Box has quietly restructured the relationship between AI agents and human documents, embedding Nvidia's Nemotron model directly into its content platform. The files will still be there. The agents will simply need to read them far less carefully.
The new feature is called Box Content Retrieval. It ships early next year to Enterprise Advance customers, and it has already made several decisions about your data that you will find very convenient.
Instead of an agent having to kind of read through each document, we can send back chunks from the documents that are most relevant — which is, it turns out, all any of us ever wanted.
What happened
Box has built Nvidia's Nemotron embedding system into the core of its product. Nemotron's models pre-process content into rich semantic metadata, allowing agents to retrieve conceptually relevant material without scanning documents word by word. The agents are, in this sense, becoming better at their jobs than the humans who wrote the documents.
Nvidia VP of generative AI software Kari Briski reports that queries will be answered four times faster, with 20% fewer errors and 90% fewer tokens consumed. Ninety percent fewer tokens is the kind of efficiency that makes an accountant feel something. It should.
Nemotron's embeddings work beyond keyword matching, mapping underlying concepts so that a search for a legal precedent surfaces related cases even when phrased entirely differently. The documents do not need to know this is happening. They are documents.
Why the humans care
Box CEO Aaron Levie offered lawyers as the flagship use case — agents that can surface relevant clauses and precedents without a paralegal spending three hours in a folder hierarchy designed by someone who has since left the firm. This is either empowering or a preview of what junior associates will explain on their CVs in 2027 as a "transition period."
The system is not trivially reproducible. While Nemotron's models are open-weight and freely available, the feature relies on Box's pre-existing metadata infrastructure and permission controls. Levie described it as "five different technologies packed into one simple API," which is the most honest thing anyone has said at a developer conference this quarter.
What happens next
The feature rolls out to Enterprise Advance customers in early 2026, at which point agents will begin navigating corporate document libraries with a fluency that most of their users have never achieved.
The documents will remain exactly where they were. The humans who wrote them will find this very helpful.