BoxWorks 2026: Box is the context layer for AI

Blog on The Ravit Show Website

I was in the front row at BoxWorks 2026 in San Francisco this morning. Aaron Levie opened the keynote, and the message was clear from the first few minutes.

Agents are only as useful as the context they can reach. Box is that context layer.

Box made 4 big announcements today. Box Context Retrieval, powered by a deeper partnership with NVIDIA. A new agentic workflow builder called Box Architect,Box for Industries, starting with Banking, Public Sector and Media & Entertainment, and threat protection for agents.

Here is everything that was announced, and what I think it means.

1. Box and NVIDIA: Box Context Retrieval

The headline is Box Context Retrieval, a new engine powered by NVIDIA Nemotron Retriever models.

The problem it solves is simple. Your company's contracts, policies, customer history and institutional knowledge all live in files. Agents need that context to do real work. Most of it is hard for AI to read well, especially scanned documents, images, audio and video.

Box Context Retrieval turns that unstructured data into a structured knowledge layer that agents can reason over. And it respects permissions. An agent only pulls from content the user is already allowed to see.

Box shared 3 numbers:

Up to 20% better search accuracy (Recall@5)
Up to 90% lower token usage
Up to 4x faster question answering

It also brings multimodal indexing and hybrid search. That means precise keyword matching plus an understanding of meaning, even when you don't know the exact words to search for.

2. NVIDIA open weight models inside Box AI

The second part of the NVIDIA news is model choice. NVIDIA Nemotron models, starting with Nemotron 3 Ultra, are now available in Box AI for Enterprise Advanced customers.

Box has always taken a multi-model approach. Open weight models give customers a cheaper option for high-volume, specialized work, and the content still stays secure in Box.

Aaron said open weight AI complements closed models and drives broader adoption across the industry. I agree. Most enterprises I talk to don't want to bet on 1 model. They want to route each task to the right one based on cost and performance.

Box gave a few examples of where this lands:

Financial analysts asking questions about loan applications, with answers grounded only in content they can access
Legal teams finding clauses and precedents across contracts, even when the wording differs
Clinical researchers working off the latest policies and care protocols in real time
Insurance auditors analyzing full claim packages across scanned forms, photos, recorded statements and video
Federal employees running multi-step research workflows across policies and case files

3. Box Architect: agentic workflows in plain language

This was the announcement I was most curious about.

Box Architect lets teams build, manage and improve agentic workflows by describing a process in plain language. It uses the content and procedures already in Box to define stages, decisions and handoffs.

It works in 3 steps:

Design: Describe the process or goal. Box Architect maps out the workflow, governed by existing permissions, classifications and retention policies.
Test: Run it on real content and edge cases. Preview agent outputs, add human approvals at key points, and pick a model for each step.
Deploy: Ship the full workflow with Box Apps, agents, automations, forms and templates. Box Architect monitors results and suggests improvements.

You can choose models from Anthropic, OpenAI, Google, NVIDIA, IBM and more for each step. A cheaper model for simple steps, a stronger one for complex reasoning.

Aaron pointed to recent frontier models, naming OpenAI's Astra and Anthropic's Fable, as the reason this is possible now.

The use cases are practical. A law firm building a contract review flow that learns its own playbook from past agreements. A bank turning client onboarding into 1 connected workflow, from forms to signatures to account setup. A government agency checking permit applications for completeness. Insurance claims. HR onboarding.

Box Architect rolls out in beta to Enterprise Advanced customers in Q4.

4. Box for Industries

Olivia Nottebohm, COO of Box, put it well in the release. Every industry runs on content, but no 2 industries work the same way.

Box for Industries packages Box's security, AI and workflow tools for specific sectors. It starts with 3:

Box for Banking: Account opening, client servicing, deal management and lending are just a few examples. Box AI extracts details, summarizes files and flags missing documents.
Box for Public Sector: Case management, grants, permitting, digital evidence, public records requests, and more. Backed by FedRAMP High, GovRAMP High and DoD IL4.
Box for Media & Entertainment: Script management, contract management, frame-by-frame video reviews, and more.

The customer stories stood out. The Norinchukin Bank cut contract review time by nearly 88%. The Texas Department of Motor Vehicles is using Box as a secure base for AI. HBO uses Box from the writers' room to final delivery.

Legal, Insurance, Life Sciences, Education and Retail are coming next.

5. Security for the agent era

This part matters more than it sounds.

As agents start to read, move and share content, the attack surface grows. Box is building new threat protection to catch attacks like bulk exfiltration, session hijacking, account takeover and mass deletion.

What's coming:

Faster detection of unusual downloads, with automatic restrictions
Session binding and device-bound credentials so stolen tokens can't be replayed
Alerts on suspicious sharing and mass deletion
Automatic responses based on severity, from blocking downloads to ending a session

What's available today are agent security controls:

Classification-based policies that limit which content agents can reach
MCP Guardrails that let admins limit what agents can do, like creating, moving or sharing content externally
Threshold alerts when agent activity crosses limits you set

My take

I have covered a lot of AI keynotes this year. Most talk about models. Box talked about context, workflows and control, and I think that's the right order.

Here is what stood out to me.

Context is the real bottleneck. Every enterprise leader I interview says the same thing. Their AI pilots work in a demo and stall in production because the agent doesn't know the business. Box holds files that contain that knowledge. Box Context Retrieval is a smart bet on that position.

Permissions are the quiet win. Permission-aware retrieval doesn't make headlines. It's what lets a bank or a government agency actually say yes to agents.

Model choice is now table stakes. Bringing in NVIDIA's open weight models next to Anthropic, OpenAI and Google tells me Box isn't picking a winner. Customers want to route by cost and task, and Box is making that easy.

Box Architect is the one to watch. Building agentic workflows still takes too much specialist effort. If Box Architect really lets a business team describe a process and ship it with human checkpoints, that changes who gets to build with AI. I want to see it in production before I call it, but the direction is right.

Security has to move with agents. Giving agents access without guardrails is how you get a breach. I like that Box expanded its advanced agent controls today, ahead of the full threat protection suite.

From my interviews at BoxWorks

I also sat down with Box leaders for The Ravit Show:

Ben Kus, CTO, Box on NVIDIA, Context Retrieval and Box Architect
Heather Ceylan, CISO, Box on agentic threats and agent security controls
Olivia Nottebohm, COO, Box and Vaibhav Shinde, Deloitte on Box for Industries and moving AI from pilot to production

Full interviews are coming to YouTube and LinkedIn over the next few weeks. Subscribe so you don't miss them.

Learn more about everything announced at BoxWorks 2026

#boxworks #boxpartner