Pagaya Co-Founder and CTO Avital Pardo Speaks at the 2023 JMP Securities Inaugural Artificial Intelligence Forum


Artificial Intelligence (AI) has seemingly taken the world by storm in the past year. Although the benefits of adopting and integrating AI appeal to many banks and financial service providers, implementation can be daunting.


As a visionary with longstanding expertise in the transformative power of artificial intelligence, Pagaya Co-Founder and CTO Avital Pardo was recently invited to speak at the inaugural Artificial Intelligence Forum hosted by JMP Securities.


In a compelling conversation with David Scharf, a Research Analyst at JMP, Pardo discussed the evolution of Pagaya’s unique business model, including the infusion of artificial intelligence into every facet of operations, as well as shedding light on Pagaya’s strategy and future growth over the next five to ten years.
During the discussion, Pardo painted a vivid picture of the “evolution of AI and its use cases, [which] now revolutionize the credit decision-making and enablement process.”


He shared how AI has shifted from being a technological novelty to fundamentally changing how decisions are made across industries. However, he cautioned that barriers to entry make it difficult for banks that are not heavily invested in technology to realize the full benefits of the marriage of technology and data.


“This is why we built our business model,” explained Pardo, “to help banks serve their clients using our technology without taking additional risks, helping them utilize AI without building it themselves.”


Pardo went on to offer a glimpse into Pagaya’s ambitious vision for the future, including insights into the company’s strategic long-term roadmap. He envisions a future where the expansion of AI-powered lending and risk management solutions with partners across industries will redefine how services are delivered, making them more efficient, accessible, and profitable.


Listen to the entire session below:

Full Session Transcript


David Scharf (0:03):

Welcome everybody. Good morning, at least out here in the Bay Area! Good afternoon to most of the people listening in or participating. I’m David Scharf, a research analyst at JMP covering consumer finance and fintech with a heavy emphasis on consumer lending and marketplace lending. As part of our inaugural AI forum, we welcome Pagaya Technologies to participate here for a nice chat. We actually launched coverage on the company late summer. 

We’ve actually been highlighting, In the course of a lot of investor conversations, a lot of unique aspects of their marketplace model, particularly the use of pre-funded securitizations. But, today, we want to focus more on the technology side, specifically AI – how it’s incorporated into their business, some insights in particular about how it’s changing and evolving, and how we think it will look over the next 5-10 years.

We’re welcoming with us Avital Pardo, who, in addition to being the co-founder of the company, is also its Chief Technology Officer. Avital, before I launch into questions, why don’t you take just a minute or so to introduce yourself and your background?

Avital Pardo (1:38):

Sure. And thank you for having me, and thanks to everybody for joining. I’m Avital, CTO and one of the co-founders of Pagaya. I come from a tech background and have a master’s in Math. I’m a graduate of one of the elite tech units in the Israeli army called the Talpiot. Served for around 10 years. This is the place where I found my passion for data. I spend most of my service utilizing Big Data for different things, especially in the Intelligence Corps. I left the service around ten years ago and went and worked at a small Fintech start-up in Tel Aviv in the lending space. My job was to fully automate the lending process. Over there, I realized the huge impact AI can have on credit and credit accessibility. This is where I started thinking about building Pagaya, a company that would focus on building AI credit and credit-decisioning in technology to expand credit accessibility and serve consumers in a better way. 

On a more personal note, I recently got married. I spend my time between Tel Aviv and New York.

David Scharf (3:12)

Terrific. Thanks for joining us. You’re obviously the right person to be talking to on this topic. I’m going to start with probably either it’s the simplest question or it’s most complicated question. Certainly, in the last year, we’ve been hearing the term “AI” or “artificial intelligence” more and more. It’s moved beyond technology and financial services almost into popular culture now. But at just at a very high level, how do you define AI as a technologist and, more specifically, within the context of Pagaya’s business?

