch.

The announcements land in a week when every enterprise software vendor is racing to put agents inside customer workflows. Salesforce rolled out Agentforce Coworker at $50 per user per month, and Perplexity launched Teammates for its $200-a-month Max subscribers. Box is taking a different route. Instead of selling a new agent, it is selling the layer underneath: retrieval, workflow building, and security controls for agents that run on content already stored in Box. The catch is timing and price. The most interesting piece, Context Retrieval, does not reach general availability until early next year, and Box has published no price for the plan it requires.

This article covers what Box announced, what is available today, what the performance claims actually mean, and what it will cost, as far as Box has said.

What Box announced at BoxWorks 2026

Box made two announcements at BoxWorks 2026, its annual customer conference, on November 5, 2026. The first, in a joint release with NVIDIA, introduced Box Context Retrieval, an engine that connects AI agents to enterprise content, and brought NVIDIA Nemotron open-weight models into Box AI. The second introduced Box Architect, a tool for building agentic workflows in plain language, plus new threat protection for agent activity. Both releases were published on Box's blog the same day, and Box CEO Aaron Levie and NVIDIA CEO Jensen Huang presented the NVIDIA partnership in the opening keynote.

Three things are available now: NVIDIA Nemotron models, starting with Nemotron 3 Ultra, inside Box AI for Enterprise Advanced customers; agent security controls for Shield and Enterprise Advanced customers; and the underlying Box AI features the new tools build on. Two things are not: Context Retrieval, which reaches general availability early next year, and Box Architect, which begins rolling out in beta in the fourth quarter of 2026. The new threat protection capabilities arrive early next year as well. That split matters for anyone planning a budget or a rollout, because the headline features are promises with dates, not products you can turn on today.

Box Context Retrieval: the engine that feeds agents

Box Context Retrieval is a retrieval engine powered by NVIDIA Nemotron Retriever, a collection of open models that turns enterprise data into what Box calls a structured knowledge layer that agents can reason over. In practice, it does the job every retrieval system does: when an agent asks a question, it finds the passages that answer it, instead of handing the model whole files.

Three details define how it works. First, retrieval is permission-aware. Queries evaluate the user's live Box permissions, so an agent only sees content that user is authorized to view. That is the feature that separates an enterprise retrieval tool from a consumer one, and it is the reason Box can pitch this to banks, law firms, and hospitals. Second, it uses hybrid search, combining keyword matching with semantic search powered by NVIDIA Nemotron embeddings, so it can find a clause even when the search terms do not match the document's wording. Third, it indexes content that has historically been hard for AI to interpret: scanned documents, images, audio, and video, not just text files.

Box also says indexing will be faster and lower-latency, making new and updated content available to AI agents in real time. The company frames the use cases by profession: financial analysts asking questions about loan applications, legal professionals finding clauses across contracts, clinical researchers working with newly updated care protocols, and insurance auditors analyzing claim packages that span scanned forms, property images, recorded statements, and video evidence.

The 20%, 90%, and 4x claims need a closer look

Box says customers will have the potential to improve search accuracy by 20 percent on Recall@5, reduce token usage by 90 percent, and accelerate question-answering speeds fourfold. Those are the numbers that will travel furthest from this announcement, so they deserve care.

All three figures come from Box's own recall benchmarks, as reported in the announcement and in coverage of it. They are company projections, not independent measurements, and Box's own wording is conditional: customers "will have the potential to" see these gains. No independent test of Context Retrieval exists yet, because the product is not generally available. The 90 percent token reduction is the most consequential of the three if it holds up, since token consumption is the main cost driver of agent workloads that read long documents. But until Box publishes benchmark details or customer results at general availability, treat all three numbers as the vendor's target, not a measured result.

There is also a limit worth stating plainly. Context Retrieval improves the odds that an agent works from the right passages. It does not verify that the passages are correct, current, or complete, and it cannot fix a document that was wrong when it was filed.

