AI Systems
Document intelligence: first, know what you're holding
A document is data that arrived as a picture
Most business paperwork carries a handful of facts that belong in a system: a supplier, a total, a permit number, a serial. Getting them there means someone opening the file, deciding what it is, finding the facts and typing them in. It's slow, and the mistakes are quiet.
Document intelligence turns that into a pipeline. It reads the text, using OCR only when a file has none; recognises the type from a fixed list, with a confidence; extracts the fields that type needs; checks them against each other and your records; and hands the result to the workflow that owns it. When it isn't sure, a person decides.
Four documents, one pipeline
inv_4471.pdf
- arrived
- email to accounts, PDF with a text layer
- read as
- text and table layout
- recognised as
- supplier invoice · 0.98
- extracted
- supplier, invoice number, date, PO-114, 6 lines, total $1,286.40
- checked
- lines add up to the total ✓ · PO-114 exists ✓
- goes to
- accounts payable, matched to PO-114
scan_0193.jpg
- arrived
- scanned at the office, no text layer
- read as
- OCR, then layout
- recognised as
- permit approval · 0.95
- extracted
- permit number, address, approval date, inspector
- checked
- address matches job J-436 ✓
- goes to
- permits tracker; J-436 cleared to schedule
IMG_2210.jpg
- arrived
- photo from T5's phone
- read as
- image, text on the plate
- recognised as
- equipment nameplate · 0.91
- extracted
- brand, model, serial, manufacture date
- checked
- serial format ✓ · date in the past ✓
- goes to
- equipment record for the job's address; warranty check
document.pdf
- arrived
- PDF in a customer email
- read as
- text layer
- recognised as
- unknown · 0.42
- extracted
- nothing yet
- checked
- not applicable
- goes to
- the office lead's queue, with the email for context
Six stages, each one checked
01
Parse
What's on the page?
Text, tables and layout, read directly from files that have a text layer. Structure is kept, because a total means nothing without its table.
02
OCR
What if there's no text?
Scans and photos are read as images, only when needed. Printed text reads well; handwriting gets lower confidence.
03
Classify
What is it?
A type from your fixed list, with a confidence, or “unknown”. Each type has its own fields and its own destination.
04
Extract
What does it say?
The fields that type needs, each with the place on the page it came from.
05
Validate
Does it hold together?
Lines add up, references exist, dates make sense. A document that fails a check isn't passed on as if it were fine.
06
Hand off
Where does it go?
To the workflow that owns that type, with the original attached. Anything uncertain goes to a person.
What document intelligence hands off
In
Out
Worth building when the same paperwork arrives every day
Good fit
- Documents arrive daily in a handful of recurring types.
- Someone currently opens each one and retypes what matters.
- The fields feed a system: accounting, scheduling, records.
Not the right fit
- Every document is different and needs expert reading, such as contracts under negotiation.
- A few documents a week. Retyping is cheaper than the pipeline.
Questions about document intelligence
Do all our documents need OCR?
No. Files with a text layer, like most PDFs from accounting systems, are read directly, which is more accurate. OCR is for scans and photos.
Can it read handwriting?
Often, but less reliably than print. Handwritten fields get lower confidence and go to a person until tests on your own forms show they can be trusted.
How accurate is it?
It depends on your documents. We measure it on a sample of your own paperwork that a person has checked, field by field, and only let a type go straight through once it's proven.
What tools do you use?
Often Docling for parsing, with OCR and a model where the documents need them. These are starting candidates, not requirements.