Resources

Newsroom

Back to Media & Press Releases
Media

AI Foundry and radius financial group inc. Partner To Automate Residential Mortgage Loans

Forbescom

Radius has made substantial progress toward the “no human touch” goal. They now process more than 90% of all documents through AI engines.


Tom Davenport | Forbes.com 
November 6, 2019

There are thousands of business niches that can benefit from artificial intelligence (AI), but perhaps none more than residential mortgage processing. It’s almost always a slow process now, but most people who apply for mortgages would like to have the process be much faster. It’s highly information-intensive, with information coming from a variety of sources: banks, employers, real estate agents, assessors, etc. There are substantial fees involved, so there is both money to invest in improvement and a need to make to make things cheaper.

So I wasn’t surprised when I heard that AI Foundry, a company based near Boston, was pretty far along in automating the mortgage process with AI. Steve Butler, a serial tech entrepreneur, has been working on that problem since 2015. Mortgage processing is a document-rich process, and fortunately AI excels at extracting information from documents. But mortgages involve thousands of different types of documents, and they have multiple media and formats—paper, fax, electronic, and so forth.

A couple of years ago Butler and his colleagues decided that deep learning models might be the key to higher levels of accuracy in extracting information from documents. They’re sure they were right about that hypothesis, but training all the models is a bit of a slog. So far they can identify and extract information from 300 types of documents—including loan estimates, closing cost disclosures, W-2s, and appraisals—but Butler would like to get to at least 600. Training the models for each type requires substantial amounts of anonymous, curated, and labeled data, which AI Foundry gets from its customers. They have a few hundred thousand training data documents, but more will be required to master the other document types.

The Workflow and Analytics of Mortgage Decisions

The deep learning models are primarily about extracting a couple thousand data elements from documents, but there are other aspects of the mortgage process that are more workflow-oriented. When a document comes in, it must be recognized, marked as received, checked for signatures, put through several compliance and quality checks, and entered into a mortgage company’s loan origination system. Robotic process automation (RPA) systems are perfect for that kind of work, being a combination of workflow engine and non-human user of multiple information systems. There is a lot of contingent branching based on the data that most commercial RPA systems find challenging, so AI Foundry developed its own RPA capabilities (and interfaces with other RPA systems as well). The company already has developed a loan processing robot, and will eventually have them for compliance and underwriting as well.

On average, processing and approving a mortgage today takes 40 days, which is improved slightly from the 47 a few years ago. The average cost per mortgage is about $9000, roughly half of which is from back-office labor costs. Butler at AI Foundry believes that they can process mortgages in at least 1-2 weeks, at 10% of the mortgage bank’s back-office costs. It’s a no-risk proposition for banks offering mortgages; they pay by the loan to AI Foundry. Since mortgage demand is quite cyclical, they don’t want to keep people around when they aren’t needed.

Applying AI Foundry Tools at radius financial group inc.

One of AI Foundry’s customers is radius financial group inc., a small but rapidly growing mortgage finance organization based south of Boston. radius does about $25 million in annual revenues and has 140 employees. It does between 2500 and 3000 loans per year. For an organization of its size, it is very aggressive in its embrace of new technologies. In addition to AI Foundry for mortgage processing, it makes use of business intelligence, robotic process automation, predictive analytics, and other forms of AI.

Keith Polaski is co-founder and Chief Operations Officer of radius; he focuses on processing the loans, and his business partner focuses on customer acquisition. He said that the company was AI Foundry’s first customer for mortgage banking, starting with it in June 2018. “These are not out-of-the-box solutions,” he commented. “They require a lot of customizing and integration. For example, we have our own RPA system, so we don’t use AI Foundry’s. But they are an important partner of ours, and we are migrating to their newest platform.”

radius’ goal is to deliver all of its loans without a human touch to secondary mortgage market buyers like Fannie Mae or Freddie Mac. They’re using AI Foundry’s capabilities for three key steps:

  • Indexing and categorizing all the needed documents
  • Extracting information from the documents
  • Underwriting the loan using machine learning algorithms.

Polaski feels that AI Foundry has the first two steps largely under control, but the third—making the actual mortgage decision—is still evolving. This is partly because radius has its own criteria for underwriting mortgages, and it attempts to “hand-build” mortgage loans that other mortgage finance organizations can’t. This requires a high level of customization of the underwriting criteria in the AI system.

Despite this evolution in a critical component of the system, radius has made substantial progress toward the “no human touch” goal. They now process more than 90% of all documents through AI engines. They hired two people to be loan exception handlers—Polaski refer to them as “professors of machines”—to deal with issues that the system can’t address. Now they are down to only one as the number of exceptions decreases with system and process improvements.

Polaski is confident that the AI-based approach is already paying off. His two largest work groups are loan processors/analysts and loan officer assistants, and data on the tasks they work on suggested that 30-35% of their workday was spent on indexing and classifying documents. After 20 months of investment in the AI system, they know just how many person-hours they have saved. Overall, they’ve already driven their loan manufacturing cost down by 70%.

Polaski and his management team refer to the AI and RPA capabilities as their “digital workforce.” It works hand in hand with the human workforce, and is positioned as focused on menial tasks. Employees are not worried about job loss—they’ve been told that they’ll be needed as radius grows its business—and they are the primary source of new tasks for RPA robots to perform. Polaski commented, “We’re in a refinancing boom now, and we’ve had a 2.5X increase in volume. Yet we lost two FTEs to attrition and we are able to process it all. The automation gives us the elasticity to scale our business.”

A “NewCo” Partnership Between radius and AI Foundry

Polaski described a new business opportunity that radius is pursuing in partnership with AI Foundry. “We’re standing up a NewCo—it would be lights-out processing of document-intensive businesses like mortgage finance and insurance underwriting. We will license our workflow and loan operations systems. It’s basically outsourcing to machines—but it will all be onshore.”

AI Foundry has larger customers than radius—their average customer processes between 5000 and 25,000 loans per year—but none is more innovative and growth-focused. Both Butler and Polaski believe the possibilities for AI to power growth and speed in mortgage loans and other similar domains are unlimited.


View article online+