In this webinar, we’ll talk about and show some exciting concepts in Financial Services intelligent automation that focus on having applications take over the laborious chore of document ingestion, classification and data extraction through the use of Artificial Intelligence and machine learning.  

If you aren’t familiar with Ephesoft, we specialize in using AI and machine learning to make sense of unstructured data in the form of documents. We have two products: Ephesoft Transact and Ephesoft Insight – the first is focused on transactional document processing that is usually tied to a business process while the later, Ephesoft Insight, is a big data analytics platform that converts large document repositories into meaningful and actionable data.

The Problem

Every day, every organization on the planet deals with the same, painful issue. They receive documents in all forms from a variety of external sources: mail, email, uploads, scans, mobile, acquired or transferred content.  The challenge is that each type of document needs to be manually examined, analyzed, classified, routed and then the important data is extracted and transferred. A simple example is when part of a mortgage packet is sent via an email. It gets opened, examined, renamed, saved and the data is manually entered into a system of record. What happens if that packet has multiple documents? What if pages or signature are missing? This very basic individual process has a huge impact on organizations.

Let’s look at the problem on a grand scale for a financial services organization. Document ingestion and creation never stops. Daily, massive volumes of documents cross the organizational boundary. Internally, users and systems create new content through their daily efforts, scanning and interaction with systems. Pile this on top of the pre-existing content, content acquired through mergers and acquisitions, and the problem continues to multiply.

To cope, organizations hire more and more processors to keep up. They implement complex technical solutions that work fine on day one, but are often inflexible and rigid and cannot keep up with the pace of change. Or it only handles inbound documents from one source. Processing can get even more painful, and in many cases organizations just can’t wrap their hands around the problem, or don’t have systems for capturing documents and automating the process. This survey of business executive from AIIM outlines the low percentage of organizations that have multi-channel systems that can automate the document ingestion process.

Mortgage Pain

In no area is this pain more pronounced than in processing mortgage documents. With a massive and broad array of document types, varying form versions and different state requirements, it is difficult for even the most seasoned knowledge worker to ingest these documents. Add the fact that many of the documents are poor quality and they are typically provided in bundled PDFs or TIFFs, or large paper stacks, and they consume endless hours of processing. Compounding all of the above is the fact that you can’t just hire temps to process. Mortgages and their associated documents are complex and involved, and understanding all the relevant components is not learned in a day. Staff turnover and the need to add people to process more just is not an option.

In today’s FinServe environment, the customer expects fast service and immediate results, but there are a number of disruptors. Rapid turn cycles for mortgages, account openings and all other functions are the norm instead of the exception. To be competitive, efficiency and speed are imperative.

How do we fix it with AI and Machine Learning?

There are core technologies that can automate human tasks and relieve the pain. When we talk about AI, most people envision the Terminator or HAL from A Space Odyssey. But by definition, AI is a system that performs tasks normally performed by humans. In our case, and for the purposes of this webinar, the task is the processing of incoming documents.

Machine learning is a branch of AI, whereby a system is trained using sample data. There are two types: supervised (SML) and unsupervised machine learning (UML). SML is the most common form of machine learning and is like a student-teacher relationship. We provide known samples as input data and know the result or required output. An example would be training a system to identify pictures of mammals. We provide input photos of bears, foxes and lions (because we know what they are), and use algorithms that tie the name or label to the input data (photo). In UMLoften associated with true artificial intelligence, the machine identifies patterns on its own without supervision of guidance from humans. Using the example above, when you feed it 20 pictures of mammals, and it begins to group, or cluster, them into similar buckets. Essentially, we give the machine a data set and let it analyze, infer and find hidden structure. The systems we talk about in this webinar are SML based.

Limited Human Intervention

The good news is, the process can be fixed with the mentioned technology. If you examine the inbound document process, you can quickly identify the key chokepoints:

  1. We need to solve the multi-channel document collection problem and provide a process that can tap into any source
  2. The core functions of document analysis are repetitive, and can be replaced through an artificial intelligence technology supported through supervised machine learning
  3. Only use humans for exception processing to train the system so it gets smarter
  4. Finally, to reduce costs and decrease time to value, a prebuilt model is key

Having a “gate master” as a main controlled point of entry for all incoming documents is critical. This allows content to be consumed from any source and creates a standardized, repeatable path where all content is normalized and then passed to the AI processing engine for analysis. This engine is pre-trained using supervised machine learning to perform analysis, just as a human user would.

Here’s how the process and technology works.


The first step after import the system performs and automates is classification. Just as a human knowledge worker would examine documents, the application applies its learned model to figure out when one document ends and another begins. It auto-classifies, or identifies the document type and all the pages contained within. This function allows documents to be captured in bulk bundles, such as large PDFs or paper stacks, and does all the heavy lifting in short order that would take a human worker extensive time and effort.


Once the system has identified all the documents in the train of pages it receives, it categorizes them into individual sets. This process is similar to the task a human would perform on a stack of documents that need to be split, called separation. Separation provides the ability to output individual documents as the end result of the process. When you combine classification and separation in this automated form, using your SML model, you create massive time savings and efficiency, especially when dealing with large multi-document PDFs or large volumes of paper files. This combined process also preps the documents for our next automated step.


Manual data entry is the next human process we can tackle. Using rule sets tied to our classification, we can now auto-extract data of interest from the document. These rules can differ and vary per document, and the technology eliminates the need for a worker to manually enter data, manually name a document, or manually create folders. This data can now be routed to multiple destinations such as an LOS or repository.


Manual data entry by knowledge workers can also have a high rate of error. With smart document capture, rules can be built to catch documents missing pages and low-quality data and as well as data that doesn’t match what the machine expects. This process of validation and exceptions processing can eliminate errant data, format information and insure high quality. Note: this is the only step that requires human intervention, and these queues can be used to train the system further through an easy to use SML interface.

The Ephesoft Solution

We’ve now covered the technology required for the fix. Now let’s see how Ephesoft provides the means to automate the process.

Patented Analytics & Supervised Machine Learning

Our technology takes a holistic view of documents, just like humans. If I were to hand you three mortgage documents, an application, a bank statement and a credit report, you would scan the pages and look for identifying items to ascertain what they were. To achieve the intelligence required to perform a task like mortgage processing, when the system is trained, we look at patterns, the roots and stems of words, fonts, layout, and in total 21 different dimensions to learn how to classify and extract.

In Ephesoft Transact for Mortgage solution, we have implemented a pre-built model for over 600 common document types for classification out of the box. It automatically and accurately classifies and separates documents, before validating and exporting them into loan origination systems (LOS) like Ellie Mae and Mortgage Cadence, along with others.

This technology provides immense value, and the unstructured to structured process provides automation and efficiency, reducing the time to process and required personnel, insuring accurate extraction and valid, clean data, and all the while getting smarter and smarter through operator interaction and learning. In addition to the core operational benefits, Ephesoft Transact for Mortgage was built in the cloud. Therefore, companies can rapidly deploy the solution, and immediately start gaining value and efficiency. The offering is flexible, and can be scaled for small and large organization alike.