Smarter Lending with AI

Lindley Gooden: Hello, I’m Lindley Gooden. Welcome to our Money Live webinar, Smarter Lending with AI, hosted in association with Hyperscience. Great to have you with us today.

Lending is on the brink of a revolution thanks to AI-driven innovation. Now, that might be through more automated processes, or perhaps it’s through faster credit decisioning. The key question is, how and where should financial institutions invest, and how will different lending products change as a result?

We are here to answer those questions and many more to cover a wide range of topics in the process. We’ll ask how AI and automation can optimize the lending cycle by speeding up document verification and by accelerating decision-making and delivering seamless personalized experiences. We’ll talk about AI workflows that can improve business processes for staff and also for customers. And we’ll look at key learnings from AI-driven lending transformations of the past.

Lots to talk about. Let’s meet the panel who’ll be helping us to get through all of those subjects. First of all, we have Esther Dijkstra. Esther, great to have you with us. How are you today?

Esther Dijkstra: I’m very good, thank you. So I’m Esther Dijkstra. I’m the Managing Director of Intermediaries at Lloyds Banking Group, which means I look after our mortgage intermediary relationships, but also after our housing relationships.

Lindley Gooden: Lovely to have you with us. Next up we have Sarah Stroud. Sarah, who are you and what do you do?

Sarah Stroud: Hello, I’m Sarah Stroud. I head up the mortgage strategy at HSBC, and I focus on our capital utilization, optimization, and returns for retail lending.

Lindley Gooden: Great to have you with us. Thirdly, we have Abdul Qureshi. Great to have you with us. Tell us who you are and what you do.

Abdul Qureshi: Good morning, everyone. My name is Abdul and I’m part of the Barclays Business Banking Leadership Team.

Lindley Gooden: Lovely. And last but certainly not least, we have Chris Bloomfield. Lovely to have you here. Tell us who you are and what you do.

Chris Bloomfield: Hi everyone. I’m Chris Bloomfield, and I am the Field CTO at Hyperscience. We are an AI document hyper-automation platform, and I’m responsible for a customer’s technical journey from pre-sales through to post-sales into value realization.

Lindley Gooden: Great. What a panel. Okay. Over the next hour, we’ll be covering lots of ground, and you may well have burning questions or comments to put to our panelists yourself. It’s your opportunity throughout our time together live to ask your questions. Simply send your question in at any point throughout the session using the question section, which is over there on the right-hand side of the webpage. We’ll get to as many of those as we can towards the last 10 to 15 minutes, but we’ve got lots to do before we get there. Send in your questions and we’d love to hear from you.

Let’s kick off by taking a look at the state of lending. Where is AI being deployed really well at the moment, and what lessons can we learn to launch the next phase of transformation? Let’s not get ahead of ourselves, but Sarah, if you can kick us off. First of all, where is AI and automation already being tested well and working well inside the system?

Sarah Stroud: I would say it’s being used throughout the lending lifecycle. We are using it for upfront targeting and tailoring of offers. We are using it throughout the application process to find efficiencies, and that’s obviously to meet the customer need of getting a quick answer, an accurate answer. We are using it within the backend with Ops. It supports underwriting. It is really throughout. And then through the lifecycle, once that loan is in place, the customer servicing, collections, that kind of thing. It really is throughout the lifecycle.

Lindley Gooden: Where for you at the moment, is it being used well? Where is it really making its mark?

Esther Dijkstra: I think from the customer angle, it’s really tailoring that initial application journey. It’s making sure we’re only asking the information that we absolutely need for that customer to make a decision. We are not asking unnecessary questions. It’s tailoring it, it’s learning to refine based on the answers that the customer is giving us step by step. So pretty smart forms. It’s not wasting customers’ time. It’s making everything a lot more efficient at the front end, and I think that’s what customers are expecting more and more.

Lindley Gooden: Brilliant. So already, Chris, it’s being integrated. I think people watching us will know that wherever they are. But what lessons can we learn out of the applications and the use of it so far? Because so many possible applications, everyone will have their own experiences. I guess we need to bring that together and work out how it’s working well.

Chris Bloomfield: Absolutely. When I think about the most successful AI projects that I’ve personally been involved with or witnessed from other AI vendors out there, it’s really about attaching yourselves to the biggest business priorities that exist. But so often what we see is organizations going completely the other way, and they’re looking at probably the most complex and actually least business-value processes.

In reality, the focus should be on those high-worth and high-business-value business practices. And actually, you might have already begun your automation journey on those, but maybe you’re only getting 50% of the way there. AI can help augment that and actually take you from that 50% nearer to your 80s, your 90 percent, and drive that greater business value.

The other thing I would also say is, start small. There’s always a tendency to overcomplicate things and trying to attach from an end-to-end process. Focus on one micro area and deliver success there. And ultimately, the other thing I would say is include humans as well. You can’t just rely on AI and the outcomes of AI. You need humans in there to validate and ensure that the outcome is accurate.

Lindley Gooden: That’s already a good bit of advice. I think naturally, and we talked about this before we came on today, people might think about those big projects, a transformational project—of course, the word transformation indicates that—but those little additions and improvements, which naturally should be done as part of that process anyway to iterate and improve. That’s where you would focus some of the attention.

Chris Bloomfield: Absolutely, yes.

Lindley Gooden: What about partnerships? Esther, great to have you with us. Where are the successes? Where are the challenges in terms of partnerships? How’s it helping?

Esther Dijkstra: Good question. I definitely think that the first partnership to call out, and you might not expect that, is the partnership with your colleagues, because your colleagues will be really instrumental in making AI work and they need to be on board with the journey. You might think that of any technology transformation, but I think for AI, that’s particularly important.

We have a partnership with Cambridge Spark who have educated our leaders, and then we’ve worked with other fintechs and universities to roll out education to our 65,000 colleagues. So that’s one thing that’s really important.

