Record-low interest rates are driving unprecedented volumes of business for mortgage originators, insurers, servicers, and solution providers. The industry, which has long relied on manual processing of document-driven workflows, is looking to new automation solutions to help scale business and develop greater resilience in times of boom and bust.
RT Insights spoke with Dr. Ari Gross, one of the mortgage automation industry’s leading innovators, about the challenges of choosing new automation technology, the use of AI to solve mortgage industry challenges, and how to best assess solutions and vendors.
RTI: Why are companies in the mortgage industry interested in automation solutions?
Gross: People want a way to streamline their processing, a way that they can scale. When mortgage rates are at a certain place and the opportunity suddenly increases tremendously, many lenders have no simple way to increase their capacity. If they normally underwrite 100 loans a week, and there’s a really strong home lending market, they may need to underwrite 200 or 300 loans a week. Where are they going to get that excess capacity?
Also, if we have a pandemic and issues where people can’t come in and productivity declines, how can you supplement that? Between the increased demand that everybody saw last year and people working remotely (and maybe not as effectively), the need for automation is more acute than it’s ever been.
RTI: What are some of the key technology challenges facing companies when selecting an automation solution?
Companies really want to see clear ROI and demonstrable benefits from the technology. There are a lot of companies out there that may be automating some data collection, but don’t know how to verify if the data has been extracted correctly. In this kind of half-baked automation solution, you might not experience the desired benefit.
Other companies have never tried automation, but they really want to find a way to scale. Or they have tried a few types of technology and been underwhelmed with the results. They want to believe that automation can radically change how they do things, but they need to be convinced because they haven’t necessarily seen it in practice.
RTI: Based on your experience, what are some of the common pitfalls that these companies face when choosing mortgage automation technology?
Gross: We’ve seen a whole bunch. One is when decision-makers at the top decide that their company needs to push innovation. They might say, “We’re going to hire an AI company that’s going to automate what we do.”
But what they do includes about 12 workflows and 12 use cases, each of which has got to be finely tuned or they’ll get nothing. You wouldn’t want to play 12 symphonies at the same time. You won’t get a symphony. You’ll get a cacophony, right?
RTI: That’s true.
Gross: They might say, “Look, we want to see automation benefit us, so we’re going to spend a lot of money this year, and we want to see three or four things radically disrupted.”
I call that the “boil the ocean” syndrome. Instead of taking one or two of their chief pain points and understanding their requirements, they just want results right away. And they want to throw three or four problems at it. They have a very general notion of what they’d like done, but no clear roadmap of how to get there. Often, none of them is solved satisfactorily because there was no one on the company side to develop precise requirements.
These companies would do better to investigate their requirements first by asking questions such as: How is that process done now? What does the manual workflow look like? Where would a bot fit in? Would it be a cognitive bot or would it just be an automatic robot with no cognition?
Ideally, a company will try to attack one use case at a time. If they pick one use case & assign a lead to it who owns that piece of the business, understands the problem, and has people on his side that can be dedicated to working on solving it, then they have the best chance of solving it completely.
The company needs to have people on their end who are dedicated and will work at integrating it and getting it into their current workflow. If they don’t either bring in a really skilled integration team or have people dedicated within their company, it’s not going to work because there’s no one on their side to give it the necessary resources and integrate it into their workflows with minimal process disruption.
RTI: Some mortgage industry companies choose to hire an AI company, but not all AI companies understand the mortgage industry.
Gross: That’s a very good point. You have some AI companies that just make a very broad, horizontal learning engine. Usually it’s nothing more than a classification or clustering agent. And it’s about 2,000 steps removed from automating the loan decisioning process.
Automating home lending is a multi-step process. There is classifying data, extracting data, believing in the correctness of the data, and making sure that all the data adds up. The data is a story about the person’s financial life. It’s chronological, and you have to unwind it. You can’t just throw a general AI bot at it and get the answers you need – the problems are too complex.
There are AI solutions that can help with really general problems like, “does this transaction look fraudulent?” You can train that technology on normal transactions. If a person normally charges something between zero and $100 dollars on a credit card and suddenly there’s a Lamborghini for $289,000, the technology identifies fraud. You can take it out of the box, and you’ll get some kind of result from an AI bot, because such fraudulent purchases are not hard to detect. They’re really simple concepts to learn.
But when you think about what an underwriter does, it can be really, really complicated.
Maybe the applicant’s salary is stable. Maybe it’s semi-stable. Maybe it’s trending up. Maybe it’s trending down. Are they self-employed? What do their assets look like? There are a million things you want to look at carefully, so a horizontal bot that understands nothing about the mortgage space is not going to really play effectively. And it won’t really save you time.
RTI: What happens when companies try to apply general-purpose AI technology to the needs of the mortgage industry?
Gross: You can’t expect general AI solutions to solve these complex knowledge worker problems, such as loan underwriting. To solve such a problem, a bot needs to know what the underwriter does every day. How do they do the work? That kind of information is necessary just to understand what the solution should entail. For example, is it an expert system? Is it rule-based? Is it a black-box machine-learning neural system or deep learning system? Is it a combination? Is it a hybrid? Do I need a certain set of features or preceptors that are not even in place yet?
If you throw a horizontal solution at a nuanced problem, it won’t work at all until somebody actually sits down and works with the knowledge workers that you’re optimizing for, understands their daily work, their processes, what lends itself towards automation and what doesn’t.
The vendor and the client need to sit down together and say, “Which problem do you need automated? Let’s understand it. Let’s see if it can be, and then if it can be, is it fully automated, semi-automated?” You have to go through all that.
RTI: It sounds like you’re adding a lot of time, effort, cost, just to take that AI solution and make it work for the mortgage industry.
