Plaid CEO Zachary Perret on LendScore — a real-time credit score built on cash flow and network data

works for you and we have Zach Bray from Plaid in the Reream waiting room let's bring in the TVP and Ultradome. How are you doing guys? Nice to see you. Thank you for having me. Good to see you. Looking sharp. Are you in DC today or is it just you're just dressed for the occasion of of uh TVPN? TBPN.

I mean, you all dressed so nicely when you're you're you're doing the news, I figured I had to dress up, too. No, I'm I'm in DC. It's DC. There you couldn't tell, we we borrowed a bunker to to to film this from. So, it looks great. Yeah. Very very very sound padded. Uh yeah, give us give us the update.

Everyone in the audience is familiar with the company, but uh what what's the latest and greatest in your world? So today's a big day for us. We we do uh two big product releases each year. This is our fall product release. Um lots of new exciting stuff that that just came out.

Um probably chief among it is uh a product that we call lend score. It is a new credit score. We've been working on a bunch of credit analytics products for many years. And this lend score gives a really robust picture of a potential borrower. Let's say that you got a new job. Uh your income went up a lot.

We could actually recognize that kind of thing in real time. uh let's say that you sign up for seven loans all in the last 24 hours. It's probably pretty sketchy. You can recognize that kind of thing in real time. So, a big upgrade to to to the credit scores.

Also launched a bunch of stuff around um risk and fraud analytics, payments, so on and so forth. So, big big day for us today. How do legacy fraud scores work?

Like I'm I'm familiar with, you know, the it's out of 800 and you get a random number and if you pay off but there's all these like weird like uh there's these weird things and you hear recommendations like oh don't pay off this loan early because that'll actually hurt your score.

It always feels very uh very indecipherable but like what what what did you not like about the current system? What's currently broken? Well, the the way that you feel is the way that everybody feels about credit scores in the past. um they're they're kind of a 350 to 850 score.

It's largely based on your your history of payments and the amount of loans that you've taken out in the past and it's not responsive to the changes in your life in real time. It also skips a lot of important data inputs. Um now the the lending industry has been built on these credit scores for for decades.

Um and they work okay. Um but when you think about uh an average lender, they want more information because today the fundamentals of lending have changed. There's a lot more attempted fraud in lending.

Um there's a lot more um new insights that you could be gathering and frankly consumers change their patterns a lot more. So the seven-year look back doesn't make a ton of sense.

I myself, I had um gone to an optometrist and that optometrist self sent the bill to an old address for me and my credit score was hit because I never got that bill for years and then I finally figured it out. I had the same thing happen. Yeah.

My first my first company, we had a whole bunch of like Wi-Fi hotspots on Verizon contract. This happened to me. This exact same happened to me, too. My credit got dinged because I sent the router back and they were like, "We never got it. " Yeah.

I was like, "So, you lost it and you have my card on file and you didn't just charge me for it and then just deal with it. " Yeah. It's very, very frustrating. I'm sure there's so much frustration. I have a bunch more questions, but sorry I cut you off. Yeah.

I guess uh my my main question is like okay obviously the industry and consumers deserve like a more modern credit score but it feels like these the existing scores are heavily entrenched and it's it's like you can make a better product but how do you actually get it to mass adoption among you know uh lenders broadly?

[Music] It's a great question. Um the the new product that we launched is focused on looking at consumers in real time. So, um, how do we look at your income? How do we look at your expenses? You can do the math to figure out free cash flow.

And that should be the amount that a consumer can pay back in a given month or year. Um, the way that we're launching it originally is actually in partnership with a lot of folks. So, we'll partner with the existing bureaus with with a lot of the existing lenders um to add this alongside FICO or alongside Vantage Score.

So, you can look at a FICO and then you can say um, oh, but this user said that they just changed jobs. Let me get the data from that new job.

Uh so we can add that in and then there's this interesting insight that we have which which only we have because we work with all of the lenders and we can see the velocity of loans applied for. We call this network insights.

There's a bunch of other data points that we can see just based on the traffic that you're doing across the entire plaid network. Um so it's kind of a unique vector that people can now lend against.

Um the net effect is we're seeing you know call it often times 20% lift in the quality of outcomes uh from from this new one.

Um, and we're seeing a lot more people actually be eligible for loans because let's say they have a really high income and good uh rent payment history that actually should indicate that they're a better loan. How do you think about I I imagine that there's a very complex algorithm to actually decide the score.

