Short answer

A reliable document workflow does not start with the model that reads the PDF. It starts by identifying each document with a fingerprint, keeping the original untouched, storing alongside each extracted data point the evidence of where it came from and letting a person approve what matters. With those four pieces in place, it makes no difference whether extraction is done by rules, OCR or a language model: the result can be checked, corrected and explained.

Key points

  • A reliable document workflow rests on four pieces: a fingerprint for each document, an untouched original, evidence alongside each extracted data point and human approval of what matters.
  • A cryptographic fingerprint of the content detects duplicates even if the file name or folder changes, and proves that the approved document is the one received.
  • If reviewing a data point highlights the fragment of the PDF it comes from, review takes seconds and errors become visible.
  • It is best to start with deterministic rules, which are cheap and predictable, and reserve an AI model for the documents the rules cannot resolve.
  • When a rule is weak, the system should say so rather than treat the result as correct: an honest warning is worth more than a silent match.

The fingerprint stops you paying twice

The same document arrives through five routes: email, the supplier’s portal, a scan, forwarded by a colleague, downloaded again. If the system calculates a cryptographic fingerprint of the content, the duplicate is detected even if the file name or folder changes.

The fingerprint also prevents the same tender document or call for proposals from being processed twice, and makes it possible to prove later that the approved document is exactly the one that was received.

The original is never touched

When a document is approved, it is best to archive a copy with a standardised name (date, supplier, number) and leave the original intact. An inventory of a file server should be mounted read-only: classifying and extracting must not be able to modify or delete anything.

That immutability simplifies audits and grant reporting: the document always exists exactly as it arrived, and the standardised copy is a convenience, not the source.

Every data point with its evidence

An amount, a date or a tax ID extracted without any indication of which part of the document it came from is hard to review. If hovering over the data point highlights the fragment of the PDF it comes from, review takes seconds and errors become visible.

It is best to start with deterministic rules, which are cheap and predictable, and to reserve an AI model for the documents the rules cannot resolve. If the model runs outside the office, it should be switched off by default and enabled document by document, with that decision recorded.

Review that does not slow things down

Human review does not have to mean reading everything. It can be an exceptions inbox: whatever meets the rules goes through flagged as a proposal, and whatever does not match the expected format, or fails twice, lands in “review by hand”. Every approval records who approved it and when.

When a rule is weak, for example when matching a payslip to a bank transfer by amount and approximate date, the system should say so instead of taking the result as correct. An honest warning is worth more than a silent match.

Frequently asked questions

How do I avoid processing or paying the same invoice twice when it arrives through different channels?

By calculating a cryptographic fingerprint of each document's content: the duplicate is then detected even if the file name or folder changes. The same fingerprint prevents a tender document or call for proposals from being processed twice and makes it possible to prove later that the approved document is exactly the one received.

Should I use AI or rules to extract data from invoices and payslips?

It is best to start with deterministic rules, which are cheap and predictable, and reserve an AI model for the documents the rules cannot resolve. If the model runs outside the office, it should be switched off by default and enabled document by document, with that decision recorded. With a fingerprint, an untouched original, evidence and human review, the result can be checked and corrected whatever the extraction method.

How can data extracted from PDFs be reviewed without reading everything?

With an exceptions inbox: whatever meets the rules goes through flagged as a proposal, and whatever does not match the expected format, or fails twice, is set aside for manual review. If each data point shows the fragment of the PDF it comes from, checking takes seconds. Every approval records who made it and when.

Why keep the original document if we already have the extracted data?

Because an untouched original simplifies audits and grant reporting: the document always exists exactly as it arrived. On approval, a copy can be archived with a standardised name (date, supplier, number), but that copy is a convenience, not the source. For the same reason, an inventory of a file server should be mounted read-only.