Developing Financial AI Products at Experian with Vijay Mehta, EVP of Global Solutions & Analytics at Experian

00:01:03 Richie Cotton: Hi, Vijay, welcome to the show. Hi, Richie. Thanks for having me. Excellent. I'm looking forward to this and we're a few years into the Tive AI revolution now. Can you talk me through what you think the biggest impact has been on the finance industry so far? Absolutely. 

00:02:00 Vijay Mehta: Yeah.

00:02:00 I think we're now past that the hype cycle in getting into real world impacts. A great example is in engineering, generative AI now drives much more efficiency. Productivity for software engineers and throughout the development lifecycle. So everything from code generation to testing to quality assurance.

00:02:21 It's enabling software engineers to focus on the areas that matter most and broaden their skillset. I would say in addition, gen AI is now starting to impact other parts of organizations finance, legal, customer service. I know when I look internally we've transformed into a tech driven enterprise.

00:02:39 So we now generate over 35% of our global revenue from software and platforms. And I think Gen AI is part of it because it makes it easier for companies to be more focused on technology than ever before. 

00:02:50 Richie Cotton: Okay. So that's interesting that you started with the software engineering impact. Of course.

00:02:54 It's a huge thing. Now in the world of finance, I guess the, traditionally the impact of AI is very much around, traditional machine learning, it's predictive ai. So is that somewhere you think finance institutions should focus, or do you think generative AI is also important? 

00:03:10 Vijay Mehta: I think it's another tool in the toolbox.

00:03:12 So financial services companies will employ a number of different tools, advanced analytics different cloud technology tools to do modeling, decisioning and go to market strategies. So I think gene AI fits right in that toolbox, and it's one that is. Is having a real big impact. When I look internally at Experian, we've been using artificial intelligence, advanced analytics, predictive and generative for over two decades.

00:03:38 And I, I think it's, it's embedded in our culture and in our solutions. When I look at the broader market and financial services customers, I think they're gonna have to pursue where the biggest buck, the biggest bang for their buck is gonna be. And I think generative AI is part of the conversation, but not the entire conversation.

00:03:56 Richie Cotton: Alright, that's fascinating. So you've got a ton of different use cases there and it's interesting that it is both the predictive AI and their generative ai that's really important. Now, I know you've been working on a whole suite of tools that you call Ascend. Can you talk me through what are the different components of this?

00:04:09 Like what have you been building within the Ascend platform? 

00:04:11 Vijay Mehta: Yeah the Ascend platform is what I would consider a one of a kind cloud-based solution. It enables seamless integration of data analytics, artificial intelligence, and decisioning technologies to really streamline the workflow and credit and fraud prevention as well as marketing lifecycle.

00:04:31 So it, it gives customers the tool set to make smarter, faster and more secure decisions. And there really isn't an organization that's able to provide the same breadth of the Ascend platform because as I mentioned, we've been doing this for 20 years. We've been using artificial intelligence and now we're making the tools that we use internally.

00:04:50 Available externally and the innovation isn't stopping. With our latest release of the Experian assistant, we're now taking our gen AI AgTech capabilities and enabling realtime assistance that empower, empower businesses to interact with financial data in a conversational way, making complex analytics more accessible.

00:05:11 Making the model development lifecycle and governance lifecycle easier and more automated. So all of this is powered through the Ascend platform and our AG Agentic AI technology. And the value that we're seeing it's not only deriving valuable insights for our customers, but it's helping them make more informed business decisions faster, saving them time and money.

00:05:31 It's like. Providing them with a 24 7 access to our experts and giving them a, an additional workforce, virtual workforce to help them, reimagine the way that they operate and the way that they think about customer experience. 

00:05:46 Richie Cotton: Just for people who aren't that familiar with say the fraud detection tasks, like what do is, what does it involve? Can you give us examples of some of the tools you're building to solve specific problems? 

00:05:57 Vijay Mehta: Yeah. So recently we announced the integration of MasterCard's identity verification solutions. So a big portion of fraud prevention does come with identity validation and verification, and that'll happen throughout the credit life cycle.

00:06:11 So enabling customers to ensure that they are in fact working with a real person. We see. Synthetic fraud on the rise. So what Experian does with our partner ecosystem as well as the tools that we provide, we use device intelligence and we use other sources of information to ensure that you really are who you say you are.

00:06:34 When you're applying for a loan or a credit card, or it really anywhere throughout the credit life cycle, it, as we see generative AI become more democratized, the bad guys are using the same tools or using these advanced tools. So we have to combat that type of fraud with more advanced tools.

00:06:53 So that's really the forefront of our innovation is using AI at scale to prevent fraud and ensure that. The bad guys don't don't win. 

