- 4 months ago
- #docai
- #mortgagetech
- #aiinfinance
- #blend
Blend’s DocAI revolutionizes mortgage and home equity closings by automating document processing, dramatically shortening time to close and boosting customer satisfaction
#DocAI #MortgageTech #AIinFinance #Blend
#DocAI #MortgageTech #AIinFinance #Blend
Category
🤖
TechTranscript
00:00Welcome to HousingWire's Demo Day on Demand.
00:03I'm Alison LaForgia, Managing Editor of HousingWire's Content Studio,
00:06and I'm glad you're joining us.
00:08This is where we spotlight some of the most innovative technology companies
00:12in housing and mortgage, giving you, our audience, a front row seat
00:16to real product demos from the teams building the tools that move our industry forward.
00:22In this session, we're featuring Blend,
00:24a company that is working to solve real-world challenges with smart and scalable solutions.
00:31Stick around after the demo.
00:33Mark and I will dive into a short conversation where I'll ask the questions
00:37that help you connect the dots between Blend's tech and your business needs.
00:42In this Blend demo, Mark will be taking us through Blend's Doc AI.
00:48Blend is the leading digital origination platform for banks, credit unions, and mortgage lenders.
00:53From mortgages to consumer loans to deposit accounts,
00:57Blend helps financial institutions streamline workflows, launch faster,
01:02and deliver standout customer experiences.
01:05And Blend's Doc AI revolutionizes mortgage and home equity closings
01:09by automating document processing, dramatically shortening time to close,
01:14and boosting customer satisfaction.
01:17Mark, sounds interesting.
01:18Take us through the demo of Blend's Doc AI.
01:20Thank you very much, Allison, and thank you all for taking a chance to,
01:24or taking some time to take a look at this with us all today.
01:27So as Allison said, Blend's been in the market for 14-ish years now,
01:32providing leading-edge digital experiences for customers across all of those different product lines
01:39that Allison touched on, whether it's real estate products, consumer lending products,
01:42deposit products, and even banking products.
01:45And we've really embraced the evolution and the development of AI and AI-led technologies
01:52over the last several months.
01:54And it's become our goal to become the trusted AI partner for financial institutions,
01:58helping you all to accelerate revenue, deliver more personalized experiences,
02:02and reduce operational costs.
02:04And I'm going to show you a few examples of how we do those things today.
02:07So as we first started evaluating the role AI could play within the real estate lending life cycle,
02:16we took two very specific lenses to solve that problem.
02:20The first is any application of AI needs to provide very unique and specific differentiation to you
02:27and the experiences you have for your customers.
02:30And the second is it needs to deliver some form of very tangible business velocity in its application.
02:38And so we'll go through a couple of different examples of how we are leveraging AI within our experiences
02:44and working with our customers to deliver on those two specific outcomes.
02:51So the first is taking a look at a very typical use case,
02:56one we've been trying to solve within the industry for, I don't know, 20 years now,
03:00and looking at how generative AI can actually help provide maybe a final answer to this challenge
03:07and at a significantly reduced cost to you as a financial institution
03:13and also deliver outsized value for your customers.
03:16The example I'm talking about is document classification, data extraction,
03:21and then helping to provide the customer with meaningful insights and useful data
03:26about the outcome of the document they just tried to upload.
03:30So what we're looking at here is Blend's borrower home, right?
03:34After your customer has applied for a mortgage or a home equity loan or another financial product,
03:38this is where they keep up to speed on what's going on with their loan,
03:41where they're at in the process,
03:43and if there's anything specific that's needed for them to do.
03:47In this example, our borrower needs to provide some documents.
03:50So they've been asked for a W-2.
03:54Now, we've also added into our technology unique and specific consents
03:59to ensure that your customers understand when, how, and where AI is being used
04:05in order to enhance or elevate their experience
04:09and capture the specific consent to allow for the use of those tools
04:15as a part of their application, right?
04:18This becomes additive to the things like credit consent,
04:22income and employment verification consents that you all are used to putting in front of your customers today.
04:27But at Blend, we take security very, very seriously
04:30and want to ensure that your customers always know when, how, and why their data is being used.
04:37So by providing that consent, I'm now presented with my task, right?
04:42I need to upload a set of W-2s.
04:46So as a customer, we might be taking a picture with our phone,
04:50or I'm going to navigate down to where I've got my W-2s saved on my desktop or on my computer at home,
04:57and I'm going to try to upload a copy of that document.
05:01First thing I'm going to do is upload something objectively wrong.
