A new paradigm for document understanding in banking


The old ways of working aren’t working for banks anymore. While the pressure to deliver differentiated customer experience continues to grow, most banks have focused their investments in the top of the tech stack. This includes the front-end experience, web portals, and onboarding. But behind the slick appearances of mobile banking apps, middle and back-office workflows and operations are riddled with manual steps and inaccessible data that can be summed up in a word: documents. These documents come in the form of all kinds of unstructured information––PDFs, emails, Slack messages, Zendesk tickets––and underpin customer journeys.

In recent years, banks and insurers have been investing in robotic and business process automation to automate and streamline their businesses. However, this effort has largely failed with just 10 percent actually implemented at scale.[1] These technologies are only suited for structured data––just 20 percent of the information large banks and insurers handle.

instabase logoSo, what’s happening with the other 80 percent of data? Whether it’s bank statements and pay stubs for retail banking, standard settlement instructions and broker confirmations in investment banking, or insurance claims, these all still require humans to manually review, sort, and understand data that is largely inaccessible. For example, an average mortgage application goes through thirty-five manual handoffs before completion.[2] This data deluge directly impacts customer experience, making document understanding a foundational differentiator for banks. In addition to preexisting competition from neobanks, more recently, the pandemic created a new urgency to solve this long-existing problem.

Obstacles to Full Automation

In trying to automate middle- and back-office workflows banks are essentially answering three key questions that come with their own obstacles:

How do we understand data?

Current approaches –– which often include in-house solutions –– all fail to move the needle for one simple reason: the variability of documents is nearly endless. Even worse, the variability creates more manual work. There are three ways banks typically approach extracting data from unstructured sources:

  • Template-based extraction: creating templates for each type of document. As long as the layout of information in a document is exactly the same way as it was originally, then the data can be understood by the template.
  • Rules-based extraction: creating a set of rules to extract data from a document with clear labels and consistent attributes that can be programmed. For example, writing a rule that says, “find date below the word ‘date.’’”
  • Machine learning extraction: use sample documents –– as many different types as you can get your hands on –– and train a general purpose model to recognize data fields within those samples. Any unfamiliar documents require more training.

How do we configure an end-to-end solution to manage the data?

Using all your data to create a seamless customer experience requires integrating external systems with the rest of your business. However, the inflexible nature of a bank’s on-premise infrastructure means using technologies that are highly integrated with its existing solutions. A point solution simply will not do, and a bank’s tech stack consists of disparate tools for doing different things. Documents in different file formats are spread across different departments and systems, which are limited to consuming only a very specific structure. To configure an end-to-end process requires stitching these various tools together in a seamless way or doing it manually with a team in-house.

How are we able to put the data to work?

Fundamental to putting data to use is scaling it across your organization and other branches worldwide. Modern technology isn’t typically able to run at scale in a way that large banks and insurance companies need using secure on-premise technologies. Secondarily, today’s digitally-focused customers require more services and software be delivered to them more frequently. To keep pace, banks and insurers not only need to be able to extract actionable insights from all sources of incoming customer data but also shift to agile development methods to build solutions faster.

A new approach

The data problem for banks and insurers is huge. But just like any other overwhelmingly large problem, the best way to solve it is by breaking into manageable parts. Here’s an example of this idea in action:

Say your bank wants to understand the information in an invoice. Instead of creating a template for an invoice or training a big machine learning model on invoices, you can forget for a second that it’s an invoice. In fact, it doesn’t even matter if you’ve ever seen an invoice before, because you’ve already seen all the components within it––names, addresses, signatures, tables, dates, etc. You have different best-in-breed technologies that you can package to apply to each component of the invoice all at the same time: computer vision to find the signature; natural language to find the names; structural detection to find the tables, and so on. The result is a more accurate, clearer understanding of the invoice.

This is a new vision for automation that accelerates how documents turn into insights that turn into processes that turn into experiences. When assessing automation solutions that leverage this approach look for those that:

  • Provide best-in-class accuracy for data extraction from unstructured data – not just a simple OCR application
  • Offer easy configuration of the plumbing for end-to-end management and allow for leveraging many different technologies for specific solution needs
  • Enable an modular approach to deploy solutions across your organization and to other branches
  • Are able to run in a hybrid environment using new-age integrations like Kubernetes and are highly secure and scalable

Document understanding at scale

In the race to deliver superior customer experience, banks and insurers need more than traditional general purpose tools to truly understand trapped data. The focus on the cost savings potential of automation is now shifting to top-line value. And with this shift, a new way of working is quickly emerging.

Next-generation solutions will package best-in-breed technologies into solutions that package computer vision, natural language processing, and other extraction technologies into building blocks. In turn, these building blocks can be used to build end-to-end applications workflows: document packet splitting, classification, extraction, validation, review, through integration to the right downstream system. For customers, what once took days, like a loan application approval, can now happen instantaneously. This kind of digital transformation of customer experience from the inside-out is how legacy firms stand to gain the biggest competitive advantage.

To learn more about how the right technology can transform how you work with complex documents and data, check out instabase.com for further details.

[1] Growth in the machine, Capgemini, 2018.

[2] Trivaeo, “Automating Back Office”, March 2013