The second one being in the mortgage area, I wanted to call out a partnership with Synectics that is around mortgage fraud prevention. Really important area to do the work in as well. And finally, our engineers are using Google Cloud Vertex AI to use their guardrails to test and learn about AI.

Lindley Gooden: Can I ask you just an additional question? We all know that partnerships are forged through good relationships and through shared values, right? What’s the gold dust that makes these partnerships work? Is it just having a similar vision and a similar outcome? Because one of the things that we need to work out as well as what we’re doing is why we’re doing it and what we’re trying to achieve for the customer. So what’s that thing that makes those partnerships work?

Esther Dijkstra: I do think having aligned outcomes really helps. So what is it that you’re really focusing on and what are you trying to achieve? Because getting that crisp and clear is really important. However, I think a second point I want to call out is trust. Because in this process, there needs to be trust in terms of capabilities, but also trusting you’re sharing data if it’s aligned to your business strategy. You do need to build on those. I think those two elements are really important.

Lindley Gooden: It’s important to put that upfront because those relationships are going to make this work, however it works in an organization. Abdul, great to have you with us too. Let’s get into the details slightly. How are we reducing manual workloads through end-to-end automation? It seems like you’re going too far, too fast, but how do you start to transform in terms of the way you work?

Abdul Qureshi: Thank you. If I may, can I take a step back in terms of framing how the banks and the borrowers are going to think about this sort of end-to-end journey? Essentially, there’s two sides to it. One side is what the customer experiences with the customer interface. The other side is what happens within the bank. Now, the customer interface we’ve touched upon with chatbots; better human language processing is improving, but it’s only as good as what happens on the bank side.

Broadly, if I frame what happens on the bank side to simplify it, essentially three things happen. Number one is Knowing Your Customer. And Esther kind of touched upon that. I think that is actually not where a lot of the pain point or friction is at the moment, because that’s pretty much being done. Most banks and financial institutions have automated that, and AI is already playing a big part, constantly improving.

If I look at the second and the third components, they are essentially affordability—can the customer or the business afford the lending or the borrowing they need? What are the different circumstances? What’s the right product fit and all the regulatory compass there. And last but not least is the credit worthiness.

And again, if you think about more complex areas like business lending, there would be a sector-specific context. So what’s going to happen in commercial real estate? Are offices going to be impacted by people coming back into workspace? There are those macro factors as well as very micro factors which come back down to the individual business. So across those pieces, there’s huge amounts of opportunity because the ability for AI to not just process the structured data, but also to really make sense of the unstructured data, the way humans have done in the past.

To conclude, the way I see it is once we really simplify those aspects of the journey internally within institutions, I think then the ability for our humans, our colleagues, to be augmented with that power of AI is what’s going to be transformational for the customer experience.

Lindley Gooden: That sounds like exactly the right approach. I mean, I know that in the world, and we see in the news that the number of AI applications and the number of failed AI applications is high. But taking that cause and effect, you really look at the problem and you don’t rush it through. You don’t just throw money at lots of potential AI models. You actually go through and say, “Use case, use case, what could we do?” and change it and update it. That seems to be a very sensible approach.

Chris, what about complex documentation? So how can generative AI help to sift through, especially the complex stuff?

Chris Bloomfield: It’s a really interesting question. I always think about the right model for the right job. Now everyone is jumping on the Gen AI bandwagon, and my boss would say, “Actually, you don’t need a helicopter to cross the street.”

So actually focusing in on smaller language models that are curated on your dataset is probably the best place to start when you want to extract key specific pieces of information. Where I think Gen AI really helps is that kind of “large language problem.” Something that is highly semantic. It requires a lot of cognitive thinking. And then combining that and having the ability to orchestrate the mixture of the smaller focused or narrow language models with your Gen AI larger language models that can drive those outcomes.

But Esther mentioned trust, and that is imperative. You need to be able to trust the output of each of those models. And I think right now with Gen AI, there’s still a lot of lees around that output and how it’s achieved.

Lindley Gooden: I would love a helicopter to cross the street. What a great thing to do. Let’s talk about AI and automation and workflows. How is the industry succeeding in starting to integrate those together? I think we’re hearing some examples of it in some ways, but how are we doing as an industry?

Sarah Stroud: I guess banking industry is built on trust. We’re talking a lot about trust. We need to trust the sources. We need to be managing the risk. That’s I think the number one role that the banks play. The first step in managing risk is understanding what that risk is and scaling up based on the level of risk within the operation and within that workflow.

So if you are looking at lower risk interactions with customers such as FAQs and you can have chatbots, that’s kind of an easy step into that world. Higher risk activities—the underwriting, the fraud detection—you need to be building up that trust with your models and building on to then use those models to drive more efficiencies in the more complex and higher risk activities. But it’s certainly possible.

Lindley Gooden: Presumably the scale is very important there as well. The amount of data that you have to take in. So at scale, absolutely, the use case is clear. But then you’ve got to be careful with the high risk where the customer might say, “That experience didn’t go well, I’m going somewhere else for my services.”

Sarah Stroud: Yeah, and it’s the learning that you do within your bank, within the wider industry. And the building of the trust.

Lindley Gooden: Talking about customers, Esther, what does all this mean for customers? Are we seeing the effects of it yet? Will we soon? And how will we feel about it?

Esther Dijkstra: I would like to break that down in three ways, because as I set out at the beginning, your colleagues are really important in that as well. And I think where it’s used fast and who will notice it the most are colleagues. And to give a specific example, we’ve built a Gen AI model for our internal colleague policies. Because we are a bank with a 250-year history of all sorts of policies. And now it’s just really easy to ask via model any questions that you have, which of course makes that colleague perform better, which ultimately will service customers.

Lindley Gooden: No more 500-page manuals.