Gross: Yes. It’s more cost-effective and efficient to choose a solution developed for this industry. A mortgage bot can understand the mortgage documents automatically and how to extract the data. The solution understands how to check the bot’s work and see whether its data fields are actually correct or not.
These solutions build in cross validation to understand the use cases. They consider which use cases are solvable by a bot and which need a human in the loop. They have to be carefully laid out in a design, much like an architect would build a house. You wouldn’t build a house by just dropping some brick and mortar on your property. It’s not going to self-construct, right?
You need a designer to come and figure out how you’re going to orchestrate the project. Some things, machines can learn. Others, they can’t. So unless you sit down and really spec out the problem and what you think that solution should look like – it may contain learning agents and AI components – you won’t create a solution that all makes sense.
And it should be prebuilt for mortgage. If it’s not prebuilt, then you could lose years of time right there. If you don’t know anything about the mortgage documents, the data fields, that can take years to configure.
RTI: Why should readers think about AI now to help address their mortgage business issues? It seems that as AI has evolved, as it’s matured, it has really come to be able to support intelligent mortgage automation.
Gross: For a long time, AI has been touted as being able to do all these things. I think it’s been heavily oversold, to be honest. Now we’re reaching a critical mass in the mortgage lending industry where AI is not just a fad, and it’s not going to go away. Things are transforming in a major way.
Part of this is the digital transformation story. If you go back 15 years, there was no digital data around a loan. Now, more and more of the data you need for a loan is either digital or you can make it digital. Once you move it to PDF or to electronic, you can really introduce bots and AI to start deciding things automatically.
We’ve also seen a lot of improvements in certain technologies. For example, in OCR technology the error rates continue to decrease. Some of the OCR engines that are online and see billions of examples every day train faster. So OCR technology is getting better.
There’s also been progress in AI technology itself, where we can run different learning agents against each other and see if they converge. This technology can run cross validation and make sure all the data is consistent.
Also, the machine hardware has gotten better. For example, if I want to run multiple parse agents on a document and see if they all agree, that all takes time. But if I run it on a cloud, I can run five servers in tandem and make sure they converge. If the machine tries two different methods and they both give it the same solution, it’s correct. We didn’t have the machine power to even do that five years ago or even three years ago.
RTI: Let’s turn to advice. How can companies choose the right technology for their very specific automation needs?
Gross: The best likelihood of success comes when you pick something that is already close to what you need. Perhaps you define two or three use cases in home lending that you really want to automate. Let your prospective vendors show you demos of what their solution can do out-of-the-box, on their data and maybe even on your data. The closer that is to already giving you ROI and benefits, the more likely you are for success three, four, or six months later.
When you’ve licensed their solution and they’ve integrated it for you, if they can show you clear results out of the box without even optimizing, that’s a good sign. If even at that state, it delivers value, you can definitely assume you’ll be seeing value when it’s fully optimized.
RTI: Any other technology advice that you want to offer?
Gross: More clarity is better when companies are looking for automation.
Some companies identify a process, and they’d love a bot to do it, but they don’t necessarily have a clear set of goals. You have to define that process exactly. Sometimes companies don’t even know what their critical documents are or their critical data fails, and it takes months for that to evolve. So the more the company has a really well-defined notion of one or two critical use cases and the more knowledge they have of the process, the better.
The other consideration is whether there should be a human in the loop or no human in the loop. The greatest benefit is realized when a process can be moved to no human-in-the-loop or what we call “touchless.” Often, to do that, you must break a very large process – which may have three or four parts – into sub-processes, because those sub-processes are more likely to be fully automatable.
There are three sub-problems here: classification, versioning, and data extraction. Then there’s verification of income and the analytics. If you break it out into its components, some of those will be fully automated and will deliver much more value. It helps to reduce the complexity by understanding each sub-process and how automated each sub-process could be.
RTI: Once companies know which use case they want to pursue, how can they pick the right mortgage automation vendor?
Gross: One, check whether that vendor already has a solution close to what you need. If they’re building some AI solution or a OCR solution from whole cloth, it’s going to be a very expensive process. The risk also goes way up. Will they ever be able to model all these complex forms? Will they ever get recognition rates where you’re going to see a clear ROI?
If you choose a vendor without a prebuilt mortgage solution, there’s also more risk that whatever they end up building will not be good enough to really satisfy your operations group. Eventually those team members will say, “This is not helping me. This is slowing me down.”
Second, because there’s a lot of construction in AI, I would also look to see whether this company has a very clear research arm that would help them adapt their technology.
You may have specific requirements, and a very large company isn’t going to tailor their solution to your problems. If you choose a group that is very technical, and particularly if they have a research arm, they can tailor it to fit your specifics. If the out-of-box is a good fit for you, that’s a good indicator, but often you want to optimize the solution.
You need a certain level of recognition and mortgage-specific taxonomy, which is not exactly the out-of-the-box taxonomy. A vendor that is a quick study – one that can learn your environment and can optimize for your use case and your environment – can bring a lot of value.
RTI: Would you also ask to hear about times that they’ve done this for other clients?
Gross: Absolutely. I think that’s really important. Have you done this before? Maybe can you show me your work? Can I talk to somebody that’s satisfied?
RTI: Is there any other advice you’d like to offer for people who are considering automating their mortgage operations?
Gross: I would just say take your time. Haste makes waste here. Be sure you understand what you want to automate. Spend some time being sure you’ve got that first use case or two.
Again, the worst case is the innovation officer who just wants to automate the entire bank and doesn’t get into the nuances of what should be automated. My advice would be to pick problems that you want to solve. Understand them well. Find the stakeholders. Check out the requirements, and then find somebody who looks like they’ve already done it or has capacity to do it.