There's probably some machine learning or AI in there. Um, but you also want to deliver something that's like a human readable insight so you can tell someone why their score dropped or what they can do to make their score better.

Maybe I I have a friend who works in it's not high frequency trading, it's like out like it's quantitative uh trading so it's like a couple days out and uh and he was like we'll crunch all these numbers.

We'll come up with some like thesis for buying a billion dollars of some company in a day and we'll be selling in like three days.

Um, and I was like, "Is there any way that you could say, "Oh, well, it's because like if gold goes up, like, could you put it in business logic or like if statements that a human could understand? " He was like, "It's impossible.

" Like, there's so many different parameters that go into finding like alpha in the financial markets. And I feel like there's a world where a consumer would be very disappointed if you're like, "Yeah, your score went up one point, but we have no idea why. And we can't tell you the first thing about why it went up.

just these tensors, these weights activated in this massive model and your score went down. Sorry. Uh so how do you think about actually educating the consumer on like how how their score goes up or down?

So we we do a bunch of machine learning and AI to obviously like clean up the data and make it structured in a way that we can build our models on top of it. But ultimately our models are representative roughly of the financial health of a consumer.

So almost everything that that a consumer would do to better manage their money will end up with a better credit score. Um almost everything they would do to uh more poorly manage their money will end up with a worse credit score.

Um and so the nice thing is when we build it, we're actually able to show it to the consumer first. They're able to review it if they want to click submit and then it goes to the lender so the lender can then subsequently make the decision on the back of it. That's cool.

So, so it's very lender friendly in that they're opting in if they want to just run with their Experian score or whatever whatever options they have, they can they can do that or they can add this on as a way to kind of improve their application. Exactly. Yeah.

So, the over time though, over the over time though, the lender would realize, hey, we're getting actually much better insight hereal. We want to just, you know, roll with this. I imagine that's what we hope. Um but the goal at first is that consumers feel like they can get more and better loans.

Lenders feel like they can make more and better loans. So it's kind of opening the market initially and then over time um you know we hope that this becomes one of the most important if not the most important way that a lender makes their decisions. How does buy now pay later fit into the ecosystem at this point?

I imagine that they all have like their own systems but then they would love to just have more data because that probably enables better underwriting. How do they fit into this ecosystem? So, buy now pay later are sophisticated lenders. There's a scale of sophistication of all lenders.

Um, you know, um, there are some lenders that will just use an out-of-the-box FICO score or an out-of-the-box lend score, um, to make a lending decision. Um, the highly sophisticated lenders, they'll want a zillion inputs.

Yes, they'll still get maybe a FICO, they'll they'll get a plaid lend score, but they want to look at the underlying inputs, the underlying raw data, and then they'll build their own models on top of it. I would say the BNPL lenders are quite sophisticated in the way that they lend.

Um, and so they're they're pulling a ton of data into the picture and then building kind of their own custom models on top of it. Now, our inputs are yes, still inputs to their model, but they're not looking at our score as the sole determining factor. Yeah.

Is the business model for the lens score the same as the previous credit scores or different in some way? How do you think about that? Pretty similar. Yeah. We we sell it on a kind of per loan application basis. Uh our customers are the lenders. They pay us for it. Got it.

Um there's a handful of things that we can do to actually help them track uh creditworthiness on an ongoing basis. So whereas FICO or Vantage often times is a one-time snapshot. Um sometimes a lender will want to maybe make offers to a consumer if they realize they want to increase their credit line.

um or um maybe they want to kind of give better preferential payment terms to a consumer um if they think they can pay a loan down faster. Um and so ours is an ongoing credit score. That's maybe the only big difference. How do you think about uh what the consumer's interface is to the credit scores?

I remember like years ago you used to have to actually go to the three credit go to the credit store. Yeah. you had to go to the credit store and like check out and pay like 30 bucks and then you'd get that and then you'd take that to a loan application sometimes. Sometimes they just do it on the back end. They'd pay.

Uh but then over time I feel like the credit scores uh just became almost commoditized or free. Like it just shows up in my, you know, banking app a lot.

Uh, and there were also like Credit Karma and different companies that figured out ways to basically pay that as a like a customer acquisition cost for free and then they give you a bunch of products on top that they would monetize.

How do you think the the what do you think the long-term consumer adoption of the lens score looks like? Well, we hope it's pretty similar actually. We want consumers to be able to get a clear view of their credit profile for free. Um, we're working on building ways that a consumer can come in and and and look at that.