Esther Dijkstra: Exactly. And you don’t have to rewrite them all, so that saves a lot of work as well. Also in that, our engineers use it to translate legacy codes. As you know, big bank, lots of legacy systems. That makes them work more efficiently as well so they can focus on the new stuff.

I think then the second area is where colleagues service customers. So we’re using that Athena knowledge-based system just to help those colleagues perform journeys quicker, easier, more efficient, more effective, which you will notice as a customer when you go through any of our lending processes.

And then finally, I would say currently we’re more in copilot mode than autopilot mode. So it is still supporting colleagues, but there are some areas where you can see we start to get into customers dealing directly with it, like ID&V. Because many customers will know, but also brokers and other intermediaries, that quite often has to happen a lot of times through the process. And we’ve already in mortgages done some tests where we can reduce it to just having to do that once, and then take that data through.

Lindley Gooden: Before we came on today, I asked the “stupid question” which I love to ask, about when will customers really see the full effect of these tools in action on the surface? And I think Sarah, you kind of answered that in some ways. It is that risk in terms of the type of service you’re offering using AI tools, bots, and so on. And the scale. And so the copilot mode sounds interesting. It’s simplifying for staff and then the copilot, and perhaps eventually when the risk is right, you’ll see the effects of that in the real world customers.

Well, let’s speed through. That gives us a really good sense of the state of play in AI transformation, including the benefits that it’s delivering so far. Also, the challenges that the industry faces, I think they’re pretty evident. Let’s now look towards the future and think about how AI and automation will improve the lending process. Let’s get granular for a moment and think about those individual pillars. Ab, first of all, what common delays and pain points are borrowers facing every day of the week?

Abdul Qureshi: Excellent question. Can I take a step back in terms of framing how the banks and the borrowers are going to think about this sort of end-to-end journey? Essentially, there’s two sides to it. One side is what the customer experiences with the customer interface. The other side is what happens within the bank. Now, the customer interface we’ve touched upon with chatbots, better human language processing is improving, but it’s only as good as what happens on the bank side.

Broadly, if I frame what happens on the bank side to simplify it, essentially three things happen. Number one is Knowing Your Customer. And Esther kind of touched upon that. I think that is actually not where a lot of the pain point or friction is at the moment, because that’s pretty much being done. Most banks and financial institutions have automated that, and AI is already playing a big part, constantly improving.

If I look at the second and the third components, they are essentially affordability—can the customer or the business afford the lending or the borrowing they need? What are the different circumstances? What’s the right product fit? And all the regulatory compass there. And last but not least is the credit worthiness. And again, if you think about more complex areas like business lending, there would be a sector-specific context. So what’s going to happen in commercial real estate? Are offices going to be impacted by people coming back into workspace? There are those macro factors as well as very micro factors which come back down to the individual business. So across those pieces, there’s huge amounts of opportunity because the ability for AI to not just process the structured data, but also to really make sense of the unstructured data, the way humans have done in the past.

To conclude, the way I see it is once we really simplify those aspects of the journey internally within institutions, I think then the ability for our humans, our colleagues, to be augmented with that power of AI is what’s going to be transformational for the customer experience.

Lindley Gooden: That sounds like exactly the right approach. I mean, I know that in the world, and we see in the news that the number of AI applications and the number of failed AI applications is high. But taking that cause and effect, you really look at the problem and you don’t rush it through. You don’t just throw money at lots of potential AI models. You actually go through and say, “Use case, use case, what could we do?” and change it and update it. That seems to be a very sensible approach.

Chris, what about applications from customers for lending? How can we make that easier and reduce drop-off rates?

Chris Bloomfield: I think Abdul mentioned there around friction. And it’s having a look at that end-to-end process and identify where those friction points are. Now typically the focus—and we were speaking off-air about this—that applying automation technologies or Gen AI is typically where the dollars are being thrown, at the front office. The reality is actually the big business benefit is in the back office.

So if I look at the friction points when it comes to applications, that’s fine. You might have a lovely web form, I can fill that in, that’s great. But there’s all this support and evidence that I need to provide in order for you guys to be able to make a decision as to creditworthiness, et cetera. So how do we get data out of those kind of analog documents at the highest levels of accuracy? So what we tend to find is we focus on these specific friction points.

We’re working with a large Dutch bank for business applications. On average it was about 30 days. So there’s a huge dropoff rate, I think about 60%, simply because of the manual process around validating different directors, et cetera. And also in mortgage and income verification, working with Australia’s largest bank right now. We’ve halved their process time, and funny enough, we’ve actually reversed their NPS. They had the worst NPS, so it was actually negative NPS, to now the highest NPS in the Australian banking industry.

Sarah Stroud: And I think to build on the income validation, because coming from the mortgages sector, that’s such an important part. And when Open Banking came along, a lot of lenders, including us, were all about automated income validation because it takes roughly eight to 10 days for those mortgage customers to get the payslips, et cetera. And it’s really important because customers value certainty. When you are doing the biggest purchase probably of your life, you want that certainty. Can I get that mortgage? And that’s such an important piece.

But we found it hard to do automated income validation because you’ll be unpleasantly surprised how complex actually a payslip is. I think an average NHS payslip has like 18 entries. So it’s not easy stuff to do, and therefore AI can really help to change those. But still some way to go because there are complexities.

Lindley Gooden: Absolutely agree. It’s a good example of streamlining, Esther, but what you also are doing by using an AI tool effectively like that is offering the all the permutations and combinations of possible payslips can be dealt with in that way. I guess, but you have to be really, really… you have to train that model so well. So I wonder, are there any challenges around that? Are there any things that can cause a problem? Because it needs to be really reliable, especially as you say, with the biggest purchase you’ll make certainly today.

Esther Dijkstra: Absolutely. You can’t, and it’s like Sarah said, as a bank you have the responsibility for the risk, but also in terms of mortgage lending, you are the lender of last resort. So you have to make sure that you either have the models trained or that you pick certain parts to improve already.