Nemotron 3 Ultra is live now, for Enterprise Advanced only

The piece of the announcement that is available today is the model choice. NVIDIA's Nemotron family of open-weight models is coming to Box AI, beginning with Nemotron 3 Ultra, and it is now available to Enterprise Advanced customers. Box's pitch is cost and control: open-weight models can be cheaper to run than proprietary models from Anthropic and OpenAI for specialized, high-volume workloads, and customers can pick the right model for each task instead of taking one vendor's default.

Box AI already supports models from multiple providers, and Box Architect extends that choice to workflow steps, letting teams use a lower-cost model for simple tasks and a more capable model for complex reasoning. The open-weight angle also matters for transparency: because the model weights are publicly released, customers and auditors can inspect what they are running in a way they cannot with a closed model.

The limit is the plan gate. Nemotron models in Box AI are available to Enterprise Advanced customers only. Box publishes no per-token price for Nemotron usage inside Box AI, and no separate charge has been disclosed for it. Usage is metered through Box AI Units, which are covered in the pricing section below.

Box Architect: workflows described in plain language

Box Architect is the second announcement, and it is the one aimed at people who do not write code. Teams describe a process or a goal in plain language, and Architect designs the workflow, drawing on content and procedures already stored in Box, then builds it from Box tools: Extract for pulling data from documents, Automate for routing, Apps for interfaces, and Forms for collection.

The workflow has three stages. In design, Architect defines stages, decisions, and handoffs, with every step governed by existing permissions, classifications, and retention policies. In testing, teams try the workflow on real content and edge cases, preview agent outputs and routing, add human review and approvals at key decisions, and choose the model for each step. In deployment, Architect monitors results and recommends improvements, while administrators publish and standardize workflows across the enterprise.

Box's examples show the intended audience. A law firm can ask Architect to build a contract review workflow that learns the firm's approved playbook from past agreements, compares incoming contracts against it, flags exceptions, and routes high-risk clauses to attorneys. A bank can turn client onboarding into one connected workflow that collects applications through Forms, verifies documents with Extract, generates and signs agreements, and starts account setup with Automate. Government agencies get permitting workflows; insurers get claims processing; HR teams get employee onboarding. In each case the pitch is the same: the process is built from the company's own documents, and a person stays in the loop at the decisions that matter.

Architect begins rolling out to Enterprise Advanced customers in beta in the fourth quarter of 2026. Box has disclosed no price for it. Aaron Levie framed the idea in the announcement by pointing to frontier models, naming OpenAI's Astra and Anthropic's Fable as the breakthroughs that make this kind of workflow building possible, and arguing that the opportunity is not helping individuals work faster but rethinking how entire business processes get done.

The security half: controls now, threat detection in 2027

Box paired the two product announcements with security work, and the timing split is the same. Agent security controls are generally available now to Shield customers and those on Enterprise Advanced plans. The new threat protection capabilities arrive early next year.

The controls available today do three things. Classification-based policies govern which agents can reach which classified content. MCP Guardrails let administrators set boundaries on the tools agents can use, including creating, moving, and sharing content externally. Threshold-based policies alert teams when agent actions exceed customer-defined limits, so staff can intervene before the damage spreads.

The threat protection coming early next year targets attacks that are specific to agents: bulk exfiltration, session hijacking, account takeover, and mass deletion. Box says it will detect anomalous downloads faster and restrict them automatically, strengthen protection against session hijacking with session binding and device-bound session credentials, identify suspicious sharing and mass deletion, and apply automatic severity-based responses, from restricting downloads to requiring step-up authentication or terminating a session.

Box is also working with NVIDIA on secure agent execution through NVIDIA OpenShell, a secure runtime that gives agents the content they need within policy enforcement and supervision, combined with Box Mount. Both companies are members of the Open Secure AI Alliance. Box also announced Shield Pro, an add-on suite of AI-powered security capabilities, at the same conference. It requires Box Shield to purchase and carries an additional cost, which Box has not published.

Pricing: the number Box did not publish

The pricing story of this announcement is what is missing. Box publishes no price for Enterprise Advanced, the plan that all three new features require, and no price for Architect or Context Retrieval themselves. On Box's own pricing page, read live on November 8, 2026, the Enterprise Advanced card reads "Starting at" followed by a blank space, then "per user/month," with annual billing only and a minimum of three users. That is not a typo in this article. The number is simply not there.

