Killing PDF re-typing: how document extraction actually works

"AI reads the PDF" is the marketing sentence. The engineering underneath is more specific, and the specifics are what make it trustworthy on a document you've never seen before.

How does document extraction work? Read, classify, extract

1. Read: text-based PDFs the fast way, scanned pages through OCR.
2. Classify: terms-and-conditions, penalty tables and fee annexes are set aside, and only the data-bearing pages go forward.
3. Extract: a language model pulls the fields that matter: booking number, vessel and voyage, POL and POD, ETAs, equipment, charges and dates.

Carriers don't agree on layout, so the extractor carries line-specific rules. It knows the ETA printed near the port of loading isn't the destination ETA, that some lines pack multiple labels into one field and that a terminal suffix on a port name should resolve back to the base port.

Confidence is the honest part

Every field comes back with a confidence score. That number is what lets the system decide what to trust and what to route to a human — and it's why we won't quote a single platform-wide accuracy percentage. The right question isn't "how accurate is it", it's "how sure is it about this field, on this document, right now".

The result: no one re-types a booking confirmation from the inbox. The fields arrive validated against your masters and tariffs, and the exceptions — the genuinely ambiguous ones — are the only thing left for a person to look at. That is where a forwarder's edge from AI actually comes from.

The hard cases are the whole point

Clean, text-based PDFs are the easy part. The value is in the rest: a scanned bill of lading photographed at an angle, a booking split across a transhipment with two vessels and two ETAs, a carrier that abbreviates ports in its own private shorthand. Generic extraction falls over exactly here — which is why the pipeline carries freight-specific knowledge instead of hoping a general model guesses right.

It also knows when to stop. If a field is genuinely ambiguous, the honest move is not to invent a confident-looking answer but to say so — low confidence, route to a human. A system that knows what it doesn't know is worth far more on a real operations desk than one that's occasionally, silently wrong.