PDF-Reading AI vision technology hits at new productivity

PDFs are an old technology that Adobe introduced more than two decades ago. The format was helpful for transferring documents over the internet and making them edit-proof. However, they came with a snag: it was hard to cut-and-paste information from them. 

Later, Microsoft introduced tools to make this easier with written documentation. It became possible to highlight text, provided the PDF allowed it. However, PDFs were still incompatible with modern data management systems. Even if you wanted to extract tables and charts and put them into databases on SQL, you couldn’t. 

Enter PDF data extraction, a technology that fundamentally changed the landscape. Thanks to AI, software is now able to scan PDFs, pull necessary details in context, and then transcribe them into a more shareable and amenable format.

“Machine vision and learning technology has come a long way in a short space of time, and today’s vision systems are significantly better than they were even five years ago,” explains document solutions developer, Apryse. “In fact, systems are now so good that companies can simply read PDF documents, extract data, and then use that information in their business intelligence systems, regardless of format.” 

Why Machine Vision Improved

Machine vision was able to improve significantly in the early days because of new approaches to intelligence. Instead of showing computers who objects, researchers broke the process down into elements that computers could then reconstruct, similar to how the human brain is believed to detect objects. 

First, computers break up images into sections and look for edges or differences in quality between one part of the square and another. Then, they look for patterns among those edges, including things like color, texture, gradient, and polygonal shape. Finally, they bring together multiple cases of these structures in a single image and then calculate the probability that what they’re seeing is an object they can categorise. 

This process is significantly easier for text and numbers than it is for cats and dogs because there are so many examples of the former (even compared to the latter). Learning algorithms are able to detect these shapes with a high degree of accuracy and low error rates, leading to much better calculations and computations. It means that extracting information from a PDF document is significantly less challenging than it ever was in the past. A piece of code can look at the document, find shapes that it recognizes, and then convert these into a format that is useful for the user. 

“The level of the technology now means that businesses are benefiting from higher productivity,” explains Apryse. “Software is able to go over documents, find information, understand the context of that information, and then recreate it somewhere else in a format that business systems can actually use. It means there’s less dead data hanging around, and more opportunities for effective exploitation, even if it is currently in a non-conventional format.” 

Real World Impact

The real world impact of this technology can sometimes be hard to imagine, but it is certainly there. Already companies across the U.S. and beyond are using PDF data reading and extraction to save time and money, and simply get more done. 

For example, legal teams are looking at ways to turn 100-page contracts into structured data in seconds. PDF readers eliminate the need for clerks and paralegals to conduct painstaking reviews and summaries, many of which may miss key points. 

Finance departments are seeing similar benefits. Auto-extraction is allowing them to lift tables of financial information from scanned invoices and reports with near-100% accuracy in a way that simply wasn’t possible before. Ultimately, what took days and was soul-crushing copy-paste work is now something that software can essentially automate and do by itself. 

Five years ago, optical character recognition was brittle. Tables would often break scanning software and cause it to lift incorrect information or misread the context. But today, the situation is different. Scanned documents are no longer the nightmare they once were. 

“Layout-aware AI has been a significant breakthrough,” explains Apryse. “It’s this little change more than anything that has given PDF readers new clout. Many systems no longer need carefully cleaned environments to perform their functions. Even handwritten notes can be read with astonishing accuracy, meaning that virtually any type of document can be read, not just those written in PDF format.”

Apryse explained that many of these readers actually have higher-than-human-level accuracy, meaning that more than 99% of characters are read in the right way. This level of accuracy is possible because the pattern-recognising algorithms operating under the surface are actually more effective in this highly-trained niche than for people, which remains shocking for many companies that deal with a lot of clerical tasks as part of their daily operations. 

Interestingly, though, it’s not just about pulling information from PDFs anymore — which was a breakthrough in itself when it became reliable a couple of years ago. It’s also about fully automating the document workflow. Once information is structured properly, it can immediately flow into business intelligence systems, ERPs, and LLMs. Companies can use their data as they want, retaking their sovereignty over it. 

Over the next 12 to 24 months, further changes are expected. Multimodal models like Opus and Claude 3.5 are enabling real-time collaboration with PDFs. The goal is to reduce the necessity for exporting images and text, allowing workers to simply interact with these documents how they want and then adjust layouts and text accordingly. Systems may even start anticipating the data that needs extracting or entering based on an employee’s role, something that hasn’t been seen before. 

“PDFs were designed in 1993 to freeze information in place,” Apryse explains. “But more than thirty years later, artificial intelligence is providing the tools to un-freeze it and allow companies to use it in the way they want. What’s more, the possibility is now here that companies can actually query the data contained in their PDFs, using advanced AIs to lift information and use it the way they want.”