Abdul Qureshi: Absolutely. And just to build on that, completely agree. I know many of the viewers would be thinking about the ethics and the bias in AI. We all know there are biases because AI is essentially as good as the data has been trained on. So I think it’s hugely important for us to have the right guardrails and the right kind of human expertise complimenting some of those processes.

I don’t think anyone would advocate like a black box end-to-end. To lean on the helicopter analogy, nobody wants to be teleported across the street. And I think we just need to take a step at a time and make sure the right checks and balances are in place.

Lindley Gooden: It seems as though again, we’re looking at the risk element here, aren’t we? So I suppose it’s just a sharp focus and having the entire team, Sarah, behind that sharp focus. It’s something that we weren’t planning to talk about in terms of the communication between those teams to make sure that everybody has a similar focus. That must be really important too.

Sarah Stroud: Yeah, I’d say so. And I think as you know, the machines learn and they do continue to learn. Those exceptions will exist, but they’ll become more nuanced because the exception of yesterday will become learned by the machine today. That’s the way the progress will be made.

But there’s always going to need to be the human interaction with the machine, if we can put it that way, to steer it in the right way and make sure that the outcomes are as the bank would expect. And that’s to protect obviously the customer, as we say. Massive purchase. The customer is taking on that risk of the borrowing, but also from the bank side, from fraud protection and things like that. I mean, you can have the document scanning, but you can also use now AI to verify that that document is real. So there’s really strong enhancements being made on the fraud side as well.

Lindley Gooden: Also, personalization. We need to look at personalization and of course, Gen AI is going to help you to personalize this. Is this actually one element of personalization? How can it, when everything’s really ready, help the personalization, the “P word” that everyone’s been chasing for such a long time?

Esther Dijkstra: That’s where we’ve done it in quite a blunt way before. You know, we’ve put customers into “you are a first-time buyer,” “you are a remortgager.” And I think that’s how we start. But that will also get more refined. So you can have a full range of first-time buyers. Some might be more financially savvy than others. Some might need a lot more hand-holding. And that can be picked up by AI. They can see that that’s the tone that the customer’s setting and respond in a similar tone and take them to more information pages, that style of thing.

And I think that’s something that is more exciting in a way, that you can match the customer tone. So if there’s a transactor coming in, they’re buy-to-let, they’re a savvy borrower, all they want is to get a really quick answer and get the process done as quickly as possible ’cause it’s part of their everyday understanding. The system can pick up on that and just be like, “Right, we’ll push you straight through.” I think that’s what’s from the customer side is the coolest.

Lindley Gooden: The coolest, as long as it’s favorable and the rates are good. That’s where I think the point we made about risk I think was interesting to think about. That’s going back to my very first point of bringing colleagues on a journey, educating them, because risk functions but also colleagues in the business who are ultimately responsible are the front line. They need to shift and learn and adapt to what risks are you willing to take when you have those models in.

Sarah Stroud: In mortgages as well, there is the element of the solicitor, which can take a long time. And that’s where things break down. So trying to avoid those breakdowns as well. Abdul?

Abdul Qureshi: Yes, just building on that. I think the point is well made. I personally don’t think that the holy grail or the ultimate end state should be a segment of one that we once used to think about in our business schools. I think where building on the point made, if you think about sectors like business lending, unlike the personal lending where there are relatively simplistic segmentations, there are hundreds of types of businesses operating in geographies, some internationally trading, some not. And I think the segmentation can be much better refined and it enables banks to think about those businesses in that particular niche in a much better way. And I think that’s where it can be transformational. And we know the SME sector drives the economy, so I think there’s a huge connection to the national growth agenda as well.

Lindley Gooden: And also, one thing that’s often missing—and this would be very useful in terms of Gen AI—I run an SME and I would love it if that service from whichever lender I have would really help me to go to where I want to go. My segmentation today is not where I want to be tomorrow. And sometimes actually I think that’s forgotten. It’s just “have a loan.”

But let’s talk about carrying with ongoing support. Maybe we’re in this space now, Abdul. How can all of these systems help with ongoing customer support? Because I want help. I need it from you. If I don’t get it from you today, I’m going somewhere else tomorrow.

Abdul Qureshi: Great question. So look, I mean, this is an ongoing process. If I go back to my analogy earlier, I think the ability and the tech to be able to converse with customers in a human language is there today, which does help because it’s 24/7. Tech never sleeps. And it sort of gets through that challenge of SMEs as an example, are very busy people. They’re entrepreneurs looking to disrupt things, build things, run a business, so they’re not there nine to five. So I think that’s a big enabler.

However, it’s going to be only as effective as the data that’s being fed into it. So the real opportunity is to really either reimagine or reset all the backend processes which happen within the organizations to feed that customer interface to truly transform that experience.

Lindley Gooden: Chris, you seem to agree.

Chris Bloomfield: Yeah, I totally agree with that viewpoint that this is all about a data estate that is built on accurate data. I mentioned earlier that you’re getting all of these documents flowing in as part of your underlying business process, but they’re in analog form. So it’s all well and good having digital intake. But how do I know (A) that actually the document is real in the first place? Then (B) that the actual data I am extracting from these documents is real? Once you have that accurate data estate, then everything becomes a data problem. And then you can segment off of the back of that.

Lindley Gooden: Brilliant. Let’s talk about security compliance for a second. I’ve been in this very room watching an AI model look at multiple millions of variations of a driving license to onboard correctly. How are you seeing Gen AI helping to reduce hassle, reduce friction, improve the detection of things like fraud?

Esther Dijkstra: It’s a lot of what we talked about already. So we use it for fraud, for example, because it can easily compare documents to see in terms of mortgage fraud in particular, where are the differences? Is that correct? Should that not be? So that’s really being deployed already.

