There's a lot of talk about automation in our industry, but what does it really look like to take humans out of the mortgage process? We sat down with Ari Gross, CEO of SoftWorks AI, to talk about how lenders can automate without sacrificing accuracy in the process.
HousingWire: Why is automation so important right now?
Ari Gross: Digital lending in other verticals is progressing rapidly. For example, student loan and automobile lease lending decisions are largely paperless and made in minutes or hours. The average time to close on a mortgage is about 40 days.
HW: Why are lenders still using humans to oversee supposedly automated processes?
AG: Naturally, lenders and insurers and other mortgage stakeholders are very concerned about the cost and time involved in on-boarding, analyzing and pricing a loan opportunity. However, the first priority is to make sure that the loan diligence process is rigorous so that each loan is given the level of scrutiny that’s required.
We believe that digital lending is like any other industry with respect to knowledge worker automation. Automation is always desired provided that the quality of the process does not degrade and, hopefully, improves.
Applying automation methods to mortgage lending has been very much biased towards having the system maximally classify documents and extract data, without a corresponding emphasis on touchless automation. This produces mounds of automated data, most of which needs to be manually reviewed with uncertain benefits from an ROI perspective.
This lack of reliability and confidence in automated solutions and the underlying OCR technology is why lenders are still using humans to oversee supposedly automated processes.
HW: What are some of the problems lenders encounter using traditional OCR technology?
AG: OCR technology is strongly relied on for classification and data extraction but unfortunately is still error prone. A typical, relatively high accuracy for an OCR engine is 99.5%. That still amounts to 1 character error per 200. On a normal form or document, there are typically at least 2,000 characters that must be recognized, meaning on average there are at least 10 errors per processed page. With an average loan packet being about 300 pages, there are on average 3,000 OCR errors in the converted loan packet output.
Without more advanced computer vision and AI methods applied to mortgage processing, it is very hard to advance the state of automation.
In particular, maximizing touchless automation requires that the system knows the accuracy around each operation it performs, including classification, splitting, stacking, and data extraction.
Moreover, OCR technology for mortgage processing is based largely on a static process, rather than a dynamic process which supports follow-up queries. In essence, for a typical OCR engine, a single character misread negates its ability to auto-process a document. A more robust OCR process is clearly needed.
HW: How does your Trapeze software go beyond OCR to get more accurate results?
AG: Our Trapeze software goes beyond OCR technology, incorporating computer vision and AI methods. Typical OCR is static so that if OCR fails to recognize all the characters on a page, there is no way to recover this content. Computer vision, by comparison, is the technology used in driverless cars and other smart devices to dynamically understand captured images, often in 3D, iteratively trying to understand the relevance of each pixel.
AI technology can be used to determine if the OCR output of a document page is considered less reliable, and computer vision can interrogate documents that seem incomplete and can “pull” or force OCR information from areas where the normal OCR process failed. This ability to detect likely OCR failures and force/pull OCR information leads to much higher process automation rates.