My company currently does this using enterprise-grade software, and I'm deeply familiar with the problem area. The article quotes the guy as hearing the problem and saying:<p>--“My instant solution was, ok we have these big devices nowadays… why can’t we use them to capture these invoices, use technology to convert that image into text and extract information out of these invoices,” he adds. “It was pretty much a bottoms up approach.”--<p>Yeah... no. I'm afraid it is much, much more challenging than he realizes. Here's a quick, non-exhaustive summary:<p>1. Most paper invoices are printed and then either mailed or faxed. Sometimes the vendor's accounting staff manually annotate (e.g. hand-writing) the printed paper with additional information that didn't exist on their accounting system (e.g. purchase order number). By the time the invoice arrives at the AP department, the quality of writing on it can be atrocious, which greatly affects the accuracy of OCR. So you need to use complex image processing on the invoice first, and then OCR it. You may also need ICR to read hand-written information.<p>2. A lot of vendors change their invoice formats on a regular basis (e.g. by changing to a different template on QuickBooks). This means that whatever machine-learning you use to process a vendor's invoices will be invalidated when the format changes.<p>3. Smaller vendors, such as mom-and-pop shops that restaurants sometimes deal with, are very undisciplined about their invoicing, and can forget to put key pieces of information on the invoice, or make typos.<p>Over the years, I've come to realize that the only way to efficiently deal with invoices is to make sure they never become paper-based. We have been pushing our customers and their vendors to use EDI instead, which is a LOT easier to control and standardize.