How to make the job of an OCR engine really difficult

Our new AP Automation software

My accounting department bought a new software for AP Automation and data capture for invoices. I hear it is pretty good getting data off scanned invoices, better than when I key it into SAP. Let’s see if that is true.

AP Automation and especially data capture software has become so good, it is almost a commodity. Scanning an invoice and getting all the key data off it, like invoice number, invoice date, total amounts, and line items, is usually a piece of cake. I think it is more fun to see how I can mess it up. Better than a human? Tsk, I’ll show you! Oh my, I have so many ideas, so let’s get at it!

Let’s start with the paper, even before it gets digitized. I rip open the envelope and see the 3 pages are stapled together. I rip the staples out, there isn’t much useful information on the small pieces of paper that I rip out with them.

Oh no, a page falls to the floor and I step on it. Now it has a very faint footprint on it. I pick it up again which slightly crumples the paper, so I straighten it as well as I can. Should I scan it in now? No! We have processes here, so I stick to my good old “we have always done it like that” routine and take my pen. First, I scribble today’s date on the top. Then I review the invoice and circle a few of the higher amounts that need to be reviewed later. I also write the account code next to one of the line items of course.

Oh, those stamps!

Next comes my good old “RECEIVED & REVIEWED” stamp. Why do I stamp that onto every invoice? I have no idea, we always do that. It is pretty worn by now but it still does its thing. So I ink it. Where do I put it? This invoice is so full of text, no white space… Ah well, it is red ink, so I can just ram the stamp somewhere into the text. I can still read the text, so what!

Invoice pre-screening completed, so now I can put it on the scanner. Wait you say, “on” the scanner, not “in” it? Yeah well, let’s use a nice 29.99 USD flatbed scanner. Why use a fancy 999 USD document scanner? So scan it. Page 2 first, just because. I configured my scanner to save disk space of course. 100 DPI black&white BMP or something like that, right? I read that somewhere.

Or should I just use my phone and submit a picture to the software? Why not, and of course, I don’t bother cutting the document from the photo image so I submit a photo of the invoice laying on my desk, with 50% of the image being my desk and pencils and half my hand to hold the paper down. I look at the photograph I submitted and at the scanned image and well, and hey, as an accountant I can still perfectly read everything!

So this hyper-intelligent AI software should be able to do that, too.

Why did we write this?

We have seen most of these things happening when trying to optimize a customer’s document automation system. Instead of making sure the scanned document is treated well and the scanner is used in the best possible way, some customers seem to insist that as long as a human can read the document, the software should be able to do that, too. But if you want to get the best possible accuracy from your system, you should feed it the best possible image quality!