In the past 12–18 months, there’s been an explosion in intelligent document processing (IDP) tools, each claiming to have the best model for capturing data from semi-structured documents, like invoices. Every tool is keen to assert its model is the strongest and most trainable. I’ve tested quite a few, and there are some excellent cloud-native solutions available; however, there are still invoice layouts that pose a real challenge for these tools.
Take a recent example: an invoice that details quantities of products delivered by day of the week. The designer has done a great job creating a layout that’s clear and easy for a human to understand, but with the combination of consolidated invoicing and day-of-the-week breakdowns, it becomes challenging for a recognition engine to interpret accurately.
To get accurate results, each item needs a row for every day, with the corresponding order number, quantity, and price. This requires not only correctly capturing the data on the page but also performing calculations, as the correct line totals aren’t provided in the text.
No amount of ‘point-and-click’ training can overcome this issue. However, with some advanced templating logic and a script applied to the Net Total field, we can achieve the accurate results needed.
The IDP tool used in this case was https://www.abbyy.com/flexicapture-for-invoices/.