00:07:01 Richie Cotton: Okay. Yeah, certainly being able to identify, are you, doing business with a real human. That seems like a very important thing. I hadn't heard the term synthetic fraud before.

00:07:10 Is this where a bot is doing fraud on your behalf? 

00:07:13 Vijay Mehta: Yeah synthetic identities, synthetic fraud. So it is exactly that, Richie. It's effectively just using an agent to mimic a person or an individual to create. False transactions and ultimately costs our financial services customers money.

00:07:31 So it's a real problem for the industry and the Ascend platform helps to solve that. 

00:07:36 Richie Cotton: That's fascinating because I think like a lot of people worried about AI taking their job is like even fraudsters are having their jobs taken by bots now. Okay. Cool. I know you've got a lot of different products here and some of them are targeted people who are machine learning experts and some of them are targeted at executives.

00:07:51 Can you talk me through like how you go about thinking about creating products for these different audiences? I guess they've got pretty different requirements. 

00:07:57 Vijay Mehta: Yeah. I think at the end of the day when we. When we look at product design it first of all, it's about jobs to be done and the problem that we're solving for, but the design principles are the same.

00:08:07 Everyone wants a simplify interface. Our no and low code solutions apply both to the CEO as well as a as an advanced data scientist because we've been able to really democratize the access to the tooling and the data through things like natural language processing. So an example is a data scientist wants to go deep.

00:08:28 That person can run a very sophisticated query through our interface, through the Experian assistant in the platform. Or conversely, a business unit manager with little or no coding experience can do something very similar, just using natural language. So we have the option for both types of interfaces.

00:08:48 I, I would say when you go back to just kinda the principles of good design and problem solving, that's the approach that we've taken at Experian and with the platform. And I think, I know, I personally, when I log into an app, I'm very critical of how the interface looks, even though I have a computer science degree and I'm, a techie at heart.

00:09:06 I want it to be easier. And I think we are, we're seeing the same thing with the technology and the data science community. 

00:09:12 Richie Cotton: Absolutely. I have to say, there's so many pieces of where I'm like. I have no idea what idiot designed this, what I wanted to look. Yeah. And I guess design is just a fundamentally hard problem and mainly designing for different audiences.

00:09:23 That's incredibly important to get the interface right. So I like the idea that you've got a sort of low code or no code interface for people who are less technical. Then also if you wanna run code, you are a bit more technical. Then I guess what is it? Like you can run Python scripts in the background or something.

00:09:37 Just you can code in order to get your models running. Yeah. Talk me through what has, what's happening. 

00:09:42 Vijay Mehta: Yeah, no, exactly. One of our solutions, the Ascend Sandbox, is a data aggregation tool that gives customers the ability to work directly in the tools that they want to use to develop new models and analytics.

00:09:55 This is used by hundreds of customers globally. What it does is it allows a data scientist or a business analyst to go in access the data in a safe, secure, compliant way, run the queries, and then, develop the Python scripts in a Jupyter notebook or, another tool that they want to use to create a new model.

00:10:13 We're adding more tooling into it. So as new open source libraries come out, and as AI and generative AI tools become more sophisticated, we continue to add that into the environment, just enabling our customers to really use our data and their data in the safe and secure environment. So the coding process and how we think about the coding process is evolving.

00:10:34 I think there is a ton of opportunity in that space and we're really focused on streamlining how the development. It can be done seamlessly in the financial services space to go from aggregation of data into production deployment and do it quickly. 

00:10:50 Richie Cotton: Okay. So I like the idea of making things go faster.

00:10:53 We talked a bit about having a sort of better user experience. So I'm curious if you can talk a bit more about what the, the goals around having all these, like AI products are like, is improved productivity, your main goal, or is it, are there other business metrics you're targeting? Is like what's the benefit of having these products?

00:11:10 Vijay Mehta: I think productivity is one part of it. Richie, I think there's other parts of it too. When we look at kind of the acceleration of analytics as a whole, everything starts with how you think about data and the use of data. So it's going to first require our customers to look at.

00:11:28 You know how they aggregate the data and then from there it's going to, it's gonna go into the model development pipeline and ultimately what that looks like, I think there I wouldn't say we're solely focused on that, but absolutely. With our Experian assistant, with our Ascend Sandbox, with our Ascend Op solution, we are very much focused on streamlining the process and making our making things easier for our customers to use.

00:11:53 So the problem that we're trying to solve with the job to be done here is how do you take. What a traditional model development lifecycle would be. Could be 15 months or more and do it in days. And we have a lot of proof points now and a lot of customers who have been able to use our system just to do that.

