Short answer: On 5 November 2026 at BoxWorks, Box announced Box Architect, an AI builder that turns a plain-language goal into a governed workflow using Box Forms, Extract, Automate, Apps, Sign and agents, with a choice of models from Anthropic, OpenAI, Google, NVIDIA, IBM and others for each step. It enters beta in Q4 for Enterprise Advanced customers only, and no price has been published. Box also made agent security controls (MCP Guardrails, classification policies, threshold alerts) generally available now, and announced Box Context Retrieval on NVIDIA Nemotron for early 2027, claiming up to 20% better search accuracy, 90% fewer tokens and 4× faster answers (vendor claims).
Checked 7 November 2026 (Asia/Tehran) against Box’s press releases, BoxWorks blog post, pricing page and developer docs. Architect isn’t available yet, so we haven’t tested it.
Box used its BoxWorks 2026 conference in San Francisco on 5 November to make its biggest workflow-automation move yet. Box Architect is meant to let a process owner describe an outcome, such as “onboard a new client with complete documentation”, and have AI design and build the whole workflow inside Box. For B2B teams that already keep contracts, invoices, claims or onboarding packets in Box, this is a direct challenge to building those flows by hand in Box Automate, n8n or Zapier. Here’s what was announced, what’s available now, what it costs, and what it means if you automate with n8n.
What Box announced at BoxWorks 2026
| Announcement | What it does | Availability (per Box) | Plans |
|---|---|---|---|
| Box Architect | Designs, builds, tests and improves agentic workflows from a plain-language goal | Beta rollout in Q4 | Enterprise Advanced |
| Agent security controls | Classification-based policies for agents, MCP Guardrails on tools and actions, threshold alerts | Generally available now | Shield and Enterprise Advanced |
| Threat protection | Detects bulk exfiltration, session hijacking, account takeover, suspicious sharing, mass deletion; severity-based responses | Early next year | Enterprise Advanced and Shield |
| Box Context Retrieval | Permission-aware hybrid (keyword + semantic) search for agents on NVIDIA Nemotron Retriever; multimodal indexing | GA early next year | Enterprise Advanced |
| NVIDIA Nemotron 3 Ultra in Box AI | Open-weight model option for Box AI workflows | Available now | Enterprise Advanced |
| Box Mount | Mounts Box folders inside agent sandboxes; works with NVIDIA OpenShell and DigitalOcean Managed Agents | Coming soon | Not stated |
| Box for Industries | Packaged solutions for banking, public sector, media and entertainment | Announced | Not stated |
How Box Architect works
According to Box’s press release and its BoxWorks blog post, Architect works in three stages:
- Design. You describe a process or goal in plain language. Architect uses content and procedures already in Box to define stages, decisions and handoffs, and Box says it will redesign the process “from the ground up” rather than copy how the team works today. Every step is governed by existing permissions, classifications and retention policies.
- Test. Before go-live you run 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, for example a cheaper model for simple extraction and a stronger one for complex reasoning.
- Deploy and improve. Architect delivers the Box Apps, agents, automations, forms and document templates the process needs. After launch it monitors results and suggests improvements you can accept or refine in conversation, and admins can publish and standardise workflows across the company.
Box’s examples are all document-heavy B2B processes: contract review against a law firm’s playbook, bank client onboarding (Forms → Extract → Doc Gen and Sign → Automate), government permitting, insurance claims and HR onboarding.
Pricing and availability
Architect will roll out in beta to Enterprise Advanced customers in Q4, and Box hasn’t published a price for it. Enterprise Advanced itself has no public price on Box’s pricing page. Here’s what the page did show us on 7 November 2026 (it displayed euros from our connection; monthly / annual billing per user):
| Plan | Price per user per month | AI Units | API calls per month | File upload limit |
|---|---|---|---|---|
| Business | €18 monthly / €13.50 annual | Not listed | 50K | 5 GB |
| Business Plus | €30 / €22.50 | Not listed | 50K | 15 GB |
| Enterprise | €42 / €31.50 | 1K per month | 100K | 50 GB |
| Enterprise Plus | €45 (single price shown) | 2K per month | 100K | 150 GB |
| Enterprise Advanced | No public price (Buy now / Contact us) | Not listed | “Higher API allowances” | 500 GB |
All plans have a minimum of 3 users and unlimited storage. AI Unit and API volumes can be topped up (“additional volume available for purchase”). Annual billing is shown as “Save 25%”.
Enterprise Advanced is described as “AI-powered workflows, automation, and agents”, with multi-step execution with extended context, custom AI agents, MCP and custom-agent guardrails, Box Forms, automated document generation, agentic metadata extraction, Box Archive and agentic workflows. If you’re on Business or Enterprise today, Architect means an upgrade conversation with Box sales.
Agent security: what’s available today
The part you can use now is security. Box says these controls are generally available to Shield and Enterprise Advanced customers:
- Classification-based policies for agents: stop external agents from reaching files labelled sensitive.
- MCP Guardrails: admins limit which tools agents can use over MCP, including creating, moving and externally sharing content.
- Threshold-based alerts: notify the team when user, app or agent activity exceeds limits you set.
- Session governance and audit trails showing which agent acted and what it did.
This matters if you already connect AI agents to Box through the hosted Box MCP server at mcp.box.com, which uses OAuth and must be enabled by an admin in the Box Admin Console. MCP doesn’t bypass Box permissions, and Guardrails now add a second, tool-level limit.