Um, and you know, ultimately the customer then would be the lender. Like it's it's to our benefit to have consumers that understand this more. Um, it's to our benefit to have consumers that trust this score and that that that trust the things that we build.

Uh, and so we want to be as as direct and open with the consumer as we can. Um, so a lot of that stuff is to come and you'll you'll see some launches from us uh kind of over the coming coming quarters on that front. Uh, what else uh besides lens? uh was the focus today.

Uh so a big big upgrade to our anti-fraud model and we launched a kind of broader uh suite of products called protect. Um so again protect for us um it's a set of products that looks at a consumer's underlying actions that they take across um almost any financial product.

Um and we can back it into an overall score of riskiness for the consumer and then a bunch of attributes. So some of the attributes are very easily explainable such as this person just signed up for three crypto apps in the last 30 minutes.

Um some of the attributes are less easily explainable more on the the the machine learning side. Um but this stuff will feed into the the models for a lot of our customers and then they can build kind of this really fascinating step up step down requirement for users based on their relative levels of riskiness.

Um, this has been a huge unlock for a lot of the fintech companies that are trying to sign up more users faster. Um, we can create a fast lane effectively. So, we've seen you before. We recognize your device. We've seen your phone number. You just do a two-factor code and you're good to go.

Um, and then we can also create a slow lane. So, um, hey, we've never seen you before. Your device doesn't seem to match the IP address that you're you're dialing in from. Um, what's going on with all this stuff?

Let's let's add a bunch of step ups and um, you know, be really really certain that you are who you say you are. So, fast for financial services. That's what we're trying to build. This sounds like the worst nightmare of somebody doing large scale financial fraud on the internet.

What does uh what what's the shape of AI adoption at Plaid right now? Um I feel like obviously there's probably a ton of AI that you're doing on just like software development stuff, but at the same time I feel like if I was a thing with Plaid and it was like, "Hey, you want to chat with me right now?

" I'd be like, "No thanks. I'm actually good to just click a couple buttons.

" So like are you seeing a lot of like LLM token consumption or more just focused on the machine learning the custom models for underwriting where you train a specific model for that and you're not just leaning on a big like token factory for a lot of work.

I I'd say um our our use of uh tokens I would call it like medium. Uh sure we do build kind of a custom training data set um and then run a ton of the data through that that model and the dating the training data set. Um, a lot of it is around just cleaning up transactions, identifying what kind of thing is happening.

Um, like the textual analysis is very very helpful for that. So that's that's like the input into a lot of our models. Yeah. Um, but you're right, we don't build like an active chat app for any consumer. Um, a lot of the AI that we're doing is used to fight AI.

Uh, meaning there's a lot of attempted AI fraud and then we have a lot of AI on the other side to try to fight AI fraud. Um uh and then it's it's kind of building the foundational kind of data set upon which we build a lot of the models. Yeah.

So whenever you have a big pile of text in in kind of behind the scenes in the back end, you can run LLMs over that and get a lot finer grain data, categorize, understand what's actually going on with those large text dumps, but you don't necessarily even need to surface that to the to the customer, the end user, which is great.

Generally, not directly. Generally, we'll we'll service it to the consumer indirectly saying, "Hey, you had elevated bank fees. " And we won't even send that to the consumer. We'll send that to our customer.

Our customer will then say, "All right, in your budgeting app, you had elevated bank fees and look at all these different bank fees. " That makes a ton of sense. Well, congrats on the launch. Thanks so much for stopping by. Always great to see you. Have fun at Have a good one. Cheers, Zach.

Before we hop on with our second guest of the show, let me tell you about Fall, the generative media app for developers. The world's best generative image, video, and model audio models all in one place. Develop and fine-tune models with serverless GPUs and on demand clusters.

And our next guest is about to be in the reream waiting room. We can go back to the timeline. Black Rockck and Nvidia are participating in a $40 billion data center takeover also including XAI and Microsoft. That's great news. This is in the Financial Times this morning.

Invest an investment consortium that includes Black Rockck Global Infrastructure Partners and that's a great name for uh a fund that does global infrastructure investing.

Uh and of course Abu Dhabi fund MGX uh has struck a $40 billion takeover of one of the world's largest data center operators as it launches an initiative to underwrite the infrastructure for artificial intelligence. Uh and I believe we have our next guest in the