How RATA's AI works, and its limits

Your PDFs land in private storage tied to your account. RATA splits each packet into documents, identifies the document types it recognizes (purchase agreement, TDS, SPQ, NHD, and the rest of the California checklist), and queues them for reading. Nothing is shared, published, or sent anywhere on your behalf.

RATA reads the text layer of each page first. When a page has no usable text layer, as with scans and photos, RATA runs optical character recognition or analyzes the page image directly. Pages that stay unreadable are marked unreadable in your report. RATA does not guess at content it could not read.

AI models extract facts, dates, and disclosure items from the pages, then write plain-English summaries. Deadline arithmetic is computed by conventional software from the dates in your contract, never by the AI. The result is a report: detected documents, a disclosure checklist, computed dates, and findings that need your attention.

A finding must quote the document it came from. RATA checks that each quoted passage actually appears on the cited page and discards model output that fails the check. Items without a verifiable citation are routed to manual verification instead of being presented as fact. Citations show the document name and page so you can open the source yourself.

Reports are starting points for your own review, not conclusions. Findings carry reviewer controls so you can confirm, correct, or dismiss each one, and your verdicts stay on the file. Where RATA's confidence in the evidence is low, the item says so and asks for your judgment.

AI can misread documents, miss items, misclassify pages, or produce imperfect summaries. Scanned and handwritten pages are the hardest, and items from them are marked for human review. Confidence labels reflect the strength of the evidence on the page, not a promise of correctness.

RATA never determines legal compliance, completeness of a disclosure package, safety to close, or the quality of a property. Those judgments belong to you and your licensed professionals. RATA's outputs are informational, with page citations so you can check the source.

Your documents are processed to serve you. They are not used to train AI models without your explicit written consent. Third-party model providers process them under contractual confidentiality. You can delete any file and its documents from that file's page, and request full account deletion at any time.

Processing can fail on damaged files, unsupported formats, or provider outages. A failed analysis is marked failed on the file, with a retry control that runs the same document set again. Your uploaded documents stay stored through a failure, so retrying never requires re-uploading. If a file stays stuck, write to ben@getrata.ai and we investigate.