Rocket Close, Aws Collaborate On Ai System To Automate Mortgage Document Workflows

Rocket Close announced Thursday that it has significantly reduced the time required to process mortgage documents by deploying a generative artificial intelligence (AI) solution developed in collaboration with Amazon Web Services (AWS).

Rocket Close, which processes about 2,000 abstract document packages daily, previously relied on manual workflows that took up to 10 hours per package amid rising volumes. Each package averages roughly 75 pages and contains complex legal and financial records tied to property ownership and lending.

Through the new system, processing time has been reduced to less than two minutes per package while maintaining about 90% accuracy in document classification and data extraction.

“The human will step into almost all of the transactions, whether it’s verifying the data or looking at the exceptions, so that we’re ensuring that we’re doing the right thing for our clients and making sure that they have the proper homeownership rights to the property,” Nathan Schrauben, chief information officer for Rocket Close, said in an interview with HousingWire.

The solution combines Amazon Textract for optical character recognition and Amazon Bedrock for document analysis. Textract converts scanned documents into machine-readable text, while Bedrock uses large language models to classify documents and extract relevant data fields.

“The Textract product that Amazon has is one of the industry leaders. We’ve benchmarked it against other leaders in this space, and they seem to come out on top,” Schrauben added.

Removing roadblocks

Abstract document packages — which can include deeds, mortgages, liens, tax filings and court records — present challenges due to inconsistent formatting, handwritten notes and varying document structures. Rocket Close’s system processes more than 60 document types and extracts structured data across categories such as loan details, ownership history and legal judgments.

“There’s no specific format or standard in which you’re going to receive a package. So what that typically does is you need human experts on the other end in order to process these because you need human judgment. And so that’s where the slowdown happens,” said Sri Elaprolu, director of the AWS Generative AI Innovation Center.

That’s where AWS comes in, Elaprolu said. The automation addresses several operational challenges, including high processing costs, scalability limits and the risk of human error. Previously, the company required an estimated 1,000 hours of manual processing daily.

“We’ve worked closely with the Rocket team in understanding that we’re not the mortgage processing experts; we’re coming from it from a technology perspective,” he said. “Our customers have the domain knowledge of their business. Nobody knows better than them, and so our job is to collaborate with our customers [and listen to] the specific knowledge about their workflows that we’re trying to automate.”

Elaprolu said that the start of the collaboration began with a “discovery process” that allowed AWS to understand Rocket’s systems and problems before building a proof of concept.

“We then sat down and had Rocket experts validate [the concept] with real data flowing through the system, or at least simulated that data that’s pretty close to real, to see if it would give correct outcomes,” he said. “Very often, you’re not going to get it right in the first pass, so we keep tweaking and adjusting. … Our goal is not just automation but … to make sure that the AI that we’re using understands this domain, understands these databases and applies them the proper way.”

Testing showed consistent performance across multiple evaluation phases, with accuracy rates ranging from about 89% to 91% across tens of thousands of data fields.

‘We want humans in the loop’

The cloud-based system is designed to scale to more than 500,000 documents annually and handle increased volume without proportional staffing increases. AWS said the improvements are expected to reduce costs, speed up customer service and support business growth.

“It relieves the human specialists who are in the workflow today to be focused on more complex packages that a system is not going to be able to handle,” Elaprolu said.

Following a successful proof of concept, Rocket Close plans to move the system into full production and expand its use to other workflows, including loan processing, purchase agreements and title clearance documentation.

“We anticipate every two to four weeks that we’re adding a new document into our workflows to allow our clients and our team members the ability to process through much more scale,” Schrauben said.

Schrauben also said that the company intends to implement continuous improvement processes and update its AI models as newer versions become available.

“We’re not looking to get anybody out of the loop. We want humans in the loop because humans are really good at the really tough stuff, like explaining things that are not as straightforward to other clients, especially in the title space. We believe that this technology is unlocking that,” Schrauben said.