New AI Team Sorts Secret Documents 54% Cheaper Than Before

Governments, banks, power plants and other critical operations handle mountains of paperwork, and someone has to decide which files are sensitive before they get stored or passed along. Today that job is done by people, which is slow, expensive and inconsistent — two reviewers can look at the same document and reach different conclusions. Researchers in Pakistan have trained an AI system to do this screening automatically, and they say it beats the previous best approach.

The clever part is how it reads. Most AI models can only hold a limited amount of text in mind at once — like reading only the top half of a long letter and never seeing the signature or the closing paragraph. Since the most sensitive cables are often the longest, that blind spot matters. Their fix: several AI 'agents' (programs that act somewhat independently) each read a different section, then hold short back-and-forth discussions to reconcile what they found. One acts as a critic, deciding which reader to trust more on a given document.

The results: about 90.7% accuracy and 92% of sensitive documents correctly caught, up from 89.1% for the single-model version. Just as notable, the new system used roughly 54% less average computing than a comparable fixed-round approach — meaning lower costs and faster turnaround. The researchers also included tools that show which words drove each decision, so a human reviewer can check the AI's reasoning rather than trusting a black box.

The honest caveats: this was tested on a specific 16,000-document corpus, not on the messy reality of every government archive, and an error still means a sensitive file gets misfiled or a harmless one gets flagged for extra review. The authors note the accuracy gain is real but modest. Expect this kind of 'AI review team' to show up first in compliance, legal discovery and records management — jobs built on reading and sorting paperwork.