Box announced Box Content Retrieval for AI agents at its BoxWorks 2026 developer conference on 5 November, with Nvidia executive Kari Briski expecting queries to be answered four times faster, TechCrunch reported. The system brings Nvidia’s Nemotron embedding models into Box’s enterprise content platform so agents can retrieve relevant material from company files.
Key points
- Briski expects queries to use 90% fewer tokens and produce 20% fewer errors, alongside the speed gain.
- Box is combining Nemotron embeddings with its existing file metadata and permission controls.
- TechCrunch reports an early 2027 rollout for Enterprise Advance customers. SiliconANGLE names Enterprise Advanced and Shield subscribers.
- Box also announced a workflow-building tool, Box Architect, and controls intended to flag harmful agent activity.
Nemotron indexes Box content for agents
Box Content Retrieval uses Nemotron Retriever embedding models to organise material that does not sit in a structured database, including scanned documents, images, audio and video, SiliconANGLE reported. Embeddings represent the content in a form that lets a search system find related concepts even when a query uses different words.
Box chief executive Aaron Levie told TechCrunch that the system can return the portions of documents most relevant to an agent’s task, rather than making the agent read each document. Box gave the example of lawyers searching for clauses and precedents. Its existing metadata and permission controls are part of the retrieval system alongside Nvidia’s models.
The distinction matters for the product Box is offering: Nvidia makes the Nemotron embedding models freely available, while Box supplies the content platform and its established controls over files. Levie said Box was probably combining five technologies in one API. Box says agents using Content Retrieval must follow the permissions attached to the files they search.
Nvidia has also released Auto Ontology to connect AI agents with company data, as AI Affairs previously reported. The Box announcement concerns a retrieval system within Box’s own platform, using Nvidia models alongside Box’s controls for content held there.
Box and Nvidia give retrieval targets
Kari Briski, Nvidia’s vice-president of generative AI software, told TechCrunch she expects Box queries to be answered four times faster, with 90% fewer tokens used and 20% fewer errors. Tokens are the units of text processed by an AI model, so the token figure concerns the amount of material passed through a query.
The accounts differ on the accuracy measure. TechCrunch reported Briski’s expectation of 20% fewer errors. SiliconANGLE reported Box’s claim of a 20% improvement in search accuracy. SiliconANGLE also described the token reduction as up to 90% and the response-time gain as around four times.
Box’s stated use cases depend on access to corporate material such as contracts, policy documents and customer communications. Levie said agents need that context to perform work accurately, while companies are wary of granting them broad access to their data. The new system is designed to retrieve content within the permissions already applied to Box files.
Box Architect joins the retrieval rollout
Box also announced Box Architect, which it says will let staff describe, build and test multistep agent workflows in natural language. SiliconANGLE reported that the tool can connect Box Extract, Automate, Apps and Forms. One example Box gave was a contract-review process that compares incoming agreements with approved playbooks and sends flagged clauses to an attorney.
Box says new agent controls will flag activity including bulk data transfers, large numbers of file deletions and account takeovers. Administrators will be able to restrict which files an agent can access through classification policies and limit its use of third-party tools. These controls accompany the existing file permissions used by Content Retrieval.
The reports name different customer groups for the retrieval launch. TechCrunch reported that Box Content Retrieval will roll out to Enterprise Advance customers. Box Content Retrieval is scheduled for general availability for Enterprise Advanced and Shield subscribers. SiliconANGLE said Nemotron models in Box AI and agent security controls were already available to Enterprise Advanced and Shield customers.
Box plans to launch Box Architect in beta later in 2026. Box Content Retrieval and the advanced threat-detection capabilities are scheduled for early 2027, SiliconANGLE reported.
.jpg){kind=link}