00:12:11 So it is, it's automation, but it's also safety. It's also improving your compliance. It's also improving the way that you think about governance. Freeing up capital for our customers to invest in higher priority things in operational operational efficiency, 

00:12:28 Richie Cotton: speeding up a model, generation lifecycle from months to days.

00:12:32 That sounds. It sounds like a huge win. It's a bold claim as well. So I'm curious what does this new workflow look like that only takes days rather than months and have you got any concrete examples from customers who've done this 

00:12:44 Vijay Mehta: well? Yeah we absolutely do. And I think the acceleration, so what we've seen from our customers, and I'll give you some specific examples.

00:12:52 So first, the breadth of the solution where our platform is leveraged by, 1500 clients globally. Processing 14 million credit reports daily, billions of credit and fraud transactions per year. In the US alone, we have 8,000 registered users and over 12 petabytes of data. So that allows us to go in from maybe the smallest lender to the largest lender in the world and really work on the problems that we need to solve for them.

00:13:20 An example is lender, L-E-N-D-R. It's a specialized FinTech company. They do financing solutions for small businesses. They're using the platform to enhance the overall competitiveness of how they make decisions. So this goes from data aggregation all the way to underwriting and decisioning processing.

00:13:39 They've been able to use the platform to fundamentally reduce the losses that they have and. Doubled their business over the past year. And that's through ai, through advanced technology, and it's through the solution. And then I'll tie it back to our internal solutions as well, because when we've been focused on this, we've been looking at how do you democratize access and give access to the, 22,000 employees that Experian has?

00:14:06 So we first needed to focus on. Education of the masses, which is very important, something we also do with our customers. But then at the same time, really form our risk council and ensure that we've got the safety and the compliance solutions in place. So our A, why we're able to innovate at the pace that we are and use generative AI at scale is because we are agile in the way that we think about risk management, the way that we think about governance and the way that we focus on training and democratization tools.

00:14:36 So we apply that same lesson out to the market and we see the results like lender, where they've been able to drastically streamline the overall processing in their credit life cycle. 

00:14:46 Richie Cotton: Okay. That's wonderful. And you mentioned the word education, one of our favorite subjects on the show. So talk me through what kind of education do you need to provide to your staff?

00:14:55 Around yeah, around AI and around how to do the sort of new AI workflows. 

00:15:00 Vijay Mehta: Yeah, it's it's a big topic and it's always evolving. Education should always evolve. Richie as I think it's one of those cases where you gotta start with the foundation, because you're gonna have a group of people who may not understand the basics of how a prompt works and what the process is going to be.

00:15:16 I actually even go one step before that, which is, here are the things, here are the parameters or the acceptable use of how we're going to use AI as an organization. We created a robust document that every employee needs to go through, and it basically highlights the parameters for us, and I know other companies are doing this too, which says.

00:15:37 This is a system that is used for productivity and efficiency, but realize that there are certain things that we do not allow you to put into the environment. So we enforce that from the front. We educate our staff, but then we also. On the back end, within our platform, we have a set of policy and technology enablers to support that too.

00:15:57 But education is continuous. So we start with acceptable use. We start with how you can prompt successfully. So education on prompting, education on what an agent can and can't do, what an assistant can and can't do. Then we go broader, depending on the area of what department they're in or what area they're focused on, to be more specialized.

00:16:20 So a person in the finance organization who is working with a large amount of data is going to need a bit more advanced training than say, someone who's just in customer service and needs the interaction to be more cookie cutter. So it really is you start with the foundation, Richie, and then you add on based on the specialization.

00:16:39 Richie Cotton: Absolutely. I like the idea of targeting what people learn depending on what the, what their role is. 'cause as you say, if you're in customer service, it's gonna be very different needs to, if you're in a machine learning team. And I do like the idea that a lot of the, just the basics, like this is how you write a good prompt.

00:16:53 These are what the use cases are and these are not feasible use cases. That's incredibly important for everyone to get started with. You mentioned that you've got a big document of policies and guardrails and guidelines. Are you able to share anything about what's in that. 

00:17:07 Vijay Mehta: We are a regulated organization.

00:17:09 So we, we take that very seriously. So it starts with, I would say, the safe use of data and how we think about personally identifiable information, PCI information, and we adhere to all of the regulatory requirements in every country that we operate in. So that is the first part of our policy.

00:17:29 That doesn't apply to every, all of your listeners, but for us. As a data organization, as an analytics and technology company, we know that number one priority is safety of data. So that's, I would say the bulk of the policy. And then from there, there is other acceptable use terms of it.

