New AI Reads Messy Financial Reports 50 Times Cheaper

Think about the documents that run the business world: 10-K financial filings, insurance policies, scientific reports, equipment manuals. They're not clean paragraphs. They're text tangled with tables, charts, footnotes, and layouts where position on the page actually means something. Asking an AI "what was Q3 revenue?" or "what's the warranty period?" is surprisingly hard, because the answer often lives in a table or a heading, not a sentence.

The usual fix is to have a large language model — the same kind of AI behind ChatGPT — repeatedly read through the document to figure out its structure and hunt for answers. That works, but every pass costs money and adds waiting time. The cheaper alternative is to flatten everything into plain text, which saves money but throws away the layout, so answers get worse.

EidosDoc takes a third path with three tricks. First, it learns the shape of a document on its own — which block is a header, where a table sits, what belongs to what — and stores that as a compact mathematical fingerprint rather than hand-written rules. Second, it searches using old-fashioned keyword matching, layout fingerprints, and a small, cheap helper model, without calling a big AI at all. Third, once it finds good evidence, it automatically pulls in nearby and related pieces, so it doesn't miss an answer split across a page break. Tested on four standard academic tests, it reportedly beats the previous best while cutting cost 50-fold and wait times four-fold.

The catch is real: this is a published paper, not an app on your phone. The tests are academic, and real-world documents are messier — scanned pages, handwriting, or wildly inconsistent formatting may still trip it up. But the direction matters. Banks, hospitals, insurers, and law firms spend enormous sums having AI read paperwork. A 50x cost cut could put that power within reach of small clinics and boutique law offices, not just the Fortune 500.