Avital Pardo (3:58):

First of all, I’m happy that people have started hearing and talking about AI for the past year. We’ve been speaking about AI since we founded the company. Seven years ago, when we told people that AI is going to revolutionize credit, it was a more difficult conversation than it is now. I’m happy to see development in this area. There are a lot of ways to define AI. My best way is, utilizing technology and data in order to automate the things humans used to do, especially in our case, in decision-making. In our specific case, in credit decisions. Talking in the context of our business, when we founded the company, we understood that traditional methods in credit are not serving people in the best way. This is why we decided to bring AI to this industry: to better serve people and give better access to credit to more people.

This is where we spent most of our focus during the first couple of years on the company, building the team. We now have over 100 data scientists in Tel Aviv, some of the most talented people in the country. Building the team, building the infrastructure, working with the data, and spending a lot of time and effort building this technology. When we started, it was mainly around credit underwriting, credit-decisioning. Now, we’ve expanded to have AI models working more in the marketing realm, in fraud prevention, servicing, and collection. We’ve expanded it to be a bit broader than what we started with.

David Scharf (6:17):

Got it. That addressed one of the questions I wanted to cover. That is, have the use cases, or the definition of AI evolved over time, beyond just credit underwriting? Particularly for investors in financial services companies or any company with credit exposure, there tends to be, while not a monolithic focus, a very heavy focus on just “everything’s credit, it’s all about credit.” Sometimes, there are a lot of questions on how much better can one analytical engine can price or predict risk than another. On some of these other use cases that maybe we don’t hear it talked about as much, can you discuss a little bit, particularly marketing customer acquisition, as well as fraud – how AI is being deployed in those areas? Whether you would characterize that as very early stage or if it’s actually something investors should be paying closer attention to?

Avital Pardo (7:38)

In different places, we’re in different stages. Some of them, like marketing, we have been working there for a long time. In terms of investor focus, it’s a difficult question. I’m talking from a tech perspective. I can speak about how much focus we’re putting in each of these domains. The main thing is still credit underwriting and credit decisions. We’ve put some focus in marketing just by understanding that if you build better scores around marketing overall, your credit performance is better because it’s not only around once someone gets to your system – how do you score them or how do you make the decision – but also who are the people that are applying?

This is where you can have an impact on the marketing. In fraud prevention, these are things that we’ve started lately. Most of it is around improving credit performance. Some of the credit performance is actually determined by the level of fraud you have in your platform, and reducing that as much as possible, especially as we get larger, have more data sets, have more partners, seeing a much broader view of the market. It’s much easier for us to analyze to analyze fraudulent behavior.

An easy example to give is that if you see someone having some sort of fraudulent behavior with one partner in one place and then you see the same person somewhere else, then it’s easier to detect fraud that way. The fact that we’ve expanded in different industries with different partners gives us a lot of advantages, and we wanted to utilize this advantage in our favor.

David Scharf (9:38):

Understood. I wanted to ask a more detailed question on the technology. This might just be all about language or terminology. It gets back to defining AI. One of the goals of this discussion, and in general, is to not just educate investors on what differentiated technologies you’re using, but also to help them understand, is it really differentiated? A phrase that we’ve heard over the years is “machine learning.” Similarly, I myself have covered digital lenders for over a decade. It’s not like there haven’t been purely digital lenders aggregating alternative data sources over the years. 

Can you talk about, as a technology officer what is different about AI versus machine learning that we’ve heard about for 10 or 20 years? Secondly, what makes Pagaya’s AI proprietary or differentiated? I think that gets to the heart of the discussion. Are we really talking about something groundbreaking and defensible, or is this just a minor tweak of things other lenders have been doing for years?

Avital Pardo (11:24):

That is a very good question. I will touch on both questions – the defensibility piece is really important. 

Starting with the difference between machine learning and AI, I see it as the difference between a cell phone and a smartphone. It’s the next generation of the same technology. When we started the company, I used to work, still work hands-on, but I used to do it more seven, eight years ago, a lot of the things that were extremely difficult to do back then are now much easier to do. Because of the advancement in computational power and the ability to handle large data sets – to work with petabytes of data eight years was impossible Now, it is possible. 

