BLOG VIEW: Modern document and character recognition technology has been advancing for a long time. Today, it’s possible to deposit a check just by taking a picture with one’s mobile phone – or, similarly, determine who someone is in an image using facial recognition software.
Interestingly, optical character recognition (OCR) – the process by which a machine converts text images on a paper document into electronic text – predates World War I. A physicist named Emanuel Goldberg was the inventor. Later on, in the 1920s, he created the first electronic document retrieval system.
These systems were not very sophisticated, but they showed the world what was possible. It wasn’t until the 1970s that Ray Kurzweil improved the technology enough to recognize text in a variety of fonts with a high degree of accuracy. By that time, the business imperative for these tools was coming into focus.
The technology has evolved apace, and today’s OCR capabilities have become so much more advanced that they can read just about any letter that appears on a document.
Unfortunately, that’s not enough for every business use case.
Because OCR technology has been around for so long, many industry professionals incorrectly assume that any technology that reads paper documents into an electronic system is basically the same. While that may have been true at one point in history, it’s certainly not the case today.
Today’s tools have gone beyond reading the characters on the page to using machine learning and artificial intelligence to determine what type of document the machine is actually reading. This is known as intelligent content recognition (ICR) and it’s a game changer for industries with high regulatory oversight and stiff penalties for non-compliance, including the mortgage industry.
Teaching Machines to Read Beyond the Character
In complex businesses where many documents are required to close a deal, OCR is not robust enough to enable automated workflows. In the mortgage industry, for instance, where each loan package contains hundreds of documents — many of which are non-standard due to variations by region, lender, settlement agent, or vendors — industry professionals need to know what document they are examining and not just what information it contains.
Whether originating a new home loan, selling it into the secondary market or boarding the loan to the servicing platform, compliance requires lenders to collect the right data and the right stack of documentation, while automation requires systems to be intelligent enough to know which documents have already been received.
To automate as much of the process as possible, the technology must be able to recognize the documents before traditional OCR comes into play. This is exactly what ICR offers. Unlike OCR, which looks for specific words to classify documents, ICR uses artificial intelligence and machine learning and applies “fuzzy logic” to the problem to allow the machine to have a contextual understanding of the text before classifying the document.
For example, finding the words “Closing Disclosure” or “W-2” on a document gives a clue to its type and purpose. But the location of those words is also important. If the words are not in the right place, they may just be referencing other documents. To know the difference, ICR systems are trained through exposure to many sample documents until the machine can exactly recognize each type.
Educating ICR systems and then maintaining their capabilities over time has traditionally been very expensive, which put these systems out of reach of all but the nation’s largest financial services companies. However, that’s no longer the case.
The Value of an ICR Component to the Process
Going beyond OCR to ICR offers multiple benefits. These include costs savings, as well as speed and consistency of process. A key secondary benefit is the positive impact the transition can have on compliance by establishing a repeatable and auditable process.
Once the system has been taught, it will recognize documents automatically and alert the user whenever an expected document is not present. Unlike using a large staff for document indexing, ICR applies the same logic to every loan package it examines, facilitating consistency. At the end of the process, the user has an audit record that details the exact contents. This audit trail is critical in adhering to federal requirements, and prudent lenders can also use this information to keep their correspondent, closing and other partners accountable.
In addition to compliance, ICR also delivers true cost savings that allow lenders to improve their throughput. This will be increasingly important to lenders this year. As interest rates rise, the lenders who close loans more efficiently will have an edge in capturing a larger share of the market.
As for pure cost savings, we have seen clients using ICR save up to 70% of their back-office costs, from origination through closing, gaining an opportunity to deploy staff elsewhere. In addition, some clients have reduced their time to close by over 15 days. For correspondent lenders, who may process loans from multiple sources, financial benefits also include ensuring the loan adheres to requirements prior to funding. In the wholesale purchase flow, ICR lenders can afford to verify every loan package instead of auditing a small percentage.
Key Metrics to Evaluate Document Classification Tools
As more ICR tools make it onto the market, the marketing noise around these systems will increase and it may become more difficult to evaluate them effectively. One good way to think about these tools is to consider the four key metrics that pertain to automated document classification.
- 1. Automation
This is a measure of the percentage of documents that are correctly identified by the software. This is a moving target. A system that can correctly identify 50% of the documents it reads on day one will improve over time as machine learning educates the system and improves the algorithm. Lenders who are focused on growing their market share have been able to achieve automation rates above 90%. This requires a dedicated effort to ensure that the product is updated regularly.
2. Confirmation
This is a measure of the percentage of documents that require a human being to confirm the AI’s conclusion. While the goal is to decrease this metric over time, this can be an easy validation process and critical in improving the last two metrics.
3. Correction
This is a measure of the percentage of documents that the operations team must reclassify after the machine classified them incorrectly. This percentage will continue to improve as the AI improves over time with more data specific to that business.
4. Errors
This is a measure of the percentage of documents that the tool confidently identified incorrectly. In many ways this metric is more important than the others because it is a true measure of the system’s ICR capabilities. This metric must be minimized as errors made by automation can impact a great many loans and may not be caught manually.
These metrics are co-dependent. If the ICR confidence level is increased, Automation and Confirmation can be increased. However, this may also increase the percentage of errors. Some of the newest ICR capabilities include the ability to optimize this confidence lever based on business requirements. Understanding these different metrics and knowing the impact each has on the technology and the lender is critical in identifying the true value of the tool. ICR is a significant upgrade from the old OCR technology the industry has used for decades and offers many significant benefits.
A complete system for use in the mortgage world should include document recognition, document versioning, document tracking and audit report generation. The industry’s best systems go beyond these basic requirements with document management tools that will include ability to deliver custom ordered document sets as required by an investor or GSE. Given everything these modern systems can do, it doesn’t make much sense to settle for inferior OCR technologies.
Anil Bakhshi is director of product management, digital lending and originations at Fiserv, Inc., a provider of financial services technology solutions.