What Box does publish is the metering. Enterprise Advanced starts at 20,000 Box AI Units per month, with 100,000 API calls per month, unlimited storage, and a 50 GB file upload limit. Box AI Units are the company's new metric for AI usage: each AI interaction, from querying content to extracting metadata, is quantified into discrete units. Box says the metric exists so customers can forecast AI needs and manage budgets. The pricing page also warns that additional charges may apply for API calls generated and AI Units consumed, that platform add-ons are available only on a recurring annual contract, and that there are no rollovers: unused quantity expires at the end of the term.

For comparison, the free tier is free but small. The Free Developer plan costs $0 for a single user and includes 1,000 Box AI Units per month, 10,000 API calls per month, 10 GB of storage, and a 2 GB file upload limit. It is a developer sandbox, not a way to run Context Retrieval or Architect, which are Enterprise Advanced features.

Because Box publishes no dollar price for Enterprise Advanced, there is nothing to convert into pounds or Canadian dollars. For reference, November 8, 2026 market rates were 1 USD = 0.7471 GBP and 1 USD = 1.4129 CAD, verified at publication. Those rates will matter the moment Box publishes a price, or the moment a sales quote arrives. Until then, the honest answer to "what does it cost?" is that it depends on a quote, and the 20,000 AI Units included in Enterprise Advanced are the only hard number on the page.

That quote-only model is the norm for enterprise software, and it is also the reason the pricing section of this article is short. Salesforce publishes its Agentforce Coworker price, $50 per user per month, and Perplexity publishes its Teammates price, $200 a month on the Max plan. Box publishes neither a price nor a per-unit rate for AI Units, so a buyer cannot compare the cost of running agents in Box against either rival without talking to sales first. That is a real gap, and it is the first thing to ask about in any evaluation.

How this compares with the other agent plays

Box is not the only vendor turning its platform into an agent layer this month, and the differences say a lot about who each product is for. Salesforce Agentforce Coworker, announced November 5, 2026, is a $50-per-user-per-month agent that employees assign tasks to through one interface, with Teams rollout beginning and 21-fold usage growth in three months. Perplexity Teammates, launched November 4, 2026, are persistent agents that own an ongoing job, available only to individual Max subscribers at $200 a month on the web. Google's Data Agent Kit, which reached general availability on September 30, 2026, gives developers free MCP tools for more than 15 Google Cloud data services. AWS Infinity, announced October 23, 2026, is an open-source framework for running tens of thousands of agents with near-zero compute while they wait.

Box sits apart from all four. It is not selling an agent, a coworker, or a framework. It is selling retrieval, workflow building, and governance for agents that work on content its customers already store. That is a narrower pitch, and for a company whose customers include banks, law firms, and hospitals, it may be the right one. The risk is that it is also the least visible: a retrieval engine and a workflow builder do not demo as well as a coworker that answers in chat, and their value only shows up after deployment, in fewer tokens spent and fewer wrong answers given.

Common mistakes when evaluating this announcement

Three mistakes are easy to make with this announcement, and Box's own materials invite at least two of them. The first is treating the 20 percent, 90 percent, and fourfold figures as measured results. They are Box's own benchmarks and projections, and the product has not shipped yet. The second is assuming Context Retrieval is available now. It is not: general availability is early next year, and the only piece live today is Nemotron 3 Ultra inside Box AI. The third is budgeting from the 20,000 AI Units figure without knowing what an AI Unit costs to exceed. Box has not published a per-unit rate, and the pricing page says additional charges may apply, so the included allocation is a floor, not a cap on spending.

For teams evaluating Box AI, the practical questions are concrete. Ask sales for the Enterprise Advanced per-user price and the AI Unit overage rate in writing. Ask how Recall@5 was measured, on what data, and whether Box will publish the benchmark. Ask whether Architect's beta includes your industry's workflow templates, and whether the threat protection due early next year is included in Enterprise Advanced or priced separately. And ask how Nemotron 3 Ultra's quality compares with the closed models already in Box AI for your documents, because the cost argument only holds if the answers stay good.