Context Retrieval and NVIDIA: the vendor claims
In a separate announcement with NVIDIA, Box introduced Box Context Retrieval, a permission-aware retrieval engine built on NVIDIA Nemotron Retriever embedding models. It combines keyword and semantic search and indexes scanned documents, images, audio and video. Box says customers have “the potential” to improve search accuracy by 20% (Recall@5), cut token usage by 90% and answer questions four times faster. These are Box’s claims, not independent benchmarks, and Context Retrieval is due for general availability on Enterprise Advanced early next year. NVIDIA Nemotron models, starting with Nemotron 3 Ultra, are already available in Box AI on Enterprise Advanced.
What it means if you automate with n8n
Architect is strongest when the whole process lives in Box. Most B2B processes don’t: a signed contract still has to create a deal in the CRM, an invoice in billing and a message in Slack. A practical split:
- Inside Box: intake forms, extraction, document generation, e-signature and approvals, built in Box Automate today or Architect once you’re on Enterprise Advanced.
- Around Box: n8n’s Box Trigger fires on file or folder events, the Box node reads and moves files, and the HTTP Request node calls the Box API for anything else, so CRM, billing, email and chat steps stay in one place you control.
- Governance: if your agents reach Box over MCP, set MCP Guardrails first so an agent can’t share or delete content you didn’t intend.
Before you build, estimate run volumes with our n8n cost calculator, scan finished workflows with the n8n workflow auditor, and compare AI tool costs in our AI SaaS pricing comparison. For other agent-workflow launches this autumn, see AWS Step Functions workflows as AI agent tools, IBM Integration Intelligence Hub and ServiceNow workflows inside ChatGPT.
What we couldn’t verify
- Architect pricing and exact beta date. Box only says “beta in Q4” for Enterprise Advanced, without a day or a statement on whether Q4 is calendar or fiscal.
- Enterprise Advanced price and AI Unit allowance. Not on the pricing page. Press coverage cites 20,000 AI Units; we couldn’t confirm that on a Box page.
- US-dollar prices. Box’s pricing page showed euros from our connection; we didn’t use dollar figures. Enterprise Plus showed one price (€45) without a monthly/annual split.
- How many AI Units Architect workflows consume. Not published.
- Context Retrieval performance. Vendor claims only; not yet generally available.
- Box Mount and Box for Industries plans and dates. Described as coming soon or announced, without plan or date details.
Frequently asked questions
What is Box Architect?
Box Architect is an AI workflow builder Box announced at BoxWorks 2026 on 5 November 2026. You describe a business process or goal in plain language, and Architect designs the stages, decisions and handoffs, assembles the Box Forms, Extract, Automate, Apps, Doc Gen, Sign and agents the process needs, lets you test it on real content with human approvals, then monitors results and suggests improvements. Every step runs under existing Box permissions, classifications and retention policies.
When is Box Architect available and how much does it cost?
Box says Architect will begin rolling out in beta to Enterprise Advanced customers in Q4. Box has not published a price for Architect or for the Enterprise Advanced plan; its pricing page shows Enterprise Advanced with Buy now and Contact us buttons but no public price.
Which Box plans include AI Units?
Box’s pricing page lists 1K AI Units per month on Enterprise and 2K on Enterprise Plus, with additional volume available for purchase. Business and Business Plus include integrated Box AI but no AI Unit allowance is listed. Enterprise Advanced’s allowance isn’t listed on the pricing page.
What else did Box announce at BoxWorks 2026?
Agent security controls (classification-based policies for agents, MCP Guardrails and threshold alerts), generally available now to Shield and Enterprise Advanced customers; new threat protection for bulk exfiltration, session hijacking and mass deletion, due early next year; Box Context Retrieval powered by NVIDIA Nemotron Retriever, due early next year for Enterprise Advanced; NVIDIA Nemotron 3 Ultra in Box AI, available now on Enterprise Advanced; Box Mount (coming soon); and Box for Industries for banking, public sector and media.
Is Box Architect a replacement for n8n or Zapier?
Not really. Architect builds workflows around content stored in Box using Box’s own tools. n8n and Zapier connect many apps. Most teams will use both: Architect or Box Automate for the document steps inside Box, and n8n for the CRM, billing, email and Slack steps around them, via n8n’s Box node, Box Trigger or the Box API.
Are Box’s performance claims verified?
No. The figures for Box Context Retrieval (up to 20% better Recall@5 search accuracy, 90% lower token usage and four times faster question answering) are Box’s own claims, written as potential improvements. Box hasn’t published an independent benchmark, and Context Retrieval isn’t generally available yet.
Related reading
- Microsoft Copilot Studio persistent agents: what Copilot Credits really cost
- AWS Step Functions workflows are now AI agent tools
- IBM Integration Intelligence Hub
- New AI agent tools this week (6 November 2026)
- Zapier vs Make vs n8n AI agents
- How to secure your n8n instance
Sources
Box press release: Box Architect, threat protection and agent security controls (5 Nov 2026) (Business Wire release, republished on the Box blog); Box blog: BoxWorks 2026 overview; Box press release: Box partners with NVIDIA, Context Retrieval; Box pricing; Box MCP server docs; n8n Box node, Box Trigger and HTTP Request docs. All checked 7 November 2026 (Asia/Tehran).
