Service
Unstructured Document Extraction
Commercial invoices, Bills of Lading, packing lists, arrival notices. They arrive as PDFs and email attachments, in a different layout from every counterparty, and somebody retypes them into your system. We build pipelines that read them instead.
Why document parsing used to be a losing battle
Traditional extraction meant template matching: pin the invoice number to a fixed position on the page, and hope every supplier uses the same layout forever. They do not. Every new counterparty meant a new template, and any redesign broke the old one silently.
Language models changed the economics of this. A model can read a document the way a person does — finding the invoice number because it is labelled as one, not because it sits at particular coordinates. That generalises across layouts you have never seen, which is precisely where template matching failed.
What has not changed is that a model will sometimes be confident and wrong. The engineering that matters is no longer extraction; it is verification.
Extraction is the easy half. Verification is the job.
A pipeline that returns a number is useless if nobody can tell whether the number is right. Every extraction we deliver carries the checks needed to trust it — or to route it to a person when it should not be trusted automatically.
- Schema validation
- Every field is typed and constrained before it leaves the pipeline. A date that is not a date, or a container number that fails its check digit, never reaches your system as if it were valid.
- Arithmetic reconciliation
- On invoices, line items should sum to the subtotal and the totals should agree. When they do not, something was misread, and that is detectable without a human looking.
- Confidence scoring
- Fields the model was unsure about are flagged rather than silently accepted, so review effort goes where the risk actually is.
- Cross-referencing
- Where you already hold the ground truth — a purchase order, a booking reference — extracted values are checked against it instead of being trusted on their own.
- Human-in-the-loop review
- Anything that fails a check goes to a review queue with the source document alongside the extracted values. Corrections are quick because the reviewer is confirming, not retyping.
Scanned documents and the OCR layer
Not every PDF contains text. Scans, photographs of paperwork, and faxes that have been through three systems arrive as images, and they need optical character recognition before anything can read them.
OCR quality sets a ceiling on everything downstream, so it is worth measuring honestly on your actual documents rather than assuming. A clean digital PDF and a phone photograph of a crumpled delivery note are not the same problem, and a pipeline that treats them identically will disappoint on one of them.
Where document quality is genuinely poor, the right answer is sometimes to fix the intake — asking a counterparty to send the digital original — rather than to engineer around a bad scan forever.
Where the documents come from
Most document workflows start in an inbox. A pipeline can watch a dedicated mailbox, pick up attachments as they arrive, and process them without anyone forwarding anything. Others start in a shared drive, an FTP drop, or as the output of a portal scraper.
Whichever the source, the useful property is the same: the document is processed when it arrives rather than when someone gets to it. Extraction on receipt is what turns a two-day lag into a two-minute one.
Common questions
Frequently asked
How accurate is AI-based document extraction?
Accuracy depends far more on document quality and field type than on the model. Clean digital PDFs with clearly labelled fields extract very reliably; poor scans and ambiguous free-text fields do not. Rather than quote a number that would not survive contact with your documents, we measure it on a sample of your real files during the prototype and show you the result.
What document types can you handle?
Commercial invoices, Bills of Lading, packing lists, arrival notices, delivery notes, customs paperwork, and most structured business documents. The approach generalises to layouts it has not seen, so a new supplier's invoice format does not require new development.
Do we still need someone checking the output?
For anything financial or legally consequential, yes — but only on the exceptions. Confidence scoring and validation rules mean a reviewer looks at the small share of documents that failed a check, not all of them. That is usually the difference between a full-time task and a few minutes a day.
Can this read handwriting?
Sometimes, and unreliably. Printed text is a solved problem; handwriting varies enormously in legibility. If handwritten fields are central to your workflow, we would want to test on real samples before committing to an approach.
Where does our document data go?
Nowhere except your systems. Documents are processed and the structured output is written wherever you specify — your database, ERP, or spreadsheet. We will walk through the data handling in detail before anything is built, because for commercial documents that usually matters.
Have a workflow like this to automate?
Send us the portal, the document, or the spreadsheet you are tired of filling in by hand. We will tell you within a day whether it is automatable — and what it takes.