05:05And what DocAI is now doing is receiving the document using large language models to try to confirm,
05:13do I actually have a W-2?
05:15Is this the right kind of document, right?
05:18What we've seen with a lot of AI tools that have come onto the market more recently is this is generally where they stop.
05:26They've been trying to solve that upfront classification and maybe the data extraction problem,
05:32but they're missing the final mile.
05:34They're not affecting the customer experience, the user experience,
05:39or helping provide insight into what I was supposed to do in the event that I did something incorrect.
05:45So here, I uploaded something that wasn't even a W-2,
05:49and we were able to present that specific guidance to the customer in real time, right?
05:54Let's say I get a tad closer.
05:57I got a W-2, but this wasn't the year that you asked me for.
06:02In this example, we're going to go through the same analysis.
06:06We'll review the document, classify it as the correct type of document,
06:10and then move forward to extracting the data off of the document to validate whether it is actually the correct document
06:18that was presented to the customer.
06:29Now, our outcome message looks a little bit different, right?
06:32We recognize that it's a W-2.
06:34You got the right document.
06:35We're getting better.
06:36But it wasn't the right year.
06:38We were extracting the data from the W-2, developed that analysis,
06:43and then gave very specific and tangible guidance to the customer about what they missed
06:49and what we need them to do next.
06:52Now, in the final example, let's upload the actual correct kind of document.
06:56Let's get a 2023 W-2, and we'll see the same set of analysis take place.
07:00And this time, we've got the correct document, a successful document, but we have some name mismatches.
07:09So we're showing kind of incremental progression of giving the customer all of the right kinds of data
07:16that they need in order to ensure that they're providing you with what you need.
07:21Traditionally, this type of an exchange, whether it's been around W-2s, whether it was on documents like bank statements,
07:30typically required phone calls, required emails, required days of waiting between those different moments in time
07:40where somebody could review a document, figure out what was wrong with it, get back to the customer,
07:44and get the right document.
07:45And that entire sequence is now fully automated in a matter of seconds.
07:51So we're leveraging AI to help facilitate and provide more value in these kinds of document classification
07:58and data extraction use cases, but closing that final mile and ensuring that your customers
08:04are being given the appropriate guidance within their user experience for exactly what they need to be doing
08:12and what the appropriate outcome or next steps are for them to get the outcome that we need.
08:18This is having material impact on document quality, getting the right documents in front of your processing
08:25and underwriting teams the first time, and ensuring that we can drive towards the fastest possible clear to close.
08:33The second example that we're going to take a look at is actually a little bit further along in the lifecycle.
08:40So the other areas that we're looking for our AI tools to become impactful are things like manual or automating
08:49what are traditionally very manual reviews.
08:52Think appraisal review, title review, or even a closing package, right?
08:56You get your final closing package, and then your teams are going through a checklist
09:01of all of the things they need to confirm took place within that process or were done in that package and on that document.
09:11And our tools, we hope, are going to give folks an opportunity to automate a significant part of that manual review.
09:19So in this example, I'm going to take an appraisal that we got back from our appraiser.
09:24I'm going to drop it into our tool.
09:26And what this is doing is taking that appraisal document, right, parsing it apart within the large language model,
09:32and then actually evaluating that document and all of the information that was provided within the appraisal, right?
09:40The comps, the justifications, the reasoning, comparing that to the documented Fannie Mae guidelines for a valid and clean appraisal.
09:49And then what we're doing now is developing what we call consensus.
09:52So we have the ability with large language models and how cost-effective these tools are to be able to run them through or this analysis through multiple different types of models
10:06because different models, Gemini's Flash versus OpenAI, Claude, right, all of these different models typically won't hallucinate about the same things.
10:17And when you run multiple times, you get down to a period of significant statistical significance around the evaluation being consistent and what we expected it to be.
10:30And so we're leveraging this type of consensus modeling, running through the same model multiple times, running through multiple different models,
10:36and then consolidating and analyzing the results to produce an outcome that gets your team's very significant confidence in what was traditionally a very labor-intensive manual review.
10:53So we're applying this not only to these types of appraisal and title review processes,
10:57but we're also looking at how can we do this from a loan closing perspective and dramatically impact your post-closed QC processes
11:05and the amount of time it takes between the receipt of the final closing package review and final sign-off from your investors
11:13so that we can hopefully drive down the time it takes for you to close that file out and minimize your overall hedge costs.
11:23We're looking at being able to drive days out of that time between final closing and getting paid by your investor,
11:31which we think is going to result in material savings from a secondary marketing perspective.
