Box, which many companies use to store their contracts, invoices and reports, has built a new retrieval feature for the AI agents that work on those files. It uses open models from Nvidia to index the files, and it is due for customers on Box's Enterprise Advanced plan in early 2027. The interesting part is not the partnership. It is what the feature shows about where AI agents succeed or fail on a company's own documents.
An agent only knows what the search hands it
An AI agent does not carry your files in its head. When someone asks a question, a search step picks some text, and the model answers from that text alone. If the search picks the wrong passages, the model writes a smooth, confident answer from the wrong material, and nothing in the answer shows it.
That is why the 1985 lawsuit review in the lead matters beyond law offices. Every company with a shared drive has the same weak point, and a better model does not fix it.
How the search finds meaning instead of words
Nvidia's Nemotron models read each passage and turn it into a long list of numbers that captures what the passage is about. Two passages about the same thing get similar numbers even when their words differ. When a question comes in, it gets the same treatment, and the system sends the agent the passages whose numbers are closest.
So a lawyer who asks about ending an agreement early can be shown the clause titled "termination for convenience", and the same works across languages. Box says the index also covers scans, images, audio and video, so a photographed invoice or a recorded meeting can be found like a typed document. The search is also permission-aware: an agent working for you is only handed passages you are allowed to open.
Reading less is where the saving comes from
Agents are charged by how much text they read, counted in tokens. Suppose an operations manager asks which of 200 supplier contracts let the supplier raise prices on short notice. A 60-page contract is roughly 40,000 tokens, so an agent that reads every contract works through about 8 million tokens. An agent handed the 20 best passages, at around 500 tokens each, works through 10,000.
Nvidia expects the new system to use 90% fewer tokens on average, answer four times faster and make 20% fewer errors. Questions that span many files save the most, and those are the ones people rarely ask an agent today because reading everything is too slow and too expensive. When the cost falls that far, "check all the contracts" becomes an ordinary question instead of a project.
Why a free download is not enough
The Nemotron models can be downloaded for free, yet Box's chief executive, Aaron Levie, says the product packs in "five different technologies into one simple API". What makes it hard to copy is that it leans on the labels and access rules Box already holds for every file: who owns it, what type it is, who may see it.
That has a practical meaning for any small business. An agent searching a drive of well-named files with clear access rules has far more to work with than one searching a heap of documents called "final_v3". The filing work people did for years, and complained about, is what a search agent uses.
What happened to lawyers after 1985
The lawyers in that study went on to spend decades with a weak tool. In 2012 a New York federal judge became the first to approve software that predicts which documents are relevant to a case. After that, lawyers stopped reading everything: software chose the pile, and people checked samples of it.
The same split is coming for anyone who goes through files for a living, in finance, HR, purchasing or compliance. The software will pick the passages and the person will decide what they mean. The skill that matters most then is checking the picking. Choose ten questions whose answers you already know from your own files, word them differently from the files, and see which passages the agent finds. If it finds nine, you can trust it with the tenth question you do not know the answer to.