Box Puts Context Retrieval, Architect in Enterprise Advanced

Box used its BoxWorks 2026 investor product briefing to position itself around AI agents that need access to company documents. The centerpiece, Box Context Retrieval, is designed to give agents relevant, up-to-date information while respecting existing permissions. CTO Ben Kus said it processes different file formats, creates text indexes and vector embeddings, and updates those indexes as files change, combining keyword-based and semantic search so agents can retrieve relevant sections rather than processing entire files. In Box's recall benchmarks, Kus said the tool improved accuracy by more than 20%, reduced token consumption by approximately 90%, and enabled agents to deliver answers four times faster, though he did not specify the comparison used.

The second pillar, Box Architect, is intended to use a customer's business goals and content to propose and build workflows using tools such as Box Automate, Box Apps and Box Extract. Kelash Kumar, vice president of product management for intelligent workflow automation, said the system is also meant to identify bottlenecks and refine processes over time. Levie described Architect as a layer above workflow platforms, comparing its intended role to that of a technical consultant or a Box forward-deployed engineer. He said Box is initially introducing the building phase, with broader capabilities coming in the months ahead.

Box also announced a partnership with NVIDIA to incorporate Nemotron retrieval models into Context Retrieval and make NVIDIA's open-weight Nemotron models available to Box Agents. Levie said the retrieval product was intended to complement other systems rather than compete directly with them, providing an option for customers that do not want to build and maintain their own retrieval infrastructure. Alongside the AI tools, Box launched Box for Industries with offerings for banking, the public sector, and media and entertainment, and announced additional protections for AI agents including classification-based access policies, guardrails on actions such as external sharing, activity-threshold alerts, and agent session governance and audit trails.