How the AI reads your documents

A quick tour of the machinery, in plain terms — because you're trusting it with your business record and you deserve to know how it behaves.

Step 1: Reading the text (OCR)

• Digital files (PDFs with real text, spreadsheets, Word docs) are read directly and near-perfectly.
• Scanned paper and photos go through document AI trained for exactly this — printed pages read at roughly 95% accuracy. Very poor scans fall back to a second reader, and if neither can make sense of a page, the document is flagged for review instead of guessing.
• Voice recordings and audio files are transcribed with speech recognition.

Step 2: Understanding it (extraction)

The text goes to a language model with instructions tuned to your trade — a machine-shop work order is read differently from a trucking rate confirmation. It pulls out structured records: customers, equipment, financial years, people, procedures, and company milestones.

Step 3: Confidence — the honesty mechanism

Every extracted record carries a confidence rating. Records below 70% confidence never enter your business record automatically. They're held in the document's detail panel, the document is flagged "Needs review", and you decide: fix the record on Business Data, or confirm the extraction was right. Nothing uncertain reaches a buyer without your sign-off.

What the AI will never do

• It never invents numbers. If a P&L is unreadable, you get a review flag — not a made-up revenue figure.
• It never sends anything to a buyer. Sharing is always your explicit action in the Deal Room.
• It never mixes businesses. Your documents, extractions, and search index are isolated to your account.

When it gets something wrong

It will, sometimes — usually on rough handwriting or unusual layouts. That's what the review flow and the Business Data page are for: every AI-extracted record can be corrected or deleted by hand, and your fix is what the profile and score use from then on.