AI tools enhance automation, can be delivered through the cloud, and would significantly improve the production of digital mortgages. At the same time, AI is also one of the least understood terms in the mortgage industry. This fact is keeping most mortgage industry participants from realizing its full benefits.
To understand how AI and machine learning are capable of impacting every aspect of mortgage lending, from origination to loan acquisition and servicing, you first need to distinguish the two terms from each other.
AI is a catch-all term that generally refers to the number of technologies that are capable of analyzing data and identifying patterns to make decisions and produce outcomes.
AI can be used to describe any technology capable of applying knowledge to solve complex problems and find the optimal solutions related to a specific task or a range of functions.
AI is like a parent teaching a child who eventually makes his/her own judgment calls and decisions, based on logical, cognitive reasoning.
Machine learning falls under the umbrella of AI. It also involves data mining and pattern recognition. However, machine learning is different in that it allows systems to learn and improve as new information is made available, without requiring additional programming instructions.
More specifically, machine learning involves training systems to identify and determine patterns from large quantities of datasets to complete tasks on their own. It learns through individual instruction and self-learning algorithms that are trained to recognize patterns in data.
In a more tremendous sense, AI describes the broad science of mimicking human abilities. Machine learning, on the other hand, describes the process of training machines on how to learn.
AI Adoption in the Mortgage Industry
To be sure, every industry which depends on large data sets will benefit from AI, such as insurance, transportation, and legal industries. Yet few industries depend on the sheer volume of data that mortgage companies do — nor are they required to work with data from as many disparate sources and in as many different formats.
Few industries, it might also be said, deal with as many different regulations and guidelines that dictate how they do business than the mortgage industry. This means that lenders have to be very tactical about how they leverage AI and machine learning, especially when it involves credit decisions and meeting the requirements of multiple regulators and loan purchasers.
For the past several years, banks, insurers, and lenders have been investing in AI and machine learning technologies. By 2017, according to PwC, more than half of all financial services firms had reported making significant investments in AI, while two-thirds said they planned to do so within the next three years.
In Fannie Mae’s 2018 Lender Sentiment Survey, it was reported that the use of both technologies was gaining momentum among mortgage lenders. In this survey, “How Will AI Shape Mortgage Lending? ,” the GSE found that 63 percent of lenders were familiar with AI and machine learning technology, and 27 percent of lenders were already deploying them. Fannie Mae also found differences in adoption rates between banks and lenders. Today, mortgage institutions are where one is mostly like to find the most successful and best uses of AI and machine learning technologies
According to the survey, only 9% of depository institutions stated they were “very familiar” with AI and machine learning solutions, whereas 22 percent of non-depository institutions — companies in which mortgage lending was their core business — were “very familiar” with them. As well, the GSE also found that only 6 percent of depository institutions had deployed AI solutions or incorporated some form of AI tools into their mortgage business, compared to 22 percent of non-depository institutions.
The Entry Point for Machine Learning
For lenders that are already using machine learning, many are finding that the most effective application of these technologies is in data and document processing.
What’s driving this adoption are the challenges related to current methods and the fact that machine learning lends itself well to the tasks required to recognize documents so that data can be effectively extracted.
Today, most lenders are still using optical character recognition (OCR) technology to extract data and aid in the classification of loan documents in various electronic formats. In theory, this technique saves lenders an enormous amount of time and expense that would otherwise be spent on manual indexing and retyping information by hand into the loan origination system or other systems of record.
But, as a data picker, OCR falls short from a classification perspective. Template-based OCR tools have been used to try and close this gap, but they rely too heavily on data being found in the same location on a particular document type every single time. That is why they work best on structured documents using a predefined set of templates.
For example, on a Uniform Residential Loan Application, specific data attributes will be in or near the same location consistently. For such documents, template-based OCR tools do a reasonably decent job, except where there are variations in the document structure affecting the location of specific data.
In the best-case scenarios, OCR is only able to identify a portion of the document types found in a borrower’s loan file. This means lenders must use other manual methods to classify unstructured documents and address situations where templates fail. To maintain any accuracy, document templates need to be tweaked continually, and lenders must rely on human staff to address the gaps.
Compared to current methods, machine learning, on the other hand, makes it possible to “grab” a much higher percentage of information from mortgage documents and turn that information into actionable data.