11:38But the output here was an evaluation of that appraisal file against all of those standard Fannie Mae guidelines.
11:46And you have the ability to drill all the way down into what the output was, right?
11:52What data was extracted from the appraisal report?
11:56How was that data evaluated?
11:58And then what outcome did those various models that we ran this appraisal through produce?
12:07So the description accuracy and completeness, we looked at the full document,
12:12and then we're able to output this specific evaluation that traditionally you have a human reading through all the pages of the document,
12:21checking boxes that all of the things match, the square footage matches, county records that a number of rooms match across multiple pages,
12:29that all the photos are present and visible and legible.
12:32And the AI tool completed that exact same analysis that traditionally takes multiple hours in just a matter of a few minutes.
12:41These are, in our opinion, some of the lowest hanging fruit that are available for taking these types of tools
12:48and helping them deliver practical business outcomes in the mortgage origination process.
12:54One of the things we're doing at Blend is applying this framework we call the jobs-to-be-done framework.
13:00And so we're looking at each major stage of the mortgage lifecycle
13:04and evaluating out of all of the steps that need to take place, out of all of the processes that need to be done,
13:10where can we focus time and attention on areas that revolve around manual process, manual document review,
13:19or things where we need to have folks involved in the process where we think technology can have a meaningful impact.
13:27And that's where we're focusing our attention first.
13:30I'm very, very excited to continue to develop these technologies out further.
13:34And roll out more of these capabilities over time for you all.
13:38But if you're interested in learning more, please reach out to your account executives at Blend.
13:43We've got content on our website about these technologies as we're developing them as well.
13:48And would love the opportunity to chat more with you about how we view AI developing in the space
13:53and some of the ways we're leveraging these technologies to solve practical business challenges today.
13:59Thank you all very much.
14:00Mark, I love the capability of taking a manual process and reducing it, especially when you're looking at the closing process.
14:10Closing documents are no joke, right?
14:13So let's take a little bit more into blends.ai and what it's bringing to the table and how it's helped driving the process to clear to close to be much faster.
14:25And like you mentioned, accelerate that decision-making while still delivering a better borrower experience.
14:32So the first question I have for you is, how does Doc.ai ensure that the documents borrowers upload are accurate and complete right at the start of the process?
14:44Yeah, great, great question.
14:45Thank you for asking.
14:47This is kind of the most and one of the most exciting things about this technology, right?
14:52So I drew the parallel at the very beginning of the conversation around traditional optical character recognition kinds of solutions we've been trying to use to solve this problem for 20 plus years now.
15:02And the application of AI in this space means what we don't need to have are massive document libraries trained on thousands of different kinds of samples.
15:15That's constantly maintained as document types shift, document formats change, lines are rearranged, because these AI tools have the ability to take a document in any format and analyze the content of the document to produce a reasonable understanding of what the document is.
15:40And it doesn't require that the data is in the same place.
15:42And it doesn't require that the data is in the same place every time.
15:45And so the example that we looked at earlier when we were uploading things like W-2s and bank statements, bank statements is a notorious one where every financial institution is going to have their own unique variations on how that document is structured.
15:58But you saw our tools able to take that type of a document and still produce the same level of analysis and insight that would traditionally require you to have taken thousands of samples from every FI you could possibly think of to produce that same outcome.
16:13And so we're able to use the AI tools to analyze the content of the document specifically, reason about what the document is, and then use the content of the document to produce insights for the customer about whether it was the right document, whether data matches.
16:32So we also see this having kind of interesting and unique applications for things like catching fraud more quickly, because you'll be able to identify when there's data mismatches between the document that was presented to you versus the application file without needing the traditional stare and compare of a human to raise those red flags for your organization.
16:54Let's talk a little bit more about that capability of recognizing the actual content in what ways does DocAI help reduce errors in closing documents.
17:06And let's just expand on that a little bit further and see and talk about even how that has impacted the overall time to close for lenders who are actually using the solution.
17:17Yeah, great question. So it's interesting, right? You made the comment earlier. Closing packages are intense, potentially hundreds of pages at times.
17:29And the data that ends up on a closing package comes from a wide variety of different places, right?
17:35Your lender's own LOS is producing some of it. The settlement agent is producing some of it.
17:39Somebody getting an email with data from county records is probably just manually typing information into some of it.
17:44So it comes from a wide, wide variety of different places.
17:48And traditionally, the review process is I get those hundreds of pages.
