Sitting Down With an Expert

Conversations with Industry Experts

AI in Consumer Fintech

A Conversation with NVIDIA’s Pahal Patangia

“A big chunk of cost drivers in financial institutions tie up to fraud. And it’s essential to use AI in fraud detection particularly, in order to catch all of those bad actors.” 

“What I see in Fintech is AI getting more mainstream. That will give companies a big boost in terms of data. For AI, you need to have data, and once you have the data, the use cases will pile up naturally.” 

— Pahal Patangia 

Pahal Patangia leads global developer relations for consumer fintech companies at NVIDIA. He focuses on driving AI adoption within the fintech industry and helps them to accelerate their machine learning models at scale. He also leads fintech oriented engagements with NVIDIA’s partners, cloud service providers, influencers, venture capitalists and startups in the ecosystem. Previously, Pahal spent a majority of his professional career at FICO, where he helped retail banks and financial institutions make smarter credit decisions using machine learning. He holds a master’s degree in business analytics from the University of Minnesota and a bachelor’s degree in electrical engineering from the National Institute of Technology, Trichy, in India.  

Innodata sat down with Pahal to discuss how AI and ML can enhance business efficiency, revenue, and customer service in the financial services industry. 

Innodata: Thanks so much for sitting down with us, Pahal. To start with, what would you say are the top use cases for AI in fintech? 

PP: One of the top use cases that we see in this area, particularly with customers, is conversational AI. If you are using a bank or investing app, you will have a conversational agent pop up, which augments and supports the already overloaded call center employees. And conversational AI has many different components. For example, with a talking agent, there would be an automatic speech recognition component, a natural language understanding component to interpret the words, and then a natural language processing component which comprehends the question and brings back information. Then it would have a text-to-speech component to send the signal back to the person on the other side. 

Conversational AI in general is a fascinating topic that we have been seeing in different banking and fintech companies. They have been using chatbots for customer service, at the same time, integrating the chatbots with next-best action strategies. For example, if I’m a customer and I’m talking to you about purchasing a loan, then what could be the next best rate I could get, customized for my interests as well as suited to my risk profile? This would be an integration of conversational AI and personalization. The chatbot would adaptively see the risk profile of the customer and then generate the best options for that particular customer. 

Another use case is personalization — making targeted recommendations to customers. You’ve probably heard about companies like NerdWallet and Credit Karma who have made their foray into financial services, FSI, and fintech through personalization and targeted recommendations. If you are a customer, based on your needs, spending patterns, previous purchase histories, etc., what would be the best-suited product for you? AI and recommender system models in particular excel at such catered recommendations and are a very hot commodity. We recently had a session with Capital One at NVIDIA’s GTC conference where they showcased the use of NVIDIA’s transformer based recommender system models for displaying personalized landing pages for their existing customers. So that was an interesting use case, a fresh story coming directly out of what we have been seeing. 

A third, and probably the most important, use case is around fraud detection. Fraud detection can take various shapes. It could be KYC ID verification or anti-money-laundering checks, it could be transaction fraud when you’re swiping a credit card, or it could be account takeovers. Or, if you are sending a claim, how do you identify fraud within the insurance claim? So those are some of the primary facets of fraud detection that we are seeing. It is an extremely important use case in terms of reducing losses for a bank. A big chunk of cost drivers in financial institutions tie up to fraud. And it’s essential to use AI in fraud detection particularly, in order to catch all of those bad actors. 

Fourth among the many is automation. Gone are the days when hundreds, if not thousands, of pages of documentation are manually reviewed by banking representatives. Nowadays, companies are using a combination of computer vision and NLP models to mine relevant information from such documents. Another example of automation is in insurance claims — Ping An, one of the largest insurers in the world, uses computer vision models trained on NVIDIA GPUs to classify accident claims and verify them by looking at the depths of dents, identify damage, etc., for more than 30,000 claims a day. 

Beyond this, there are many other use cases like ESG (Environmental, Social, and Governance), avatars, risk prediction, transaction monitoring, and robo-investing, where NVIDIA’s accelerated computing platform is unlocking business value via AI.  

Innodata: Could you give us some examples of results that AI-based fraud detection models have delivered for your customers? 

PP: I can give you a few anecdotes on this. The first one is American Express. We worked with American Express to help them build deep learning models using NVIDIA GPUs for solving their day-to-day credit card fraud detection challenges. With American Express, we found that they were able to make decisions with less than two milliseconds latency and had a 6% reduction  in their fraud rate. And if you think of it on the scale of Amex, which is massive, imagine the dollar savings for the bank in fraud costs, not to mention the enriched customer experience with reduced false positives. 