I also think one of the points that we haven’t made yet is a lot of the lending comes from the big high street banks who all have lots of legacy systems, which will drive a lot of that complexity. I’m not saying that’s the sole reason for the complexity, but quite often it will be. And I think using AI to reduce that legacy estate to simplify it will usually benefit customers.

So there’s almost that tension of making sure on the one hand side you will make the journeys, you could argue a bit more complex because it personalizes maybe not to the segment of one, but it will adapt to where it needs to be. But on the other hand, to make sure that that backend is simple, effective, you can use the tools to make sure that you simplify your legacy estates, simplify the products and propositions, and therefore make it easier to get the clean datasets. Although that’s probably in some cases easier said than done.

Lindley Gooden: Would agree. But of course, it is always an arms race in terms of compliance and fraud and so on. Chris, I wonder, could you take us under the bonnet slightly just to show us how we can start to reduce fraud? Is there anything we can, a little glimmer that for the uninitiated, we can see how you can identify these guys?

Chris Bloomfield: There’s a lot to unpack there. I think about Gen AI, “the Lord giveth and taketh away” in equal measures. So we look at GPT-4o now. Anyone could go online and generate fake documents now and gone are the days where I put my thumb over a piece of pertinent information and take a photo of that. These are truly realistic. So to the human eye, we cannot differentiate between what’s real and what’s fake.

So we’ve spoken earlier about AI. It’s really examining the visuals and trying to understand what’s actually happening underneath the hood. So there’s a lot of focus in the AI industry around document forensics. And really understanding, has this been created before? Do we see similar signatures? It’s actually almost a branch of cybersecurity now. So we have these kind of listening posts that are out there globally being able to ascertain “Hey, I’ve seen this before” or “This looks like this comes from this particular data farm.” We definitely see the data. There’s subtle signals in there.

There are also kind of business rules that you can apply, very simple sort of metadata about a document, and you can understand whether that metadata’s been changed as well. So a couple of those different approaches.

And then once we actually ascertain whether it’s real, whether it’s fake, then you want to start examining the underlying data when you extract that for trend analysis. How many times has this particular person claimed before, or applied for a loan before? How many times have they been rejected? It’s really fascinating.

Lindley Gooden: While we’re talking about moving on speed and accuracy as well, Sarah, how do we really engage these tools to make sure that speed and accuracy are dealt with? I think accuracy we’re starting to deal with, but speed is also dealt with.

Sarah Stroud: Yeah. I mean, I think on the accuracy side of it, in decision making in particular, I think it’s got to be run in parallel really. Because you’ve got manual systems and manual processes that have worked for you in the past. So you don’t wanna just sort of walk away from those straight away. You wanna be running an automated process in parallel for some time to make sure you’re getting the outcomes you’re expecting. And if not, and chances are you’re going to need to be fine-tuning as you go.

And it’s not trying to do the whole process in one go. A mortgage application process is very long, as I’m sure many people are aware. So there’s lots of parts to break down. Start with the lower risk side parts of it and get those speeded up. ‘Cause there are plenty of opportunities within a mortgage application process where a system can just do that without absolutely moving fast.

Lindley Gooden: And in terms of credit risk models, Abdul, how can we improve the accuracy and just make the process a lot more smooth, not only for the borrower but also for the lender?

Abdul Qureshi: I mean, it’s a huge topic. I guess if again we think about the credit risk models, there is a regulatory side to it. So the big capital models and how they impact the way banks can operate. Which again is a whole separate webinar probably.

Lindley Gooden: Every answer today could be another one!

Abdul Qureshi: Absolutely. I know there is thinking happening and work happening in terms of how the capital models that the regulators ask banks to operate can be improved and automated and fundamentally reset. But we’ll park that for now.

So coming back internally, building on my panelists’ comments, there are a number of things we can do. So we touched upon credit worthiness earlier. I think the ability for AI to take many more data points and corroborate and build a 360 picture of the client. Now imagine as an example, we touched briefly upon some of the commercial real estate. What we’ve seen in the commercial real estate market is there’s a geographical side, there is a human societal side in terms of are people coming into city centers? Are we coming back in? Are we still working from home, et cetera? There are macroeconomic factors. There are multiple factors which help the lenders build a complete picture of the risks involved in a certain asset class or with a certain proprietor. And I think that’s where the opportunity is. And it can be completely transformational. But it’s gonna take some time.

Esther Dijkstra: Just to add on that, I think that is also the customer’s expectation. We get a lot of feedback that customers banked with us for a very long time, and we’ve got a lot of their history. We’ve got so much data about that individual, but we’re not necessarily looking at it when they then come to us for a loan. And the customer is expecting us to; you know, they would say, “But you know me.” And you have all this information, you see my salary coming in. But actually when we’re doing, in the olden days, that mortgage application, we’re only looking at the data that they’re giving us on that day.

Lindley Gooden: How end-to-end can we make it?

Chris Bloomfield: Well, that brings me to three pillars, basically. Data quality, data architecture, and data governance. You actually have to do all three really well, and not just from a centralized function somewhere, but this really needs to be in the hearts and minds of every colleague. Because that’s the only way you can ensure that you do achieve that data quality, that your governance is there, that your oversight is there, et cetera. And your architecture of what are your golden sources.

And that sounds sometimes simple, but we found some real practical examples. For example, with our mortgage brokers, where you would think establishing what is the business address of that broker should be fairly easy. But you then go, “Oh, do you use Companies House? Do you use the FCA website? Do you use the day-to-day input?” What about their network? There are various sources, so many sources, and then various data quality issues.

Lindley Gooden: You are redeploying your staff that they have to work in different areas because data governance becomes much more important in some ways.