Running a sophisticated model on large data sets eight years ago was very difficult. You needed to build your entire tech stack from zero. There were lots of barriers. Now, it’s much easier. As we started gaining more and more data – and now we’re holding and holding enormous amounts of data – being able to run sophisticated and deep algorithms on this vast amount of data was almost impossible 5-7 years ago. You’re talking about 10-20. Ten years ago, it was very far from being possible. The ability to take a more sophisticated model, run it on larger data sets, and grasp everything you can get from the data extracted to the end – this is the kind of advancement that we’ve seen over the last couple of years, and this would not have been possible without like advancement in computation and modeling. Also, things that are usually not talked about in the entire tech stacks. There was not a lot of shared code in doing so 5 or 10 years ago, and there’s much more now. It’s gotten easier.

Specifically talking about Pagaya, there are different advantages we have starting from the team. Which is one of the largest, and I want to say is the most talented team in the industry. Second is the data – working with different partners in different industries. And now we’re active in four different industries. We’re active in personal loans, auto loans, point-of-sale lending, and credit cards. We’re seeing a lot of cross-data from different partners in different industries. This gives us a huge data advantage that can be utilized. And, of course, our ability to build over time better and better models. 

If I compare us to ourselves a couple of years ago, we have much better technology now. We keep advancing at a fast pace to become better and better. Especially as we’re working with more and more banks, getting access to more data, and seeing more behavior. Some of it is based on seeing more and more macro environment changes, which have been quite dramatic changes over the past couple of years. That is the difference between AI and machine learning, specifically for Pagaya. 

David Scharf (15:24):

Got it. As a marketplace lender, one of the unique aspects is that you’re the only B2B2C model out there. You just referenced, you’re aggregating data from multiple lending partners, whereas a lot of direct-to-consumer marketplaces, whether it’s a Lending Club or an Upstart. They are only learning from their own data set. You seem to be learning from all of the declines that your various lending partners make. To put you on the spot, I know you’re the CTO, but if you had to rank, which is more important in developing your credit analytics, is it the technology or is it the fact that you’re in the unique position of having access to underwriting decisions of so many different lending partners? Or is it impossible to really differentiate the two?

Avital Pardo (16:39):

I want to say both. When we started, we didn’t have the access to so much data. So we can see how much improvement it gave us. We focused a lot on building the technology. The access to unique, proprietary data gives us a huge benefit. What people usually don’t realize is that we don’t only see declines of other partners; we see a lot more that. Within data exchanges, we see much broader of what our partners do and have done over the past couple of years. We can learn from that. The advantage that we get from working with multiple partners, and not just multiple partners, the multiple industries is also extremely important. People behave differently when they go across industries. This is very beneficial for the underwriting and also for fraud modeling, as well, which I discussed a bit earlier.

David Scharf (18:00):

Got it. You mentioned some macro changes as well. That’s something I also want your perspective on. One of the unique aspects of not only companies such as Pagaya, but really so many fintechs, is that they were all founded within the last ten years during a time in which there were unprecedented low interest rates, there was unprecedented low inflation, both of those things with full employment. The last 10 or 15 years represent a fairy tale that never existed before then and will probably never exist again. Yet so many of these technologies and the credit analytics that developed during that time were born and maturing during this, quite frankly, unrealistic macro environment. Is there anything about AI that makes it better equipped to respond to macro changes, like what we’ve seen over the last year? Or, is a 30-year old, 40-year old credit card issuer who says, “Listen we’ve been through a lot of cycles; we’ve seen all this before. Our models are better because we weren’t born in the last 5-10 years.” How would you respond to that devil’s advocate market?

Avital Pardo (19:49):

I think that AI models are learning much faster. This makes them react to changes much faster. We benchmark everything with AI models and with traditional models as well. We can see that the AI models are adopting much faster, and this is due to the case that their ability to intake data is built into the system. Everything we do, all the data, there’s a feedback loop – everything that comes from loans that we issue, loans that our partner’s issue, and general loans – are immediately taken into the model. So all new data – this month’s issuing and this month’s payments from all loans are immediately taken into the models and updated. This is a huge advantage for companies like us. We update credit models really often.