00:17:47 It's easy for people, and this is a really a, maybe a bit of a tangent, but it's easy for people to start associating with an AI agent and asking it questions. You can go a little bit off topic or go down a rabbit hole. I know I've done it. I'm sure you've probably done it too. So it's making sure that people stay business focused and there's a language around that as well.

00:18:08 So don't. Don't ask it about, medical advice, don't ask it about things like that. But boundaries are primarily focused around acceptable use of data for our organization. And we've had very good luck with that. 

00:18:20 Richie Cotton: Okay. Yeah, certainly. Understanding like what is personally identifiable information, what can you do with this, the respect way ai, that seems like a pretty.

00:18:28 Essential thing for I think almost all organizations. It's interesting that you don't allow personal use cases of gen guy at another, it was a chief data officer on the show recently. He was like we know people are gonna try and plan the whole day for this thing. It's we're okay with that as long as they actually use it to do some work as well.

00:18:45 So I think that's an interesting sort of cultural distinction about like where you draw the boundaries of what you can do with ai. 

00:18:51 Vijay Mehta: And we're pretty flexible on that point. Just to jump in. With Microsoft co copilot, I think nearly every one of our employees has access to copilot and that boundary between your schedule and how you think about the integration with, your day-to-day tools.

00:19:04 We're flexible on it, but when you get into more of this specialized agents that we have and the assistance that we have. I think that's really what I was referring to. 

00:19:12 Richie Cotton: Okay. Yeah, you don't necessarily want to be having like some massive, like workflow swarm of agents doing your own business bidding for personal use cases.

00:19:20 Okay. That makes a lot more sense. Does having more powerful AI tooling, has that changed the skills that you need your employees to have? 

00:19:28 Vijay Mehta: I think it has to some degree. And I think we're still at the early stages of it, as, as an industry, as society. Logically, if you're going to give someone a tool that.

00:19:40 Can be used to do complex tasks that required an advanced degree in the past you're gonna see a shift from the amount of, say, data scientist or technologists that you need. It's going to, it is going to change now. Is natural language processing right now, changing that skill? Diversity? I would say we're, it's starting to I would say it's starting to, it's not, I wouldn't say it's a standard yet, I don't think, I don't think any organization out there is saying, yeah, we're now able to take.

00:20:13 All of our engineers and all of our data scientists and only hire people who are prompting. I think that's hype still. And the reality is, will we see it change? Yes. Because with our Experian assistant, we're able to visibly see the transactions and see how people are interacting and it no longer does it require.

00:20:33 A person with an advanced degree to go in and create a new business int intelligence solution, a new report or even a new model in the environment. So again, it's early stage. We'll have to see how it goes over the next. In the next 10 years. 

00:20:48 Richie Cotton: Okay. So you're not hiring vibe coders just yet, but maybe one day in the future?

00:20:53 Yeah, 

00:20:53 Vijay Mehta: I think that's the, I think that's where all of us are in the industry right now is we know that there is going to be some shifting in skills, and we know that the democratization of natural language processing and generative AI is going to change that dynamic. But I haven't seen anyone who's taking significant action and saying, okay, we're only gonna hire English majors now.

00:21:14 Teach 'em to prompt versus hiring data scientists. The actual, like the data that we have and what we've seen is the people who get the most benefit of this are people who are educated in data science or in Python who really know how to code. They're the ones who are taking the most advantage of generative ai.

00:21:31 And it's the early stage folks who might be right outta school who are still navigating it because you know where we, I would say where we are as an industry is. The assistant process is geared more towards understanding. You get understanding the fundamentals. You understand odd Victorian programming, you understand objective functions and feature engineering already not coming in with a blank slate and saying, I'm now going to be the best feature engineering expert.

00:21:57 I'm gonna now be the best Python expert without at least having the foundation. Again, it's, it'll be a, it's a fun academic exercise to have Richie and one where I think there's gonna be a lot more discussion on it over the coming years. 

00:22:11 Richie Cotton: Absolutely. It is definitely an interesting situation where if you have more skills already than having AI is gonna provide that exercise multiplier.

00:22:19 But if you had, you're learning stuff from scratch, then. AI can, it can sub, it can boost your productivity a little bit, but you're not getting the benefit because you don't know when it's doing stupid stuff. So since you mentioned feature engineering, I was trying to think of what are those trickier, more annoying. Basically the machine learning workflow. And I think feature engineering is one of those things where it requires like technical skills that also requires a lot of domain knowledge to do well. Do you have any tooling or that's gonna make it easier for finance professionals to do feature engineering in the models more easily?

