Indecomm Automates Complex Mortgage Document Processing with Amazon Textract

The most time-consuming and critical tasks associated with mortgage loan origination—commonly referred to as the loan application process—are reading, analyzing, and comparing information across a large repository of documents. Lenders spend an inordinate amount of time manually reviewing documentation, which reduces productivity and lengthens the time required to obtain a loan.

With decades of mortgage industry experience and millions of data points stored, Indecomm has a deep understanding of the loan lifecycle. The company’s mortgage automation products help lenders, insurers, and financial agents streamline back-office operations so they can spend more time improving the borrower experience. Its Genius product suite addresses many inefficiencies in underwriting and other preliminary stages of loan processing. In 2019, Indecomm sought to drive further automation in document processing and analysis by developing an improved data extraction solution using machine learning (ML). 

Over a period of three years, Indecomm evaluated in-house and third-party alternatives to support the development of its Intelligent Document Extraction (IDX) ML tool. The tool goes beyond simple optical character recognition (OCR) to identify and classify documents; extract, validate, and certify data; and enrich data as needed. IDX serves as the underlying document extraction technology powering three Indecomm products: IncomeGenius, DecisionGenius, and AuditGenius.

Scalability, document turnaround time, cost, and configurability were leading considerations in solution evaluation. The business eventually decided on Amazon Textract on Amazon Web Services (AWS) to build IDX, choosing the service for its scalability and integration with serverless tools such as AWS Lambda. Previously, Indecomm required many virtual machines to meet data processing requirements, often exceeding budget thresholds when large jobs came in. Amazon Textract’s application programming interfaces (APIs) allow parallel processing, which facilitates rapid document analysis at scale without additional delays or overhead. 

Indecomm’s SVP of Engineering, Dr. Harish B. Kamath, says, “We found that Amazon Textract offered us the highest flexibility of use and lowest cost to develop our IDX module. We utilized all the capabilities of Amazon Textract, in combination with AWS Lambda machine-learning components, to map out hundreds of mortgage industry-related documents and extract over 4,000 data fields.”

Indecomm used to experience delays of up to 5–6 hours in processing long document queues due to corrupt files, leading to increased costs and management overheads from constant monitoring. However, the integration of AWS Lambda and built-in monitoring through IDX allows for on-demand monitoring, effectively removing bottlenecks from the system. 

Prior to IDX, extracting data from a 100-page document took 30 minutes. With Amazon Textract, the new solution efficiently converts images to text at the field level and enriches data within 5–7 minutes. This has been especially helpful for mortgage lenders dealing with self-employed borrowers who often present non-standard income documentation. 

Indecomm and its clients have experienced a significant reduction in turnaround time and loan origination costs; jobs are now scheduled every 2 minutes, compared to the previous wait time of over 20 hours for sequential processing of 800 pages. Document classification costs have also decreased from $5 to $3 for the same 800 pages. Overall, costs, including data enrichment, security, storage, and reporting, have decreased to just 2 cents per page on AWS. 

In DecisionGenius, IDX works to automate data verifications, reducing the number of manual loan file interactions. As a result, lenders have lowered the required number of file “touches” by 50–60 percent, doubling underwriter and processor productivity. A knock-on effect of less manual intervention is higher accuracy; Indecomm’s IDX boasts a document classification accuracy rate of 100 percent and a data extraction average of 97 percent accuracy for its average loan package.