Esther Dijkstra: It does. And what we did is give colleagues the opportunity to relearn, sort of educate themselves, but also recruited lots of data scientists, et cetera. So you have that transition period where, as we said before, quite often you have to run things in parallel. So there is a lot that we ask of our colleagues in terms of upskilling themselves as well as still running your main business processes and in testing out new ones.

Lindley Gooden: Final question before we move on to our third section, Sarah. How end-to-end can we make this? Ideally we can, but of course there are still silos. There are still people, there are still ways of working, there are still lack of understanding of the systems and how they change. And I think that’s why it’s taking a while to get there.

Sarah Stroud: There are a lot of third parties involved in buying property. It’s not just us in the banks that are trying to streamline and automate. There are conveyancers there. Land registry, solicitors. There’s so many. Not very connected. And I think that’s the gap really. There’s this huge ecosystem that the customer is depending on, and actually a lot of the time, the customers don’t know where the delays are or where even the application is sitting. So I think more transparency around that, more connectivity, but all having to be extremely secure and safe and making sure that customer data is safe. This is really key.

Lindley Gooden: Yeah, absolutely.

Sarah Stroud: And it’s not easy and that’s why it’s not yet in place. But I think there’s been massive inroads recently. I’ve heard of some really exciting things that are happening.

Lindley Gooden: Again, communication would be great. But I think if we know where we are in the process, we can start to put that to one side and get on with our day. So those things are really important too.

Okay, we have to keep moving on. We’re about 10 minutes away from your questions, so please do keep sending them in on the right-hand side of the webpage while we are live, and we will try to answer them later on. Now, let’s think about getting the foundations right to take advantage of all of this potential. We’re gonna backtrack slightly through our final section. Ease of integration, Abdul, is so important. How can we, if we haven’t done it already, start to think about integrating AI into our systems, into our ways of working, even if it’s just something we haven’t done yet?

Abdul Qureshi: Great question. So if I may split it into three things. Firstly, going back to the data point. The analogy I would use is historically we have trained our colleagues, our staff, on a number of training materials and built expertise to deliver a customer experience. The equivalent of that today is the AI. So when we think about data, that’s like training the AI. Quite literally. So that’s how we should be thinking about that. And hence all the good stuff that Esther mentioned earlier is absolutely critical.

The other side of integration then is how can we complement existing processes? We touched upon some of those examples earlier. So taking some elements of the journey as it is today, applying AI to smooth out that process to enable our humans within existing processes, that is a lot of opportunity.

The third part is more transformational. That’s more radical. So there is the good old debate of whether AI will eat SaaS. Now, we touched upon some of the middleware and complex legacy systems. Inevitably in parallel, our CTO functions are looking at how do we reimagine our technology estate. How do we make sure that we use AI to go from a great AI-native data architecture straight into the customer experience? ‘Cause the ability for AI to essentially render useless some of the middleware is there, but that’s gonna take time. And that would be a transformation happening on the side.

Lindley Gooden: From a CTO’s point of view, what do you think, Chris? I mean, that sounds very reasonable. I wonder as well, something that came to mind, the old Google model where you spend one day a week working on your personal projects, which potentially is still owned by Google, but you do that, siphoning off some of the effort and the potential of everybody working in the business into feeding a new way of using AI or a new set of models. That would be quite interesting. Just to make that part of my day would be quite interesting.

Chris Bloomfield: Absolutely. Agreed. And wholeheartedly agree. It’s all around that kind of the data estate and ensuring that you can use these models and use them right. But also repeatable usage as well. So does it only exist in this part of the process, or can actually take that and use that in different parts of the process and scale that as well?

Lindley Gooden: I mentioned to Esther about are you deploying people differently? Sarah, is there an element of that that we need to look at now if we’re looking at the foundations before we go into audience questions? Yes, retraining is always part of it, but the appetite may not always be there. It’s a complicated picture to get people on side with you.

Sarah Stroud: It is, yeah. And it’s important to bring your colleagues along the journey and get them involved in whatever new developments you’re bringing in. And I think for us, going back to the mortgage interaction with the customer side of things, it’s really about getting those specialist staff, training them up on the exception side of things. I mean, we do a lot of international lending and it’s those finer skills, I guess, around cross-border regulation. It’s the specialisms, and getting those staff trained up fully in that.

Lindley Gooden: Well, look, I think it’s a deeper cultural thing as well, because if you think of a lot of people in financial services, but even in the tech industry, I think what defines you is your expertise, and that is effectively data in it. So you need to almost give them the confidence that it’s okay, that they won’t lose their identity or their jobs by actually working with it. Because it is quite nerve-wracking for people when you think of it, when you could see something being automated. But you do need their expertise.

And I think it’s then lifting it to the next of “would you redeploy the colleagues by training them, or would you say that frees them up to actually do added value things?” Whenever we talked about the customer interactions and that we’re still using a lot of humans in the loop, as we call that, I think there will always be a role for humans in that. And it’s just finding out where is that. Because even with the most perfect fraud detections, you probably still ultimately want to have someone occasionally validate that there is a real person at the other end.

Abdul Qureshi: It becomes really a wealth of context.

Lindley Gooden: Yes. The people may not be sharing… their skills may not every day be using the same thing. I mean, the phrase that you always hear is “it’s scary, isn’t it?” So reduce the fear and make it possible is probably quite an important thing to do. But all that wealth and experience, those 20 years doing deals in that area could very well be the defining character of how you use that system in the future. But it takes a bit of communication on that.

Chris Bloomfield: So I was just gonna comment really about the human-in-the-loop piece. I guess it all depends on where you’re applying that human in the loop in terms of the end-to-end business process. So later on, when you’re making important decisions, and you spoke about the 360 view as well, you need somebody that really understands that backend process and the confines of which to make a decision. But maybe earlier on in the process where I need to understand what is this document, where is the interesting data located, what’s it actually say there? Do you need somebody supremely skilled in that underlying business process? So then you might have a different profile of individual at that point as well.