00:22:49 Vijay Mehta: Yeah, absolutely. Traditionally feature engineering is, it's a mix of technical ideas and domain expertise. When we started the process, I guess about 15 years ago, we anticipated that we needed to think about not only the strength of our development capabilities, but also how we innovate and how we're going to structure the process.

00:23:13 And everything really comes down to. Understanding the use case understanding the problem that you wanna solve for and then defining your trajectory after that. And I think that gets lost a little bit in the process sometimes. Our tools, like the AI assistance that we have, Experian assistant for model risk management, the Experian assistant for data.

00:23:35 Like all of those are streamlining the way that process is happening. So again, it's making complex analytics more accessible, giving customers the ability to have 24 7 access to an expert, to to a coding buddy. Who can actually really make things much faster, boost productivity, and actually it creates also a more collaboration too, I would say, because it allows.

00:24:00 Engineers and data scientists to focus on the outputs more than the input. We're not eliminating human in the loop. We're keeping human in the loop there. I would say it is it's going to streamline the entire feature engineering process. We see it full. Fully in all of our tooling and how the way our customers interact.

00:24:20 Though, I think there's still this aspect of human interaction before you're gonna take something into production 

00:24:25 Richie Cotton: that does make a lot of sense that you want a bit of human interaction for. So go straight into production. Actually, maybe on that note, can you talk me through do you have a process for putting things into production that always seems to be like a bit of a sticking point?

00:24:36 Vijay Mehta: Absolutely. So with our Ascend ops solution. We've made a ton of progress, so the combination of our tooling, so we've got, we have our Send Sandbox, which handles all the data aggregation, the linking, the pinning. It allows customers to actually get in and do development of models with our data.

00:24:53 Their data. Third party data as Send Ops is our ML ops pipeline, which allows for the seamless integration and the model development process. It does monitoring explainability, it provides transparency. It uses containers to take a model and move it from one environment to another. And then our Experian assistant is the interface that allows you to interact with all of those environments and make it really easy.

00:25:18 So by providing the tooling, by providing the right pipelines, we're able to actually solve the problem of model deployment and take it from, many months to a matter of hours and do it in a safe way because we still have. The appropriate workflow, the appropriate governance in place, as well as the document automation that needs to happen for anytime you're dealing with model risk management or governance.

00:25:41 Richie Cotton: Okay. So actually a lot of the stuff you mentioned there, stuff like containers model observability, like the deploying tools, like documentation creation stuff I know there are lots of existing tools for this, so I'm curious how much of this have you built in house and how much is just is tools you've purchased for for setting this up?

00:25:58 Vijay Mehta: We built quite a bit in-house, but we've also, we partner with a lot of companies. We have an entire partner ecosystem, so AWS is one of our big partners. So we have worked with them in order to enable a lot of our service-based solutions within the Ascend platform. But there's a host of other partners that we work with.

00:26:16 I mentioned MasterCard earlier. Valid Mind is another company that's helping us. So we have a great relationship with. They have a automated documentation solution, and that is built directly into our Experian assistant, allowing our customers to take away that cumbersome task of generating documentation through the model lifecycle.

00:26:36 And anybody who's worked in a regulated environment knows that is an extremely time consuming task for our data scientists and modelers because, if you're dealing with personally identifiable information. You're, have to deal with the OCC. It's hundreds of pages of documentation that we are able to automate directly through our relationship and our partner ecosystem.

00:27:00 So those are just a couple examples of how we think about the broader Ascend platform and the solution that we can provide. So it, it really is about not only just the analytics side, Richie, because we have streamlined that, but it goes further into the entire life cycle of how you think about originations.

00:27:18 How you think about collections and account management and bringing that all together with fraud prevention throughout the entire workflow and the lifecycle. So very powerful solution I would say, at the end of the day. We believe the integration of all these components, there's, there really is no one else who can do it because we are uniquely positioned, not only with our data assets, our advanced AI and our software, and we brought it all together to address real customer issues.

00:27:46 Richie Cotton: Okay, nice. It just seemed there are a lot of different parts of the lifecycle that can then that are being addressed by better tooling of all some ai. I, because since you've built a lot of these tools now, I'm curious as for other organizations that, that wanna create AI products, like what are the easy things to build?

00:28:02 What are the difficult things? What's not possible yet? What do you do first? 

00:28:05 Vijay Mehta: Yeah. It, and I think it's, where do you begin? Is always the burning question. And I. I just go through the process, right? And I think there's a buy, build analysis, but everything starts with it.

00:28:16 Define a clear use case and your in the math world, your objective function, so what do you wanna really focus on? And then you identify specific measurable problems to solve, whether it's reducing fraud or enhancing your credit life cycle or getting more approvals through your process.

