How document AI actually reads your files

"Document AI" gets used as one phrase for what's actually three or four distinct problems stacked on top of each other. When people ask how it works, or why it sometimes gets things wrong, it helps to separate the stages, because each one fails in a different way.

Stage one: turning pixels into text

Optical character recognition is the oldest piece of this stack and, by now, the most reliable for clean, typed documents. Modern OCR models handle a scanned invoice or a photographed receipt well enough that this stage is rarely where things break down anymore. Where it still struggles: handwriting, low-resolution phone photos taken at an angle, and documents with unusual fonts or heavy stamps and watermarks overlapping the text. If an extraction pipeline is producing garbage output, this early stage is worth checking first, since every later stage inherits its mistakes.

Stage two: understanding layout, not just characters

Reading the words is not the same as understanding a document's structure. A layout model has to figure out that this block of text is a line item in a table, that number over there is a total, and this stamp in the corner is a signature block rather than body text. This is where most of the real progress in the last few years has happened — models trained specifically on document structure, not just general text, that can tell the difference between a header, a footnote, and a table cell even when the visual formatting is inconsistent across different vendors' invoice templates.

Stage three: classification and extraction

Once a document's structure is understood, the system decides what kind of document it is — an invoice, a purchase order, a signed contract — and pulls specific fields out of it: vendor name, invoice number, due date, total amount. This is the stage most people actually care about, and it's also the stage where confidence scores matter most. A good extraction pipeline doesn't just guess; it tells you how sure it is, and routes anything below a threshold to a human for a quick check rather than silently guessing wrong.

Where it still needs a human

We're honest with customers about this: fully unsupervised extraction on messy, varied documents — think hand-annotated PDFs, faxed forms, inconsistent vendor templates — still benefits from a human review step for anything financially or legally consequential. The technology is good at reducing the review from "read the whole document" to "confirm these four highlighted fields," which is a real time savings, just not the fully hands-off magic some vendors imply. We'd rather set that expectation correctly than have you find out the hard way.