A supplier invoice passes through many hands: it arrives by email, is stored, has its data copied, is assigned to an account and is eventually paid. OCR handles only one part of that journey well: reading the document and preparing a proposal.
A safe workflow, from file to journal entry
- 01Keep the originalStore the PDF or image exactly as received. Any optimised version is a working copy, not a replacement.
- 02Extract with field-level confidenceEach field has both a proposed value and a confidence score. A clearly read total does not make an uncertain NRT correct.
- 03Check for duplicatesCompare the supplier, invoice number, date, amount and file fingerprint before creating anything in the ledger.
- 04Send exceptions for reviewInconsistent fields, tax amounts that do not add up and incomplete documents remain in a clearly visible review queue.
- 05Approve and postOnly an authorised person confirms the supplier, dates, taxes, account and reporting dimensions before the journal entry is generated.
The fields that deserve a second look
| Field | Common mistake | Verification |
|---|---|---|
| Supplier and NRT | Mistaking an address or the customer for the issuer | Match against the supplier record and country |
| Number | Dropping dashes, slashes or zeros | Uniqueness by supplier and financial year |
| Dates | Swapping the issue and due dates | Logical order and open period |
| IGI | Reading the invoice total as the tax amount | Taxable amount × rate and rounding variance |
| Currency and total | Missing a currency or discount | Sum of lines, taxes and payments |
Official information about IGI Govern d'Andorra — Impost General Indirecte
Privacy: process only what is needed
An invoice may contain names, addresses, tax identifiers, bank accounts and information about services provided. The capture system must restrict access, record who views the document and apply retention periods appropriate to its purpose.
- Do not send invoices to personal mailboxes or channels without access control.
- Keep the original document separate from thumbnails and extracted text.
- Do not use client documents to train models without a clear legal base and an explicit decision.
- Allow the complete case file to be exported, and delete temporary copies once they no longer serve a purpose.
- Record administrative access and limit it to staff who really need it.
Andorran Data Protection Agency Agència Andorrana de Protecció de Dades
How to measure whether OCR is helping
Accuracy alone is not enough. Measure how many documents reach the proposal stage, how many need correction, how long review takes and how many duplicates or inconsistencies are caught before posting. A system that reads quickly but forces you to recheck everything has not automated the process; it has merely moved where the typing happens.