00:28:33 You build a strong data foundation and AI is. Is only as effective as the data that's behind it. You use tools to accelerate that inva innovation, and that's where the conversation comes in on what you really wanna buy and build. With the Ascend platform, we've focused on a number of areas of tooling, but there's still things that our customers need to do themselves.

00:28:54 So how they manage their entire data lake or data ecosystem, that's an area that they need to focus on. They need to have good quality data in order to build good quality models, and that's. That's something that while we can help with our, data quality to some degree, they're gonna need to do some on their own.

00:29:12 I think the mathematics still haven't been automated. When you get right down to defining your objective function and how you look at your features, we're not using natural language processing just for that, or yet for that I should say. I think there is maybe a case in the future. But you still have to have good, a good statistical background when you're building a model.

00:29:32 It's just all of the bits and bobs that sort of enable the workflow. So a customer shouldn't have to focus on building their own ML ops pipeline. That's something that is very easily outsourced. A customer shouldn't focus on, how they aggregate the data. That's something that's easily outsourced.

00:29:50 So those development environments, the tooling, and then the tools for things like strategy developments and decisioning. I would say it this way. So within probabilistic and deterministic decisions that we have in any number of life cycles, so whether it be credit or marketing or fraud prevention the tools that support it, the data that supports it.

00:30:11 Should all, be leveraged through a company like Experian. And then you need to layer that in with your overall needs and what your problem is that you're trying to solve for. 

00:30:21 Richie Cotton: Okay. Lost to digest that. Just making sure I've understood this correctly, it sounds like when you're trying to figure out what do I do in order to if I wanna create an AI product, it sounds like really it's think about what is a good business use case to solve first, and then you figure out okay, what data do I need?

00:30:37 And then you figuring out okay how can I build a good model on top of this? And the rest is just okay, let's let's build a thing, put it in production. Does that sound about right? In terms of what you were explaining? 

00:30:46 Vijay Mehta: I think you, you summarized it in a nice, clean way. 

00:30:48 Richie Cotton: Okay, nice.

00:30:49 Yeah. In terms of creating models it just sound like anything is not part of your core business. You should be trying to outsource either through experience products or other platforms or other technology, and then you really wanna focus on stuff that's aligned your business.

00:31:01 Do you have a sense of what are simple like. Good first use cases then can you give some concrete examples of specific models you want to build or specific AI products that you might be able to build? Quite simply? 

00:31:12 Vijay Mehta: I think it depends. I think what we see is most banks and what they're coming to us for, they're going to be looking at their portfolio.

00:31:20 They're, the declines that they have within their, say their card portfolio and they wanna look at. A tool. They wanna look at solutions around reject inferencing or eligibility analysis to see how they can improve the number of loans and the number of credit cards that they might issue.

00:31:40 And I, I think that's a really clear use case for most in the financial services in the lending world. That is usually where we are coming in as a starting point. But there are other areas too. It depends on. If a customer's really trying to think about account opening specifically and how they look at the false positives that are occurring.

00:31:59 So it really does depend on that, that first problem statement. But I would say everyone, everyone has a view of this and it, it just depends on. Overall, what the overall goal of the business is. I, it does go back to jobs to be done and how you define your first statement of what problem am I trying to solve?

00:32:19 And going from there you break it into smaller pieces, build a data foundation and then apply the tooling around it. 

00:32:26 Richie Cotton: Okay. Wonderful. Yeah really, so it's gonna depend a lot on what your business is as to what a good sort of first project's gonna be. So I'd like to talk a little bit about compliance.

00:32:36 You mentioned before finance is incredibly heavily regulated industry. Just talk me through your process for how you make sure that. Any models that get built are gonna be compliant with any regulations? 

00:32:47 Vijay Mehta: Yeah, so we're a company that's, we work in highly regulated industries. We've invested in the resources to ensure that we've got proper governance generative ai, especially as it relates to any innovation that we're doing the guardrails that we have in place.

00:33:01 Protect consumer pri privacy, and that's really vital I think with all the emerging government regulations, that carry some moderate or severe penalties. Misuse of AI and data or data is becoming more and more relevant. So as we think about this transparency explainability and how you monitor transparency, explainability within AI is absolutely critical.

00:33:26 And that's the tooling that we've invested heavily in. So as part of the model governance and model risk management process, part of that is to ensure that throughout the entire life cycle, how many people are working on the model, the lineage, the history. How it executes appropriate ab testing needs to take place.

00:33:46 And having all of that documented in in an efficient manner. And that's how we think about the life cycle. So it's following good model development practices, going from feature engineering all the way into deployment, but ensuring that you still have your governance and you have your risk levels, first line of defense, second line of defense, et cetera.

