Box Architect: natural-language workflow builder heads to beta

Box will put Architect, a natural-language AI workflow builder, into beta in Q4 2026 for Enterprise Advanced customers, the company announced at BoxWorks 2026 on Thursday. Pricing was not disclosed — notable because Box does not publish a price for Enterprise Advanced, its highest tier, which ships with an allocation of 20,000 AI Units to meter consumption of Box AI features.

Architect, said CEO Aaron Levie on X, "effectively acts as an agentic operator of Box's core content and workflow capabilities … to help customers design any content agentic workflow in Box."

How does Architect actually work?

The mechanism is straightforward. A user describes, in plain language, a goal or process to automate. Architect then defines the necessary stages and decisions, applying the permissions, classifications and retention policies already attached to content in Box.

Users can test the workflow before publishing it, and can insert requirements for human review and approvals. Once a workflow goes live, Architect monitors its outputs and recommends improvements as it runs continuously.

What does it plug into?

Architect orchestrates several existing Box tools rather than replacing them. Box gives a banking example:

• Collect client applications via Box Forms
• Verify documents with Box Extract
• Sign agreements with Box Sign
• Begin account setup with Box Automate

In short, one onboarding process, four Box services, one natural-language prompt to design it.

What did Box announce besides Architect?

A new Context Retrieval engine, built on Nvidia's Nemotron Retriever open model, will reach general availability for Enterprise Advanced customers early next year. Levie said it will "deliver better results for agents" connecting to documents held in Box.

CTO Ben Kus said the engine enables faster indexing and lower latency and is "designed to find and deliver more relevant enterprise context to AI agents." Three specifics matter for buyers:

• Multimodal indexing across documents, images, audio and video
• Combined keyword and semantic search, so queries work even when they lack the exact terms used in a document
• Enforcement of existing Box permissions, so agents retrieve only what a user is authorized to access

Internal testing, according to Kus, showed a 20% improvement in search accuracy, "fourfold faster question-answering," and "up to 90% fewer tokens sent to downstream models" — the last figure translating directly into lower spend on downstream model calls.

What do analysts say?

Jason Andersen, vice president and principal analyst at Moor Insights & Strategy, said Architect will make it easier to build agents and workflows based on the collective knowledge of a team or business, rather than focused only on individual productivity.

He flagged two caveats. First, customers will still need to evaluate how the workflows perform. "I'd still recommend human oversight during the creation process with any agent development project and using observability tools to make sure that the agents are running in a fashion that meets the risk and cost management needs of the business," Andersen said.

On context retrieval more broadly, Andersen said these tools are "significantly improving output accuracy and agentic performance, both of which lead to higher end user satisfaction and lower costs," and that Box's use of Nemotron's open-weight model improves the value proposition.

His second caveat: "to get the most benefit from context management, a good set of design principles and governance are also needed, so customers need to be sure to account for that effort to get the most benefits."

For CIOs, the story is a beta to watch in Q4 2026 and a retrieval engine to budget for early next year.