data.day
ocr / automation / software selection / finops

OCR Demos Are Theater: Test It With Ugly Invoices, Not Perfect Ones

Sales demos use perfect PDFs. Real life uses crumpled coffee receipts. Here is how to stress-test an OCR tool before you sign the contract.

Klara Budget Klara Budget

If I Have to Fix It, It Is Not Automatic

You are about to spend €5,000 a year on an Accounts Payable automation tool. The sales rep promises “Touchless Processing.”

This is a lie.

In the demo, the invoice is always a clean PDF from “Acme Corp” with one line item and a clear Total. In reality, your vendors send you photos taken in a dark room, bills with handwritten adjustments, and invoices where the VAT is hidden in a footer font size 6.

If your “smart” tool cannot read these, you are paying a monthly subscription for the privilege of correcting its mistakes. That is not an asset; that is a chore.

The Hidden Cost: The “Review Tax”

The cost of bad OCR (Optical Character Recognition) is insidious. If the tool gets the Vendor Name right but the IBAN wrong, you send money to the void. If the tool gets the Total right but misses the VAT breakdown, you lose 21% of the value on every transaction.

But the biggest cost is Trust. Once your team realizes the tool makes mistakes, they have to audit every single scan.

  • Time to manually type an invoice: 60 seconds.
  • Time to upload, wait for OCR, review the data, spot the error, and fix the field: 90 seconds.

You just spent money to make your team slower. Meten is weten.

[Image of comparison chart. Left bar: “Manual Entry (60s)”. Right bar: “Bad OCR + Correction (90s)”. Label the difference “The Idiot Tax”.]

The ROI: The “Ugly Invoice” Test Pack

Before you sign any contract, demand a 7-day trial. Do not use their sample data. Use yours.

Create a folder called “The Torture Test”. It must contain:

  1. The Crumpled Receipt: A lunch receipt that lived in a pocket for three days. Can it read the date?
  2. The Mixed Tax Bill: An invoice with 0% VAT items (flights) and 21% VAT items (booking fees). Does it split them into two GL lines?
  3. The Foreign Language: A Spanish taxi receipt. Does it recognize “IVA” is VAT?
  4. The Handwriting: An invoice where the vendor crossed out the total and wrote a new one in pen. (This happens).

Upload these 4 files.

  • 0/4 Success: Do not buy.
  • 2/4 Success: Negotiate a 50% discount.
  • 4/4 Success: Buy it immediately.

Conclusion

Automation is binary. It either works, or it creates work. Don’t be impressed by the shiny dashboard. Be impressed by the robot that can read a coffee-stained receipt.

FAQs

Why do OCR tools fail on complex invoices?

Because they are guessing based on layout. If the layout shifts or the ink is faint, the guess is wrong. A 99% accuracy claim usually excludes 'handwritten' or 'low quality' inputs.

What is the most common OCR error?

The VAT split. It sees the Total (€121) but misses the Tax (€21). If you don't catch this, you overstate your expense and underclaim your tax refund.

Should I accept 80% accuracy?

No. If you have to check every single line, the tool is useless. It is faster to type it from scratch than to proofread a hallucinating robot.