It can solve problems under conditions where rules are not easily codified. It learns about document variations and where data might be found based on the availability of a robust set of documents examples that can act as a training set.
As a result, it can consider differences in various loan documents and classify more documents more quickly and accurately than processes used in the past.
Unlocking the Potential
OCR tools still have an essential role to play in the gathering of data, but machine learning tools are being deployed to create even greater value for financial institutions who are using it in data and document processing, improving both the transparency and auditability of loan documents and the accuracy of extracting the loan data found within them.
The key to maximizing these results lies in the amount of training data available to aid in the development of machine learning applications. Unless a lender is working with a regtec provider that has amassed a rich repository of document examples, it becomes difficult, if not impossible, to leverage machine learning technologies without supplementing the training set with enough examples to achieve any consistent degree of accuracy, which takes time and internal resources.
For this reason, there is an advantage gained by machine learning vendors who have the needed “training set” to maximize performance and increase the accuracy of the results that their machine learning applications are trained to do. By blending OCR tools, machine learning, and more sophisticated data extraction programs, lenders can now accurately identify and classify a greater range of loan documents and extract data from them.
This combination of technologies has been broadly defined as Capture 2.0 technology (also described as “automatic identification and data capture” and “intelligent capture”). Simply put, Capture 2.0 solutions can translate content, make it understandable, then route the information to automated business processes.
Since its inception as a software category, Capture 2.0 has adapted itself to enable business processes and to eliminate the associated manual work, often involving keystrokes. According to one market intelligence research provider, the global capture software market is expected to reach more than $11 billion by 2026.
The Power of Capture 2.0
Within a Capture 2.0 approach, OCR becomes the starting point, creating big blocks of data capable of feeding machine learning models. These models are “trained” to recognize patterns through textual analysis that can classify both structured and unstructured documents.
From these documents, data can be more efficiently and accurately extracted and verified. This results in the form of “purified data” that can then power rules-based logic and other AI applications. Capture 2.0 involves a three-step process when transforming loan documents into actionable data:
1. OCR tools turn scanned, imaged documents into text.
2. Machine learning tools drive automated document recognition (ADR) that can classify and prioritize loan documents, assigning each document a confidence score.
3. Automated data extraction (ADE) tools then identify scalable data elements within the document, not a keyword search, but instead looking for words used in context.
For example, a borrower’s first name may be in one location on one document, but on the next document, the name could be in a completely different location. By leveraging pattern recognition and large datasets of documents, Capture 2.0 technologies are still capable of finding, identifying, and capturing that data, regardless of the number of documents it might be found on or whether they are structured or unstructured.
An even better example is a gift letter, an utterly unstructured document, lacking in any format and varying tremendously as a document type. However, there is content and context that is consistent — someone is gifting someone else an amount of money.
Using machine learning, trained algorithms, and an enormous document sample size, it is possible to classify gift letter documents and extract data from them consistently with as high as 90 to 95 percent accuracy.
By leveraging machine learning technologies in this way, lenders are now able to determine the accuracy, quality, and completeness of borrower information while it is being collected, reducing both the costs of manually retyping data and constantly asking the borrowers to resubmit their information.
The ROI of AI
By far, the most powerful benefit that AI provides lenders lies in the ability to remove costs. According to the Mortgage Bankers Association’s most recent figures, the current average cost to produce a mortgage loan is over $9,000, the highest ever recorded.
A significant portion of these costs is spent on human capital. Anywhere AI and machine learning can automate tasks can help to reduce these costs and create a much better experience for the consumer. A question we frequently encounter when discussing the impact of AI on the mortgage industry is whether such innovations will lead to a loss in jobs.
What we more consistently see as the impact of its use is a better ability to scale in support of business growth and a refocusing of skills on business process improvement and more sophisticated risk management strategies.
Regardless of whether the further adoption of AI will lead to job reductions, AI is almost sure to drive higher demand for new skills among mortgage professionals. There will be a much greater need for the ability to instruct and guide AI and machine learning tools, especially around expanded applications related to automating controls and decision making.
It will take some time before AI becomes commonplace in a mortgage lending environment, and perhaps even longer until lenders can maximize its benefits. However, it’s believed that lenders that are taking steps today to implement Capture 2.0 technologies on their most labor-intensive processes will attain competitive advantages so significant that they could alter the mortgage landscape.