00:34:05 Place to make sure that it happens. I can't stress enough. Explainability, transparency of execution is really at the core of most of what the regulators are looking for. So if you can't go into your model and explain why it's, why it's doing something then you're gonna have a problem.

00:34:22 And the tools that we have help solve that. 

00:34:24 Richie Cotton: Okay. Yeah, that certainly seems to make a lot of sense, is that if you need, if, regulator is gonna come and ask you what your bottle's doing. You need to be able to explain what's going on. It needs to be very transparent. Does that affect the kinds of models you can use?

00:34:35 I know something like logistic regression or decision trees, that incredibly easy to explain something like a neur network. Less does it affect, like what types of models you can use then? 

00:34:45 Vijay Mehta: I think it does to a certain extent. Actually boosted models have been used in the financial services world for a very long time.

00:34:51 I would say. Obviously regression models, but as you get into neural networks, I think they're used primarily in the non-production environment to have an output then, which is used in a production environment. So it's like how can you create the learnings that you need but not have an autonomous decision making process within AI for financial services.

00:35:12 And that's the, I think the boundary that we see in most credit right now is that. No one is going to have an autonomous underwriting agent that is using a neural network because it, if you're regulated, that's gonna be very difficult to explain, but you can still get the insights of what a neural network can provide and what an autonomous system can provide.

00:35:37 But you need to have a little bit more, I would say. Structured way of actually having that execute in product, in production. 

00:35:45 Richie Cotton: Ah, interesting. Is this suggesting that the agent's gonna give a recommendation, then you want a human to make the final decision? So I guess you've got some accountability there.

00:35:53 Is that the idea behind it? 

00:35:54 Vijay Mehta: To some degree.

00:35:55 I think the answer I think it depends. So when you're dealing in the fraud prevention space. You're able to do more things without the regulators. Being as strict, I would say it falls under a different category, but when you're actually providing credit to a person, when you're providing a loan to a person, there are some things that you need human in the loop still for.

00:36:18 And from the market standpoint, we're not seeing that shift yet. Because there is an appetite to say we're going to completely eliminate human in the loop for for underwriting or for providing a loan. Now, when it comes to certain like auto decisioning solutions, there are auto decisioning solutions, but what I refer to is the broader model explainability aspect of how originations will take place.

00:36:43 There needs to be before before a model is going to go into production. If there does, there is still that that final check of a human signing off on it and saying, this is okay. That's, your risk officer is gonna be responsible for that depends. I think we're gonna see a shift more towards autonomy.

00:37:01 But the regulators aren't quite there just yet. 

00:37:05 Richie Cotton: Okay. Ideally, like the regulators want a human to blame if something goes wrong rather than just a bot where it's like, what? That you can't fire the bot, okay. I guess the other thing you mentioned around this is you want everything to be very well documented now.

00:37:21 I'm curious, have you got people just writing pages and pages of documentation about what's going on? Or is this now possible to automatically generate 

00:37:29 Vijay Mehta: almost a hundred percent autogenerated? So that is one of the really neat things about generative ai and the way that we've thought about language models is we can digest the regulations from.

00:37:44 The OCC, the SR 11 seven is a document that's used widely in banking. We're then able to take the policies that a customer has, ingest that, and then create an automated process around that to generate a document for. A a governance manager or a modeler to use to drastically reduce the amount of time.

00:38:07 So we're generating all of that, using that, using a language model, with training. The model itself we use RAG to ensure that we've got consistency and vectorization to ensure we have consistency. Then that actually streamlines a lot of that process and we're having really great results with it.

00:38:23 We use it internally. If you ever developed a model you have, you've gotta go through and fill out a lot of information. And this is one of those time savers that we've found to be extremely extremely useful and a great application of gen ai. It's not. It's not academic. It's very practical because it does solve a, and it solves a problem.

00:38:45 Richie Cotton: Okay, nice. Yeah, certainly. I imagine just, having documentation about compliance with financial regulations that sounds is probably not the most fun task for humans to do. It's a good time that's mostly automated. Alright. So are there any sort of unsolved problems what's still tricky to be done?

00:39:01 Like probably open research questions at the moment? 

00:39:03 Vijay Mehta: Yeah I think we're getting to that point. You actually, you, you started going down the path a bit. I think the dynamics of when we start having mult multi-agent or swarms and the level of autonomy that the financial services industry is comfortable with, I think this is a still a debate.

00:39:23 I don't think we've landed the plane on it. The fintechs or some of the more flexible financial services companies are gonna push the envelope. At what point is there agreement on how far we go with it? I think that's unsolved. The use of neural networks and going beyond, of course we're still early stage on, on things like quantum, but heck, quantum ai and how we get into that and what that really means.

