Anatomy of a document
An invoice comes in.
What does the AI do with it?
Every invoice arrives different: an electricity bill looks nothing like one from a hardware shop, or a subscription in English. On the left you have the paper as it is. On the right, the three things the model does with it —sort it, read it and pull the data out, look for mistakes— and how long each one takes on the card you pick.
Document
Action
—
And over a whole day…
One invoice is seconds. A thousand is a department.
—
The same document, four cards
Documents a day working without stopping, 24 h.
The hard case
And if the invoice is a scan or a photo?
A «proper» PDF carries the text inside: you read it and that is that. But half the offices out there work with PDFs that are a photograph of a piece of paper. There is no text to copy: the model has to look at the image, and an image costs a good deal more tokens than its text.
Qwen3.8 27B quantised to 4 bits (Q4, the UD-Q4_K_XL format, 17.55 GB) reads images, so the same card does the work without changing model or quantisation. A page scanned at reading resolution comes in at around 2,600–2,900 tokens instead of the ~1,000–1,300 of plain text: two or three times more reading. What the model writes afterwards does not change, so the hit shows up mostly on short documents.
PDF with text
Scanned PDF · vision
The calculation engine (tw-engine.js) is not there yet. Reload the page in a moment.
Where these times come from
Each time comes from counting the tokens the AI has to read and write for this invoice and dividing them by the speed of the card you picked. The RTX 3090 Ti is measured in our lab; the others are estimated with the same sum for all of them.
How every figure is worked out: Method.