Box Builds Agent-First Content Retrieval on NVIDIA Nemotron and Claims 90 Percent Fewer Tokens and 4x Faster Answers

Content-cloud company Box has announced Content Retrieval, a new retrieval layer designed for AI agents rather than for human search, built on NVIDIA's Nemotron embedding technology. The feature is due to reach Enterprise Advance customers in early 2027, according to TechCrunch's coverage of the launch.

The pitch is aimed at a specific pain in enterprise agent deployments. When an agent is asked a question about a company's documents, the naive approach is to feed it entire files, which burns tokens, slows responses and drags in irrelevant material. Content Retrieval instead surfaces the relevant passages from across a customer's Box content, letting the agent work from what it needs rather than what happens to be in the folder.

Box claims the system returns answers four times faster, with 20% fewer errors and 90% fewer tokens than the document-stuffing baseline. Those figures come from the company's own testing, not from an independent benchmark, and their magnitude will depend heavily on how well any given customer's content is organized. Box's own materials note that the feature depends on the platform's metadata and permission controls as much as on the underlying models.

That dependency is the honest caveat in the announcement. Enterprise content repositories are notoriously messy: mislabeled files, stale versions and inconsistent naming all degrade retrieval quality, and permission structures determine whether an agent can even see the right material. A retrieval system is only as good as the metadata it searches over, which means customers with poor content hygiene may not see anything close to the headline numbers.

The NVIDIA angle matters too. Nemotron is NVIDIA's family of enterprise models, and its embedding technology is becoming a common substrate for retrieval products. For Box, building on Nemotron ties the feature to the infrastructure stack that most of its enterprise customers already buy from, rather than requiring a bespoke model investment.

The broader lesson for enterprises evaluating agent platforms is that better retrieval is often a more practical investment than a bigger model. The number to measure on your own material is not the vendor's speed claim but the full task cost: whether the agent finds the right clause, keeps the source reference, respects document permissions, and finishes the job for less money than the current baseline.