- What AI OCR makes easier and what stays
- Six steps from inbox to archive
- Why to capture only about five fields
- Three reasons the human check stays
The short answer
What AI OCR removes from invoice processing is re-typing and searching, not judgement. Copying supplier names, dates and amounts into another system gets shorter, and finding one invoice months later gets much shorter. The numbers that drive payment still go past a person. Capture quality depends on how each invoice is laid out, and a supplier who changes their format changes your results with it.
What actually gets easier
Say fifty invoices arrive at the end of the month. Until now the sequence was: open the email, save the attachment, rename it, type the supplier and amount into the accounting system — fifty times.
With OCR in the middle, saving, naming and the groundwork for the transfer move to the machine. The person's job changes from typing to checking. Work does not disappear; it changes kind. That is the honest description.
The gain depends less on how many invoices arrive than on how many different layouts they come in. Twenty suppliers with twenty formats is a different problem from twenty suppliers with three.
From inbox to archive
1. Receive the email
Point invoice delivery at one shared address. If invoices land in an individual's mailbox, processing stops on the day that person is out.
2. Save the PDF
Store the attachment in a fixed location. Leave the original filename alone at this stage — naming comes after you know what is inside. Paper invoices get scanned here and join the same folder.
3. Run OCR
Turn the contents into text. A scanned PDF is an image, so without this step nothing in it can be searched or copied. Keep the extracted text alongside the original rather than replacing the original file.
4. Pull out supplier, date and amount
Few fields actually get used later. Supplier name, invoice date, due date, total, invoice number. Those five are enough for finding documents and for reconciliation. Trying to capture every field adds checking work that the extra data does not repay.
5. Let someone check it
Show the extracted fields next to the PDF itself. Corrections made here become a reference point the next time an invoice arrives from the same supplier.
6. File it
Only checked items move to the archive. Storing both the filename and the extracted fields means people can find an invoice without remembering what it was called.
Why the checking step stays
Three reasons.
First, errors here are errors in money. Read the subtotal where the total belongs, push it through to payment, and you find out when the remittance fails to reconcile. Investigating a discrepancy takes far longer than checking would have.
Second, layouts change. When a supplier replaces their billing system, the invoice design changes without notice, and a field that had been captured correctly until then starts picking up the box next to it. A person looking at the results notices in the same month.
Third, an invoice is not only something to be read. The arithmetic can be right while the invoice still lists goods nobody ordered, a unit price nobody agreed, or a period that does not match the contract. Only the person responsible can make that call. The check is not a patch over OCR's limits; it is part of the job in the first place.
Skipping the check looks faster, and it is — in the months when nothing goes wrong. Counting the months when something does, keeping the check is the shorter route on average.
How this looks in practice
Suppose a food wholesaler with about fifteen staff receives invoices split between email and post, with one person handling accounts. The first change would be a single receiving address, with posted invoices scanned the day they arrive.
This company does not reconcile by invoice number, so it drops that one from the five above and captures four fields: supplier, invoice date, due date and total. The person reviews them on screen and corrects what needs correcting. If two days at month end go on re-typing, this can turn that time into review and exceptions. Amounts picked from the wrong field also become easier to catch and correct at the review stage.
The other thing that can change is answering questions. When someone asks what a particular March invoice came to, filtering by supplier and month can find it, instead of starting the search in someone's mailbox.
Where to be careful
If automation continues all the way into generating payment files, the cost of a skipped check rises sharply. Keep at least one point between capture and payment where a person approves.
Invoices received as electronic data are subject to rules about how they must be stored. We do not cover retention periods or the detail of those requirements here; check the guidance published by the National Tax Agency.
Finally, do not overwrite the original PDF with the OCR output. When you need to re-run capture, or when a supplier's format changes, you need the original to compare against. Keep the original as it arrived and hold the extracted fields separately.
In short
- Re-typing and searching shrink; judgement stays.
- Capturing around five fields keeps the process sustainable.
- The human check is what catches format changes.
- Never replace the original PDF with its captured text.
APOUL OCR Reader