00:39:50 And obviously that's gonna be big for everything, but it has a potential to change the change that we, the way that we operate in the traditional lifecycle that we're looking at right now. I think we have a pretty good plan of attack. It's more of the what if scenario and where do we go as this technology, as AI advances and continues to advance at the speed that we're moving at.

00:40:13 And what does it mean for, what does it mean for the industry, the finance industry, as well as as the customers who. Who wants more sovereignty maybe around their data. So how do we think about things like lineage and data ownership as well as, the identity management that needs to plug into it?

00:40:29 A long response to your short question, but I think there's a number of areas that we've gotta, we've gotta look at here and continue to focus on as things evolve. 

00:40:37 Richie Cotton: Absolutely. Certainly like lineage and ownership, these incredibly important things, particularly like for any enterprise ai.

00:40:44 But since you mentioned quantum, i've had quite a few people posting on LinkedIn recently about, oh, quantum computing's gonna be viable in like 2030, and it's LinkedIn influencers, so I take it with a grain of salt. But have you started thinking about like how quantum computing's gonna affect your business?

00:41:00 Vijay Mehta: Yeah we absolutely have. I think. Things are still pretty noisy right now. And once they figure out, I think Willow is the first real example of how, the chip makers, how Google is really thinking about the relationship. Back to things like like deepminds and how it can be used in quantum analytics or quantum ai.

00:41:21 But I, I feel like we're still many years away. The inconsistency in the noise has to be reduced for it ever to be in a point where you know it's gonna be adopted by a bank or an organization a. In the business realm. 

00:41:36 Richie Cotton: Okay. So it sounds like maybe for now, just casually research it and then it's next year's problem to figure out the details.

00:41:42 Vijay Mehta: Yeah, it's one of those, I think with every trend, Richie, we're we keep an eye on all of 'em, and Quantum is one that we've got we're staying involved in. We keep the, keep our labs engaged. Blockchain was really big for a while. Very engaged on that. Self sovereignty around identity is ano another one.

00:42:00 Just making sure that, as a company, our focus is to ensure that we've got. A very innovative culture and that we stay ahead of all of the changes that are happening in the tech and AI world. So you gotta keep your, you gotta keep your finger on on the pulse of what's changing in order to be a technology company these days.

00:42:16 And that's what we do at Experian. 

00:42:18 Richie Cotton: Absolutely. Tons of exciting things coming down the road. Is there anything in particular that you're really excited about at the moment? 

00:42:24 Vijay Mehta: I'm very excited about the evolution of a agents at scale, and we're now at, so you look at MCP, I haven't seen a protocol and I've been around for a while.

00:42:35 I got some gray hair. I know both of us have been around, we've seen this before where a new protocol comes out or something like. Happens a new interesting thing. But MCP has taken off faster than anything I've seen. I think it's only been around, for six months, but it's basically a defacto standard now, and all the big players are saying MCP enabled, and I know we are too.

00:42:55 I think that's going to create a whole new set of really interesting problems to solve and opportunities for us as an organization and as the industry as a whole, because. When you actually get to agent communication and agent to agent workflows and swarms and all that it opens up the possibility of doing really interesting deep research things.

00:43:19 But it also opens up the need for greater fraud prevention tools and identity validation and things like that. Very excited about that topic and spending a lot of time on it. 

00:43:30 Richie Cotton: Absolutely. Yeah. For anyone of the audience who, who doesn't know about MCV model context protocol, it's like a standard way for AI agents to communicate.

00:43:38 I feel it's a whole separate episode which we should cover. At some point. But yeah that's a, it's a very exciting standard and yeah, you're right. It's just taken the world by storm in the last few months. Finally I'm always on the lookout for other people to follow research. That's interesting.

00:43:52 Is there anyone in particular's work that you think you'd recommend I take a look at? 

00:43:56 Vijay Mehta: I love the Google DeepMind podcasts that Hannah Fry does. She's fantastic. And it's one of those where. It works all parts of your brain too, Richie. 'cause she'll do like things about protein sequencing and then go into, notebook, LM and sort of everything in between so you can get some really deep thinking and then you can get.

00:44:17 Maybe some more application level episodes as well. So big follower of that podcast and I, I always recommend it to people. 

00:44:24 Richie Cotton: Wonderful. Yeah. Hannah Fry's amazing. I actually had no idea she had a Deep Mind podcast. Good to know. Thank you so much for your time, Vij, it was a pleasure chatting with you.

00:44:32 Vijay Mehta: Pleasure chatting with you as well, Richie. Thank you.