Lindley Gooden: One of the most important things, and we talked about decision making in the very first link, but how do lenders guarantee that their decisions are explainable, justifiable, especially with potentially a black box in the middle?

Chris Bloomfield: Yeah. So a great question, and I’m kind of [looking through the] lens—sorry for the pun—around KYC, so “Know Your Customer,” but I think about it in terms of “Know Your Model.” So understanding what am I using actually from an AI perspective in my business process? Surfacing up confidence is gonna be a key one. So how did that model make a specific decision?

And by the way, do not confuse confidence with accuracy. So I can be highly confident, but I could be highly wrong. And understanding how those outcomes arrive as well. I mentioned there, and you’ve mentioned, human in the loop, so ensuring that you’ve got humans in the loop that can actually take that decision and actually validate that decision. And probably most importantly, certainly in highly regulated industries, is being able to record and log that information and understand it from an end-to-end point of view when a customer rings up and says, “Hey, how did you come to that particular decision?” You’ve got that full audit trail.

Lindley Gooden: Confidence versus accuracy is kind of the heart of bias. It’s the best confirmation bias on the planet. And if you just assume that you are right because you’ve done what you think you should be doing, you may have missed something very important in the process.

Finally, as we go into audience questions in a couple of minutes, what’s missing? What should lenders be doing today, do you think, to make start to get this process working better? Look, as you’ve said, Sarah, right at the start, these tools are integrated throughout the process at the moment. Not necessarily in the customer side, but certainly behind the surface. And I wonder what we could do better? Anything that would come to mind, Esther, first of all?

Esther Dijkstra: I think the thing we’ve talked about a little bit, but maybe not specifically yet, is what customer experience and expectations are. Because particularly in other industries, customers are expecting a different experience. And because we’re in a heavy regulated environment, because we have a lot of complexity, legacy, et cetera, we start more from fixing that, whilst the customer experience is already a lot further on.

And particularly when you overlay the lens of customers really valuing things that de-stress them, because we’re in a highly stressful society at the moment. A lot of risk. And that’s I think where perhaps we should focus more on. Because as you know, it’s a very different experience buying a house as buying a paddling pool. So it’s a real example. Don’t ask me why.

Lindley Gooden: But actually, the experience that I will have as a customer is going to be stressed. It’s going to start that way. There is so much that could go wrong. I don’t care if the paddling pool bursts.

Abdul Qureshi: Just building on that, I completely agree. And I think the intent is there. The technology is there. I think the intellect is there and the expertise is there. I think it requires more collaboration, more active joining up both of the industry. And I think you mentioned earlier the collaboration with fintechs as an example, with tech companies, but most importantly government and the regulator. Because we need to make sure we do this safely. And the current complexities involved can actually put a handbrake on some of this experimentation. But I know regulators are on side. We just need to collectively forge that collaboration ultimately in the interest of the customer.

Lindley Gooden: We may get more ideas as we finish up our session today, but let’s get straight to your audience questions. It’s been great discussion so far. Hopefully you’ve found the insights from our panel really useful so far. AI clearly has a million applications in lending with potentially faster, more accurate, more compliance and personalized services. But it’ll be out with the old processes and in with the new. So lots to do as you know.

Okay. Now it’s time to hear directly from you. Thanks for sending your questions through the webpage so far on the right-hand side. You can keep doing that while we’re live. And before we do get into them, please, if you could just take a very brief moment to rate today’s webinar so far. What did you think about it? There should be a link to the feedback form through a little button just below the video player. Your response is totally confidential, but really useful to us so we can keep on improving what we do here.

So let’s get straight into your questions. We haven’t seen these questions yet, so we’ll do our best to answer them. If you can’t answer it, you beat us, but that’s okay.

Question one: When transforming processes into AI models, how big would you say the data preparation part is?

Well, I would say very. Any thoughts? Sarah, first of all.

Sarah Stroud: Yeah, very. It’s interesting when I’ve worked on transformation programs in the past, it is the data architect that actually is in charge. You know, they’re the ones that you can have the business need, can have the customer experience side, but it’s the data architect that actually ultimately sets it out.

Lindley Gooden: That’s actually quite good. So strictly proofing, if you know who’s dealing, you know how important it is when somebody is leading the process.

Sarah Stroud: That it’s the most important possibly.

Lindley Gooden: I’ve got nothing.

Esther Dijkstra: I think the data, but when you look at what often happens is you look at it internally and that can be a little bit dangerous because as we identify it, particularly from mortgage lending, there is a whole front end and backend, from estate agency right through conveyancing. So I think looking at data in the context of whichever industry you are in and trying to get to standardization is really key as well. But that’s often quite hard.

Chris Bloomfield: So I was going to actually ask a question around this. So one of our biggest experiences and friction points in terms of AI adoption is around the use of customer data in order to train a model. So, naturally speaking, you cannot profit from customer data. How do you handle that and do you actually find that as a pinch point in terms of driving AI adoption?

Lindley Gooden: Sarah, first of all, any thoughts?

Sarah Stroud: So I’m not sure I’m the expert on this, but I would definitely say there’s an awful lot of masking that goes on. So there’s never any way that customer data could then identify back to an individual.

Abdul Qureshi: Happy to build on it. This is a personal view. I think the way this is gonna pan out is much more adoption of small language models, more localized models where customers have the confidence and we can give them the confidence that their data doesn’t leave the bank. I think it’s anonymized, and I think the tech is there today. But I don’t see a scenario where any large financial institution would be prepared to share their proprietary data or their customer data with any form of large language models. It’s just not gonna happen. It shouldn’t happen. It doesn’t need to happen because we can still deliver customer value by bringing tech in, using that to train the localized instances.

Lindley Gooden: Now this question is actually for Abdul, so let’s see if it’s gonna be an answerable one. Which part in the underwriting process for corporate lending did you see challenges in using AI? So where in underwriting has AI had challenges? Perhaps they’ve been overcome, perhaps they’re still in progress.