Frequently asked questions

When will Box Context Retrieval be available?

Box says the Context Retrieval engine will be generally available to Enterprise Advanced customers early next year, meaning early 2027. It is not available today. The related NVIDIA Nemotron models, beginning with Nemotron 3 Ultra, are already available in Box AI to Enterprise Advanced customers.

What does Box Context Retrieval cost?

Box has published no price for Context Retrieval, and no separate charge has been disclosed for it. It requires Enterprise Advanced, whose per-user price Box also does not publish. Enterprise Advanced starts at 20,000 Box AI Units per month, and additional AI Unit and API usage can carry extra charges. Expect a sales quote.

Can I use NVIDIA Nemotron models in Box AI today?

Yes, if you are an Enterprise Advanced customer. NVIDIA Nemotron models, beginning with Nemotron 3 Ultra, are now available in Box AI for Enterprise Advanced customers. Box has published no per-token price for Nemotron usage inside Box AI; usage is metered through Box AI Units.

What is Box Architect?

Box Architect is a tool that builds agentic workflows from a plain-language description of a process or goal, using content and procedures already stored in Box. It designs, tests, and deploys workflows built from Box Extract, Automate, Apps, and Forms, with human review at key decisions and a choice of model for each step. It begins rolling out in beta to Enterprise Advanced customers in the fourth quarter of 2026.

What is a Box AI Unit?

A Box AI Unit is Box's meter for AI usage. Each AI interaction, from querying content to extracting metadata, is quantified into discrete units. The Free Developer plan includes 1,000 AI Units per month; Enterprise Advanced starts at 20,000 per month. Box has not published a price per AI Unit, and unused units do not roll over.

What security controls does Box offer for agents?

Agent security controls are generally available now to Shield and Enterprise Advanced customers, covering classification-based access policies, MCP Guardrails on agent tools, and threshold-based alerts. New threat protection against bulk exfiltration, session hijacking, account takeover, and mass deletion arrives early next year for Enterprise Advanced and Shield customers.

Box's BoxWorks announcements are a bet that the enterprise agent race will be won by whoever controls the context, not whoever builds the flashiest agent. Context Retrieval is the most substantial piece of that bet, and it is also the piece that is furthest from shipping. The numbers Box quotes are promising and unverified, the price is unpublished, and the product that makes the whole story work arrives early next year. For Box customers on Enterprise Advanced, the Nemotron models are worth trying now. For everyone else, the honest move is to wait for general availability, a published price, and a benchmark that someone other than Box has run.

BoxWorks 2026 runs November 5-6, 2026 in San Francisco, and Box's blog carries the keynote recap for anyone who wants to watch the announcements as they were presented.

Sources

• Box, "Box Partners With NVIDIA to Deliver the Trusted Content Layer for Enterprise AI," November 5, 2026
• Box, "Box Unveils the Next Generation of Enterprise AI with New Agentic Workflow Automation and Advanced Threat Detection," November 5, 2026
• Box, "Plans and Pricing," read live November 8, 2026
• SiliconANGLE, "Box taps Nvidia's AI models to build a rich context layer for autonomous enterprise agents," November 5, 2026
• Computerworld, "Box launches Architect, an AI-powered workflow builder," November 6, 2026
• Reworked, "Box Architect Adds Multi-Model Agent Workflow Builder," November 5, 2026
• The Markets Daily, "Box Unveils AI Agent Tools and NVIDIA Partnership at BoxWorks 2026," November 8, 2026
• AI Core Updates, "Salesforce Agentforce Coworker: $50 Per User as Teams Rollout Begins," November 6, 2026
• AI Core Updates, "Google Data Agent Kit: Free MCP Tools for 15+ Data Cloud Services," October 24, 2026
• AI Core Updates, "AWS Infinity: 75,000 AI Agents on a Raspberry Pi, Zero Compute While They Wait," October 24, 2026
• AI Core Updates, "Perplexity Teammates: AI Coworkers for $200 a Month," November 6, 2026