Resources

Newsroom

Back to Media & Press Releases
Media

Mortgage Providers Look to AI to Process Home Loans Faster

Wsj Image1

“Mortgages take forever to close,” said Al Pascual, senior vice president of research at Javelin Strategy & Research in Pleasanton, Calif. “If they can cut a week off the mortgage process, that’s meaningful.”

By John Murawski, Wall Street Journal Pro | Artificial Intelligence

March 18, 2019 4:53 a.m. ET 

The home loan industry is looking to artificial intelligence to shorten the wait times for a mortgage approval,

from a month or more to a tolerable week or two.

A speedy loan turnaround not only moves capital for banks

and investors; it also reduces the reliance on costly human labor.

“Mortgages take forever to close,” said Al Pascual, senior vice president of research at Javelin Strategy & Research in Pleasanton, Calif. “If they can cut a week off the mortgage process, that’s meaningful.”

The industry took a step in that direction last month, when Ellie Mae Inc., a major player in home loans, started offering an AI option to its 3,000 customers—banks and other lenders—to speed-read mortgage applications for accuracy and completeness.

Ellie Mae, a Pleasanton-based software and services company that processes about 40% of all U.S. home mortgage applications, selected an AI technology developed by Wakefield, Mass.-based home loan data automation outfit AI Foundry.

“They will market their services directly to our customers,” said Joe Tyrrell, Ellie Mae’s executive vice president for strategy and technology.

While AI is light years faster than people in reading documents, human loan analysts will be necessary to manually validate the loan documents that intelligent machine can’t decipher. And in the home-loan business, there will never be a shortage of crumpled, corrupted or esoteric documents.

Ellie Mae’s loan platform makes sure the loan packet is complete and properly filled out, so that lenders can act on it. The company has been using automation to read credit reports, title reports and home appraisals, which are submitted by institutions and come in standardized formats.

The average loan takes 52 days to process, but Ellie Mae has been able to reduce that to 30 days, Mr. Tyrrell said. The company hopes to reduce that to 10 days by further harnessing AI, he said.

But the loan industry has been constrained by a missing link in its AI initiatives: reading personal documents submitted by home buyers. These include W2 forms, pay stubs, tax forms, divorce decrees, bank statements and other documents that validate the buyer’s finances. Many of these documents are irregular and have been more difficult to automate.

AI Foundry’s technology, called Agile Mortgages Solution, currently is being used by about a dozen banks and lenders, said Steve Butler, the company’s founder and general manager. It takes less than five minutes to read several hundred pages of documents that a human would need an hour to slog through, he said.

“The mortgage industry is ripe for this because it is so document intensive,” Mr. Butler said. AI Foundry is a three-year-old business unit of Kodak Alaris Holdings Ltd.

Still, the system also needs some human help. Users have to train the AI to recognize when it can’t read a loan document, or to ask for help when its confidence level is shaky, so that a person can intervene.

Mortgage processing takes so long in large part because lenders must comply with safeguards established after the subprime mortgage crisis and 2007-09 recession, said Craig Focardi, a senior analyst in the banking practice of Boston-based research firm Celent.

Problematic paperwork is a concern for more than just regulators. Institutional investors who buy the bundled loans on the secondary market are aggressive about spotting defects in documents on loans that have defaulted.

“The big risk is being told by someone who bought the [defaulted] loan that now you have to buy it back,” Mr. Tyrrell said. “The investor can look at the documents for anything that was missed. There’s a lot at stake if the [originating lenders] get it wrong.”

Mr. Butler said the mortgage papers come in as “one huge PDF, all connected.” So Agile Mortgages must first learn how to identify, separate and classify the incoming documents. Then AI has to read the files, extract the relevant data from the appropriate line or box, and enter it into the bank’s system.

One of the first AI Foundry customers, Radius Financial Group Inc. in Norwell, Mass., has been using Agile Mortgages since 2017. Keith Polaski, chief operating officer at Radius, said it took about six months to train the AI on three years of mortgage applications, spanning 180 different types of documents. After 20 months of ingesting data, Agile Mortgages has improved its accuracy rate to about 85% from 40% of documents it can recognize and read correctly, Mr. Polaski said.

Humans are about 97% accurate, based on industry data. Mr. Butler said the ultimate goal for Agile Mortgages is to match human performance. Mr. Polaski said it is only a matter of time until Agile Mortgages is so reliable it will be able to approve and issue some portion of home loans without human assistance.

The volume of paperwork Agile Mortgages can process in one year for Radius is equivalent to the labor of 10 loan analysts performing 6,500 hours of work, at a combined cost of $227,500, for classification and indexing. Agile Mortgages has “freed them up by two hours per day to do human brain power things,” Mr. Polaski said.

AI won’t attain perfection anytime soon. Radius doesn’t train Agile Mortgages on about 50 types of mortgage documents that are so uncommon they will always have to be handled manually, Mr. Polaski said. 

View complete Wall Street Journal article online