We can see how once things change, we can see the model adapts to that and how it behaves. In the past five years, we’ve been through some dramatic changes. We started from a rather stable environment, went into the COVID shock, then to incentives and then to high inflation, then to inflation stabilizing, and maybe now, hopefully, inflation is starting to go down. These are three, four different environments in a very short period of time. This is I think unprecedented. We’ve seen how the model adapts to all of these environments very quickly. When the model expects something, and it sees that the behavior is actually different, it learns that very quickly. This is what these models are trained to do: learn what’s happening very quickly.

But, it’s not only about general macro behavior; some of it is about specific behavior. We have a lot of examples of that. I think the most interesting one is from the last couple of months regarding student loans. We’re not active in student loans, but people with student loans are now behaving rather differently than they used to behave a couple of months ago because of the changes that occurred in this industry. We can actually test in real time how long it takes for the model to adapt to that. This is a segment – it’s not a full macro change. But it’s definitely a segment within the borrowers that we can track in real-time to see how fast the model adapts to the changes that are happening. It’s really interesting to see that it adapts very quickly.

David Scharf (22:35):

That’s a great insight. Lenders always have to respond to changes in the macro environment. To the extent that you’re able to have a better handle for the next recession or the next spike in inflation, you’ve got a certain degree of predictability learning that’s taking place in your models that helps you respond more effectively down the road.

Only got a couple more minutes, I want to close out with one last question on how we ought to think about AI within the context of competition or barriers to entry. On the one hand, you have looked at over a trillion dollars of loan application volume in terms of the lifetime of the company after just seven years. At the same time, are tools coming out that just make it easier for another start-up to enter the business? I think of ChatGPT – as a consumer, I could do things in five seconds that would take me five hours a year ago. How do you think about AI in the consumer lending and credit world as a tool for new companies to enter, build, learn, and create models in 24 months when it took Pagaya 60 months to create? Is that a legitimate, competitive issue?

Avital Pardo (24:26):

This is a very good question, and I want to answer it in both ways. The short answer is that I think the entry barriers are actually now higher than they used to be. There are two types of competition. One is inside competition within our partners building similar technology. The second is from another company. When we founded the company, as I think I mentioned, we understood the impact that AI can have over credit. Another thing that we understood, is that the ability for a large institution to adopt this technology by itself is almost impossible. This is why we built our business model – to help banks serve their clients using our technology without taking additional risks, helping them utilize AI without actually building the technology themselves. This business model is gaining more traction. We’ve seen in the last year, more and more banks have started approaching us to work with them. In the context of this type of competition, working with more and more teams inside banks, their ability to build this is almost impossible. They’re not oriented in that direction. You need to be tech-heavy to build that. From the competition side, we’ve learned over the past couple of years, is not just the models and the data, but it’s also the partner’s behavior. When we started working with Ally Bank, for example, we were not as good as we are now. 

Over time, you get much better in working with a specific partner because the AI model adapts to the behavior of the applicants and specific channels and specific partners, that are actually different than applicants or customers of other partners of ours. You gain a better understanding of how the mechanism on the other side works and AI just learns that – it learns how the mechanism on the other side works. So it becomes much more adaptable to specific partners. Another issue, of course, is getting the level of data that we have now. Starting now, I don’t want to say it would be impossible, but I would say very difficult. We’ve gained a lot of data and gained a lot of knowledge over these years, and I don’t think that someone using ChatGPT can build this faster.

David Scharf (27:28):

It comes full circle to where I started. Even though the business model of B2B2C is not necessarily a technology differentiator, it kind of is. Access to all these lending partners, in a lot of ways, is as much a technology differentiator as it is a business model.

We’ve run out of time here. I really appreciate you taking part, Avital. It’s been a fascinating journey for the company over the last few years, and I know we didn’t get to talk about what’s on the horizon for the next five years in terms of your predictions. But we’ll save that for the next forum.

Avital Pardo (28:13):

Thank you very much. Thank you, I really enjoyed the conversation.

David Scharf (28:16)

Ok. Bye now.

Avital Pardo (28:17)