17:52I have a giant checklist that I sit on my desk next to me, and then I scroll through pages on my monitor or I'm flipping through physical pages on my desk and marking items off on my checklist.
18:05Our AI tools allow that entire process to become digitized and fully automated, right?
18:12All of those things that need to happen as part of the process.
18:15Every data point that needs to be confirmed, if we want to validate that the exact same long property description exists on every relevant page within the closing package,
18:27that that description still matches what you had within your loan origination system and consolidating those different data sources and evaluating those two different data sources,
18:36all the way down to were the dates correct, did they match consistently across the package, were there initials in all the places that we expected there to be initials, were any signature lines missed?
18:47These are our really common examples that we rely on people to try to identify and correct and take care of.
18:52Or you've been using thousands of document samples and really expensive management libraries to try to use OCR to solve for some of these same use cases.
19:03And a lot of the same benefits we talked about AI having within traditional document types like W2s, bank statements, and other upfront approval document types, same rules apply here, right?
19:14We're able to take hundreds of pages in a closing package and provide that same level of analysis, that same comparison of data present on the document to source data within your LOS
19:28or expected data or expected data for what we needed out of that document, and then go the final mile, right?
19:35Produce the insight around where there's discrepancies, where there's mismatches, raise those flags and ensure that the people who we need to insert back into the process
19:45have a clear and concrete understanding of what it is they need to do to resolve the flag.
19:52And then in terms of the outcomes that we're seeing at producing, I touched on this a little bit kind of just in the general overview.
20:00But, you know, traditionally, final closing, review of the closing package, getting the findings delivered to your investors five to seven plus days in the traditional closing process,
20:12which represents a pretty material amount of hedge cost and hedge risk, depending on how you're trying to deliver those files, right?
20:20At cash window, it's a little less risky, but if you're trying to deliver in batch or in bulk to get advantageous pricing,
20:26those days slipping can mean material loss of savings within your overall bulk delivery.
20:33So we're seeing kind of early leading indicators are in the high hundreds in terms of savings from a hedge cost perspective.
20:41But depending on the scope and size of your delivery, it's starting to see savings in the thousands becoming pretty regular.
20:51So let's dig into that secondary process a little bit.
20:55I know that you mentioned it at the very end of your demo and in that last answer,
20:59but how specifically has automating document verification and quality control within Doc AI improved the borrower experience
21:10and the overall customer satisfaction scores that we see on the borrower side?
21:16Yeah, great, great question.
21:18So it's been fun to watch.
21:20So Blend traditionally, and we do this with our customers as part of every business review that we host with our customers,
21:28is review overall application performance, customer feedback.
21:32And one of the things Blend has done for years now is inject net promoter scoring into our application experiences.
21:39So every borrower who comes through the platform is leaving an NPS score at the end of their application experience.
21:46And we surface those scores to our customers.
21:47We're collecting that feedback and using it to constantly improve our application.
21:52But one of the kind of neat outcomes of this is Blend traditionally has had about an 8.8 to 8.9 NPS score.
22:01And now with the addition of these new tools and the addition of this automation,
22:06we're seeing those NPS scores climb even higher, which is a little bit silly considering they were already 8.8 and 8.9 to begin with.
22:12So from a customer satisfaction perspective, it's giving them more confidence in the lender's ability to deliver the speed that they're looking for,
22:23deliver the outcomes that they're looking for.
22:25And for our lenders, it's giving them the ability to reduce reliance on manual process,
22:33on human intervention, and drive their time to close down significantly.
22:38So when they find those customers that are in a competitive purchase situation and they need to move quickly on an approval,
22:45they need to move quickly on closing, putting them in a better position to achieve those outcomes for their customers
22:52because they're not relying on traditional manual processes and manual review of documents in order to get to the same place, right?
23:02So this allows you to redeploy your resources more effectively, reallocate your spend in a way that is more directed towards your customer
23:12and specifically getting more customers through the top of the funnel and improving top line from that perspective.
23:19And then ensuring that your business is running as lean and efficiently as it possibly can.
23:24Given our current rate environment, these are very, very lean times for everybody.
23:28So we think this is a tremendous opportunity to lean into the kind of practical environments that we're all in
23:36and really lean into the opportunities that technology affords to protect your margin and your bottom line
23:46in an environment where, you know, businesses is hard to come by.
23:50Absolutely. Mark, thank you so much for taking us through Blends Doc AI for our audience.
23:57For more information about Doc AI, click the link below.
24:01Yeah. Thank you all very much for the time today.
Be the first to comment