Abdul Qureshi: So I think frankly, across the industry when it comes to SME and corporate lending, we are still at the infancy of that adoption. The way I see it is, where’s the biggest opportunity? The biggest opportunity is to really synthesize and bring together the 360 view of the customer. That’s a key part of how business underwriting, business lending happens today. And I think AI can really fast-track that. So I think that’s where the opportunity is, and equally that’s where the challenge is because those data points that we need to build the 360 are very analog. It could be sector specific, it could be out on the web, it could be the local trading environment, it could be the way the financial statements are published. So I think that’s where the opportunity as well as the challenge is, but I think these are very much surmountable by AI.

Lindley Gooden: And still in the spirit of co-piloting that we talked about earlier on as being a real focus for the use of these tools at the moment, in underwriting, that would be very helpful to have absolutely a very educated potential, which then as an underwriter you can make a decision on, especially with a corporate client.

Okay, Esther, I think you may have seen this one on your screen over here. Let’s give this a go. With plenty of legacy systems, versions of processes, and majority of Know Your Customer data gathered on paper across the years, what did you see as the biggest challenge in implementing Gen AI?

Esther Dijkstra: Tired fingers. I call it data entry. No. So, I think what the challenge is, first of all, there’s actually a lot of opportunity and you get a little bit over-excited maybe in terms of wanting to tackle the biggest thing. And then as we’ve discussed, that’s then hard because you see the complexity. So it’s sort of creating that discipline of making sure you focus on the right projects and also if they don’t work, scrap, move on, next one, because there’s lots to go at in this space.

The second challenge is making sure that you get all your colleagues on board a hundred percent. Because they’re crucial in terms of not just being the expertise in the projects, but we talked about the importance of data quality, the data architecture, et cetera. And you can have the technical standards in place or the policies in place. But you need people to actually live and breathe it and do it every day and be familiar with it.

Lindley Gooden: I don’t wanna leave Chris outta this one. Just briefly, Chris. It comes back to our talk about complex documents and making sure the data’s good. In brief, when it’s a very complex document where there’s a complex process, what should we be focusing on in terms of translating that? Maybe just taking old documents into your Gen AI system?

Chris Bloomfield: Yeah. So I think that if I look at Gen AI, it has been trained on the corpus of publicly available information. And actually, when we’ve got the internal documentation, long-form contracts, all those kind of things, is actually embellishing that with internal systems as well and being able to read that. So in this particular realm, it’s called RAG. So actually allowing Gen AI to use internal data sources in order to be able to help actually assimilate that information and provide those types of insights.

Lindley Gooden: Brilliant. Okay. We have three minutes left. Let’s try and get one more question in. What is your approach to integrating AI into today’s banking processes while simultaneously strategizing and envisioning the fully transformed bank of the future? Okay, making AI work while also thinking of the future in the best possible situation. Esther, any thoughts from that?

Esther Dijkstra: Yeah. I have this mental picture of we are flying in an airplane, and whilst we’re flying, we need to change and fix the engine. And that’s basically what’s happening at the moment, so it’s very hard. But it is very doable because I would want to highlight that we always talk about AI like it’s very recent, but of course it has been around. It started with machine learning, automation, et cetera. So it’s been a long process, and therefore the advantage that we have is already a lot of those processes have been massively improved.

Lindley Gooden: Any thoughts, Sarah?

Sarah Stroud: Yeah, I think just to add, I’m gonna use the word agile. I’ve gotta drop it in somewhere. How have we come this far? We’ve gotta stay agile, you know, we’ve gotta keep reassessing the business priorities and making sure that whatever enhancements we’re making using AI are still meeting what those original business priorities were. And maybe that priority has changed and therefore we need to change the direction that we are pointing the AI at. But it’s staying agile.

Lindley Gooden: We’ve got one minute left. Between Abdul and Chris, what do you think Abdul?

Abdul Qureshi: Well, just gonna build on that. The way I look at it is there’s two parallel tracks. There is the tactical head and lots of AI potential that we should be experimenting in the safe way as we discussed today. In parallel, when it comes to big strategic transformation programs, we should be thinking AI-native architectures first, spanning data spanning front end. And I think that’s where we will see a parallel transformation.

Lindley Gooden: Chris, we’ve got time. One thing we can do today to keep the business working in the right way and integrating these things correctly.

Chris Bloomfield: Yeah. I think it’s partnering and finding the right partner to work with. You mentioned there about flying airplane, rebuilding the engine and stuff. I think with Gen AI it is still brand new. I think even as a software vendor, we are kind of doing that in parallel, but as long as we’re aligned and we’re aligned to the business value, I think great things can happen.

Lindley Gooden: I prefer helicopters, but a glider is very useful. Okay. Well, I’m afraid that’s all we have time for. I hope you’ve enjoyed the last hour that you coming away with some insights, some ideas and practical advice on how to transform lending using AI and Gen AI. Thanks very much for sending in your questions and sorry that we couldn’t get to them all, but they were really good.

If you do have a couple of minutes before you leave us today, please do fill in the feedback form, which should be available as a little button just below this video to give us your thoughts. We completely value your opinion. It drives what we do in the next session. So please do let us know what you thought. And if you’re still hungry for more insights, whether it’s about lending, AI, tech innovation, or the banking sector more generally, join us at Money Live’s events this year. We have more webinars and reports that you can view online and also a range of very exciting conferences. This autumn. Our North American Summit is coming in September in Chicago, and we have our Nordic banking conference in October in Copenhagen. Also, we have our Payments Europe event in Amsterdam in November.

So lots to choose from, but for now, it’s time to go. Thanks very much to our wonderful panel for being here with us and giving us so many great ideas, and also to you for tuning in and for your questions. Until next time, all the best and take care.