Another example is PayPal, where they are using graph neural networks (GNNs) to train fraud detection models on the NVIDIA platform. Graph data and graph neural networks have become increasingly popular in the fraud detection arena because they capture more interconnected network information for a customer. In tabular data you would see one row for one person, another row for another person, and a third row for a third person. In graph data you would know how person #1 is connected to person #2 and person #3, so you can determine if there’s a corrupted chain, or connections that are engaging in nefarious activities. PayPal has seen close to a 10% reduction in fraud and have reduced their computational costs by 8x, all powered by NVIDIA GPUs. 

There are further interesting stories from BNY Mellon, who have been working with an ISV called Inpher to build AI fraud detection models with more extensive and diverse datasets. BNY built a collaborative fraud detection framework that runs Inpher’s secure multi-party computation — which safeguards third party data — on NVIDIA DGX systems. BNY Mellon’s GPU-powered ML and AI models outperform rules-based models, improving fraud prediction accuracy by 20% while also preserving the privacy and residency of the input training data.  

These are just a few examples of success stories from big players leveraging AI-based fraud detection. 

Innodata: Where do you think AI will have the most impact on the user experience? 

PP: Imagine you are a customer applying for a loan or a credit card. Usually the bank or the fintech would say, “Ok, you will get approved in 10 seconds” and that’s it — yay or nay, your decision should be there. But once you’ve applied, let’s say for some reason the decision hasn’t come and you’ve been sent to a manual review. Imagine how frustrating it can be when you have to wait for two or three days to get your credit card approved. I’ll give you a personal example of how this can impact customers. Just last year, I was looking for travel cards right before booking a huge upcoming trip. I applied for a travel card, and it was crucial that it arrive on time so that I could book the flights; otherwise prices would go up. If these decisions get stuck in manual review queues, which are already piling up with the increase in digital adoption in the COVID era, it’s very difficult for banks to keep up with their growth story and have all of their customer requests serviced at the same time. That’s where AI comes in, to augment the manual review process. It reduces false positives, creating a faster time to value for customers. False positives are a big problem for companies because they are directly tied to the customer experience. Reducing these false positives builds customer loyalty in the long term. Establishing that level of trust at the point of onboarding is very important. 

In another example, we are working with Square, whose chatbot called Square Assistant can understand and provide help for 75 percent of customers’ questions. It is reducing appointment no-shows by 10 percent. Under the hood, they are training models like BERT, which are the popular transformer models for processing large language data and performing tasks like question answering, summarization etc. They are using NVIDIA GPUs to automate responses to 75% of the seller inquiries that come onto their platform, thus enabling efficient servicing and minimizing backlog. At the end of the day, all of this makes customers happy. 

Innodata: Could you give us an example of a fraud detection false positive, and how AI could prevent it? 

PP: Let’s say that I live in the US, but I’m on vacation in Europe, and I make a large transaction. My card gets blocked because it’s an unknown location and may not be legitimate. But I’ve had a great history and I’ve been doing everything right so far. How can the bank ensure that I am a legitimate customer making a legitimate transaction, and this is not fraud? This is where building AI machine learning models helps. These models learn the patterns in the data from an existing pool of customers, which is why they are better at recognizing what’s fraud and what’s not. This is a more accurate and efficient way of detecting fraud versus a manual review or a rule-based model. The AI model, having understood the existing trends, can look at a new event and decide that ok, this one is an anomaly, I have to wait it out, or I have to tag it as fraud, etc. 

Innodata: In the example of a cardholder going overseas and making a large purchase, how would an AI model treat that differently, compared to a traditional rule-based system?

PP: What works there is seeing your old patterns in terms of where you have made transactions. You might have made a transaction at the airport, or booked a flight or a cab. You have likely shown signs that you were going to travel to that place, and now you are there. So that is the pattern which an AI model would see in the data. This could be viewed as an indicator that this person had an intent to travel, so likely it is the same person and not fraudulent. 

Innodata: Since data-centricity is a current AI trend, could you talk a little about NVIDIA’s comparative focus on data vs. models, and how they handle data sourcing and preparation? 

PP: Let me frame it this way. To start building a model, you need to have the right data. How is that ensured? By having the right kinds of checks. Are all the distributions that you are seeing in the data broadly expected? How are you selecting your data sources? Do those sources make sense? From where else is the data coming in? If this data is being tied into one format, does it tell a comprehensive story? It’s very important to pick and choose which data source to start with. Everybody in the data org, right from data engineers, data scientists, and ML engineers, needs to have a sense of what kind of data is coming in. The utmost care and data governance policies need to be applied before you really get started.   

Innodata: Are you seeing a need for synthetic data in fintech? 

PP: Yes, absolutely. Synthetic data is getting hot in fraud, mainly because there isn’t much data available from the rare class, the fraud class. It is necessary to generate samples which look like fraud, to help distinguish between what’s fraud and what isn’t. This is where synthetic data generation and the use of GANs (generative adversarial networks) come in. A very recent example was with Swedbank and a community called Hopsworks, where they used GANs for anomaly detection for fraud and we put up a whole developer blog. There, they are using graph neural networks as well, to catch anomalous/fraud cases. These are neural network models that are used for synthetic data generation. 

For banks, it would mean developing a roster of synthetic data which would show what it looks like to be a good customer vs. a bad customer — in cases like these, data is a problem. You have to simulate data so that you can have enough samples of fraud and non-fraud classes, and there needs to be a large enough pool of data fed into the machine learning model so that its results make sense.  

This is also the case with anomaly detection, insurance, auto claims fraud, etc. They generate synthetic data in the form of images, for building and training their models to identify which claim images are fraudulent and which are not. 

Another great use case for synthetic AI data generation in fintech and FSI is simulations for risk calculations in capital markets, like Monte Carlo simulations, where they test out different scenarios of how commodities will interact with each other. Synthetic data is used to stress-test models on extreme values and worst-case scenarios.  

Innodata: What is the most exciting use case for AI in fintech that you’ve seen lately? 

PP: Something that we are seeing very recently is the use of avatars in financial services. It has been catching on, and a lot of banks have been asking about how they can integrate avatars into their point-to-point operations. I think that’s something that we wouldn’t have imagined using a few years ago. But it makes sense from a financial services perspective. Banks are interested in using avatars right from customer onboarding to customer service, and also for employee training. These avatars can be so close to mimicking a perfect human that they give you a full immersive experience. Plus, they are omni-channel; not just web, not just mobile, maybe also AR and VR. We are seeing avatars pretty much across industries, in retail as well as other domains. There are some great applications here, and I’m excited to see this use case develop in the next few years. 

Innodata: Where do you see AI taking fintech in the next ten years? 

PP: This is a very tricky question. What I see in Fintech, particularly, is AI getting more mainstream. Companies will have different points of engagement with customers and they will have a holistic view of a customer’s needs, spends, and interests. That will give them a big boost in terms of data. For AI, you need to have data, and once you have the data, the use cases will pile up naturally. 

Innodata: To wrap things up, could you give us an example of how AI will bring in a competitive edge? 

PP: People will have different touch points. Let’s say you use Google Pay. Whatever interactions you have outside Google Pay, Google has your history on those. In your Google Pay, for example, they would have your spend patterns as well. So you see the concept of embedded finance coming into play. 

This is prevalent in companies trying to become super apps. These are companies like Google, Apple, and Walmart, who want to come into financial services. There are these new tech players, then there are the already disruptive fintechs, and then there are the banks, who are incumbents, but have the full-blown capacity to cause a lot of disruption at speed. It’s critical for these companies to have AI done right, to get a competitive advantage using AI. 

All of these companies have different ways to capture holistic views of customers’ interests, spending, and more. Once they have that at scale, they’ll be able to keep adding use cases like conversational AI, personalization, fraud detection, and document automation. That’s what will make these next few years very interesting. Whoever invests in AI, not just from a model-building perspective, but from a full stack perspective; who has not just the libraries, but also the hardware and the end-to-end pipeline; not just the data collection but the data prep, training, and inferencing pieces all in one, like a platform, they’ll be succeeding the most. And that’s where I see AI bringing in the most positive change in fintech. It will help banks, financial institutions, and fintechs get that extra edge over their competitors. 

Learn more about NVIDIA’s AI and HPC solutions for financial services  

follow us

Ready to Scale and Train Your AI Models?

Solve Your Toughest Data Engineering Challenges Using Artificial Intelligence and Human Expertise

(NASDAQ: INOD) Innodata is a global data engineering company delivering the promise of AI to many of the world’s most prestigious companies. We provide AI-enabled software platforms and managed services for AI data collection/annotation, AI digital transformation, and industry-specific business processes. Our low-code Innodata AI technology platform is at the core of our offerings. In every relationship, we honor our 30+ year legacy delivering the highest quality data and outstanding service to our customers.